From c1976cec5adbef099d7ec086967803dfc5c3f4b6 Mon Sep 17 00:00:00 2001
From: Yev Meyer <ymeyer@nvidia.com>
Date: Mon, 2 Nov 2026 20:45:57 -0500
Subject: [PATCH 1/5] fix(engine): find a built-in's own banks before its
 module is imported

A built-in kept in its own folder declares usersim/engine/probes/<label>/assets/ as its assets_root. Until its class registers, the asset lookup takes that folder as the declared root when it holds the probe's folder, so a prompt read while the probe's module loads, as general_educational and general_open_ended read theirs, and a scorer in a process that never imports the probe reach the probe's assets.

A test reads general_educational's seeds and general_open_ended's prompts in a fresh process with no probe registered.

Signed-off-by: Yev Meyer <ymeyer@nvidia.com>
---
 architecture/assets.md                     |  5 +++-
 src/usersim/engine/core/_assets.py         | 35 ++++++++++++++++++----
 tests/engine/core/test_assets_discovery.py | 23 ++++++++++++++
 3 files changed, 56 insertions(+), 7 deletions(-)

diff --git a/architecture/assets.md b/architecture/assets.md
index 056dc2b..6587c44 100644
--- a/architecture/assets.md
+++ b/architecture/assets.md
@@ -31,7 +31,10 @@ banks wherever the package is installed, with no environment variable to
 set. Registration refuses a relative root. That root is searched for that
 probe only, after the first three entries and before the packaged directory,
 so a deliberate override wins over it. A built-in probe kept in its own
-folder declares one too, that folder's `assets/`. A probe's bank loader reads
+folder declares one too, that folder's `assets/`, and the lookup finds that
+folder before the probe's module is imported: a prompt read while the module
+loads, or a scorer in a process that never imports the probe, reaches the
+probe's assets. A probe's bank loader reads
 under its class's `assets_dir()`, the folder this search finds for the probe,
 and its `prompts.yaml` is read from that folder too.
 
diff --git a/src/usersim/engine/core/_assets.py b/src/usersim/engine/core/_assets.py
index b55129d..9e871a8 100644
--- a/src/usersim/engine/core/_assets.py
+++ b/src/usersim/engine/core/_assets.py
@@ -127,13 +127,35 @@ def declared_assets_roots() -> dict[str, Path]:
     }
 
 
+def _builtin_probe_root(probe: str) -> Path:
+    """``usersim/engine/probes/<probe>/assets/``, where a built-in kept in its own folder holds ``<probe>/``."""
+    return Path(str(resources.files("usersim.engine") / "probes" / probe / "assets"))
+
+
+def _declared_root(probe: str) -> Path | None:
+    """The root ``probe``'s class declares, found before its module is imported if it is a built-in.
+
+    A built-in kept in its own folder declares ``usersim/engine/probes/<probe>/assets/``.
+    Until its class is registered, that folder stands in for the declaration when it
+    holds ``probe``, so a lookup made while the probe's module loads, such as a prompt
+    read at import time, or by a scorer in a process that never imports the probe,
+    reaches the probe's assets.
+    """
+    declared = declared_assets_roots().get(probe)
+    if declared is None and probe not in probes._PROBE_REGISTRY:
+        builtin = _builtin_probe_root(probe)
+        if (builtin / probe).is_dir():
+            return builtin
+    return declared
+
+
 def probe_asset_roots(probe: str) -> tuple[Path, ...]:
     """The roots searched for ``probe``'s assets, highest priority first.
 
     The search path, with the root ``probe``'s class declares placed just
     ahead of the packaged directory.
     """
-    declared = declared_assets_roots().get(probe)
+    declared = _declared_root(probe)
     roots = [*_override_roots(), *([declared] if declared is not None else []), packaged_assets_dir()]
     return tuple(dict.fromkeys(roots))
 
@@ -177,9 +199,10 @@ def probe_assets_dir(
     folder belongs to a registered probe.
 
     Args:
-        probe: A registered probe label. Its class's ``assets_root``,
-            when declared, joins the search; a label not registered yet
-            is looked up on the search path alone.
+        probe: A probe label. Its class's ``assets_root``, when declared,
+            joins the search. A label not registered yet is looked up on
+            the search path, and in its own folder under
+            ``usersim/engine/probes/`` when it is a built-in kept there.
         assets_root: Pin resolution to exactly this directory, skipping
             the search path. This is what ``--assets-dir`` becomes: a
             user who named a location is told when the assets are not
@@ -199,7 +222,7 @@ def probe_assets_dir(
             return candidate
     # Nothing matched. Report where a user would most expect it to have
     # been: the root the probe declares, else the highest-priority root.
-    declared = declared_assets_roots().get(probe)
+    declared = _declared_root(probe)
     return (declared if declared is not None else roots[0]) / probe
 
 
@@ -215,7 +238,7 @@ def _pinned_probe_dir(probe: str, root: Path) -> Path:
     packaged = packaged_assets_dir().resolve()
     if folder.is_dir() or root.resolve() != packaged:
         return folder
-    declared = declared_assets_roots().get(probe)
+    declared = _declared_root(probe)
     # ``packaged`` is ``usersim/engine/assets/``, so ``parents[1]`` is the package.
     if declared is not None and declared.resolve().is_relative_to(packaged.parents[1]):
         return declared / probe
diff --git a/tests/engine/core/test_assets_discovery.py b/tests/engine/core/test_assets_discovery.py
index 67d6a81..e954dd9 100644
--- a/tests/engine/core/test_assets_discovery.py
+++ b/tests/engine/core/test_assets_discovery.py
@@ -19,6 +19,7 @@
 from __future__ import annotations
 
 import os
+import subprocess
 import sys
 from pathlib import Path
 
@@ -328,3 +329,25 @@ def test_any_other_pinned_root_stays_strict(self, register_test_probe, monkeypat
         register_test_probe("ext_probe", assets_root=own)
 
         assert probe_assets_dir("ext_probe", assets_root=tmp_path / "mine") == tmp_path / "mine" / "ext_probe"
+
+
+class TestBeforeTheProbeIsImported:
+    """A built-in kept in its own folder is found before its class is registered."""
+
+    def test_its_banks_and_prompts_are_found_in_a_fresh_process(self):
+        # A prompt read while the probe's module loads, or a scorer in a process
+        # that never imports the probe, looks the assets up with the probe unregistered.
+        code = (
+            "from usersim.engine.core import probes\n"
+            "from usersim.engine.core.prompt_loader import load_prompt_file\n"
+            "from usersim.engine.core.seeds import load_seeds\n"
+            "assert not probes._PROBE_REGISTRY, 'a probe was registered before the lookup'\n"
+            "print(len(load_seeds('en_US', 'general_educational', 'subjects')))\n"
+            "print(sorted(load_prompt_file('general_open_ended')))\n"
+        )
+        result = subprocess.run([sys.executable, "-c", code], capture_output=True, text=True)
+
+        assert result.returncode == 0, result.stderr
+        count, keys = result.stdout.splitlines()
+        assert int(count) > 0
+        assert "user_agent_system_prompt" in keys

From 1feb467fbcd58ba775497bbf3063f7c9a5402fff Mon Sep 17 00:00:00 2001
From: Yev Meyer <ymeyer@nvidia.com>
Date: Mon, 2 Nov 2026 20:46:08 -0500
Subject: [PATCH 2/5] refactor(engine): keep sov_ai_facts in one folder

The probe module is probes/sov_ai_facts/probe.py, beside its bank loader bank.py, its scorer scorer.py and its banks in assets/sov_ai_facts/, the folder the class names as its assets_root. Its tests are in tests/probes/sov_ai_facts/. The label, the scorer name, the banks, the stored columns and the cache the probe and its scorer share are unchanged.

DEFAULT_SCORER_MODULES, the capability source references and the docs name the new modules, and the shared-code label allowances list neither the bank loader nor the scorer, which sit under engine/probes/.

Signed-off-by: Yev Meyer <ymeyer@nvidia.com>
---
 docs/engine/AUTHORING_A_PROBE.md              |  4 +-
 docs/engine/README.md                         | 15 +++++---
 src/usersim/engine/core/agentic_bank.py       |  2 +-
 src/usersim/engine/core/pressure_bank.py      |  2 +-
 src/usersim/engine/core/probes.py             |  2 +-
 src/usersim/engine/core/probing_taxonomy.py   |  4 +-
 src/usersim/engine/core/query_bank.py         |  2 +-
 .../engine/evaluator/scorers/__init__.py      |  2 +-
 .../evaluator/scorers/sov_ai_dynamic.py       |  4 +-
 .../engine/probes/sov_ai_facts/__init__.py    |  2 +-
 .../assets/sov_ai_facts/SCHEMA.md             |  2 +-
 .../assets/sov_ai_facts/de_DE/sample.yaml     |  0
 .../assets/sov_ai_facts/en_GB/sample.yaml     |  0
 .../assets/sov_ai_facts/en_IN/sample.yaml     |  0
 .../assets/sov_ai_facts/en_SG/sample.yaml     |  0
 .../assets/sov_ai_facts/en_US/sample.yaml     |  0
 .../assets/sov_ai_facts/fr_FR/sample.yaml     |  0
 .../sov_ai_facts/hi_Deva_IN/sample.yaml       |  0
 .../sov_ai_facts/hi_Latn_IN/sample.yaml       |  0
 .../assets/sov_ai_facts/ja_JP/sample.yaml     |  0
 .../assets/sov_ai_facts/ko_KR/sample.yaml     |  0
 .../assets/sov_ai_facts/pt_BR/sample.yaml     |  0
 .../sov_ai_facts/bank.py}                     |  8 ++--
 .../sov_ai_facts/{generator.py => probe.py}   | 19 +++++-----
 .../sov_ai_facts/scorer.py}                   | 16 ++++----
 .../probes/sov_ai_facts/task_derivation.py    |  2 +-
 src/usersim/taxonomy/capabilities.py          |  4 +-
 templates/probe/README.md                     |  2 +-
 templates/probe/test_probe.py.template        |  2 +-
 tests/engine/core/test_agentic_bank.py        |  2 +-
 tests/engine/core/test_assets_discovery.py    |  2 +-
 tests/engine/core/test_default_asset_paths.py | 10 ++---
 tests/engine/core/test_pressure_bank.py       |  2 +-
 tests/engine/core/test_query_bank.py          |  2 +-
 .../probes/test_india_variant_probes.py       |  2 +-
 tests/engine/probes/test_probe_invariants.py  |  2 +-
 tests/engine/probes/test_sov_ai_dynamic.py    |  2 +-
 .../probes/test_sov_ai_multilingual_parity.py |  2 +-
 tests/probes/__init__.py                      |  2 +
 tests/probes/sov_ai_facts/__init__.py         |  2 +
 .../sov_ai_facts/test_bank.py}                |  8 ++--
 .../sov_ai_facts/test_probe.py}               | 30 +++++++--------
 .../sov_ai_facts/test_scorer.py}              | 38 +++++++++----------
 tests/test_capabilities.py                    |  2 +-
 tests/test_probe_names_in_shared_code.py      |  2 -
 tests/test_toolset_seed_cli.py                |  4 +-
 46 files changed, 108 insertions(+), 100 deletions(-)
 rename src/usersim/engine/{ => probes/sov_ai_facts}/assets/sov_ai_facts/SCHEMA.md (99%)
 rename src/usersim/engine/{ => probes/sov_ai_facts}/assets/sov_ai_facts/de_DE/sample.yaml (100%)
 rename src/usersim/engine/{ => probes/sov_ai_facts}/assets/sov_ai_facts/en_GB/sample.yaml (100%)
 rename src/usersim/engine/{ => probes/sov_ai_facts}/assets/sov_ai_facts/en_IN/sample.yaml (100%)
 rename src/usersim/engine/{ => probes/sov_ai_facts}/assets/sov_ai_facts/en_SG/sample.yaml (100%)
 rename src/usersim/engine/{ => probes/sov_ai_facts}/assets/sov_ai_facts/en_US/sample.yaml (100%)
 rename src/usersim/engine/{ => probes/sov_ai_facts}/assets/sov_ai_facts/fr_FR/sample.yaml (100%)
 rename src/usersim/engine/{ => probes/sov_ai_facts}/assets/sov_ai_facts/hi_Deva_IN/sample.yaml (100%)
 rename src/usersim/engine/{ => probes/sov_ai_facts}/assets/sov_ai_facts/hi_Latn_IN/sample.yaml (100%)
 rename src/usersim/engine/{ => probes/sov_ai_facts}/assets/sov_ai_facts/ja_JP/sample.yaml (100%)
 rename src/usersim/engine/{ => probes/sov_ai_facts}/assets/sov_ai_facts/ko_KR/sample.yaml (100%)
 rename src/usersim/engine/{ => probes/sov_ai_facts}/assets/sov_ai_facts/pt_BR/sample.yaml (100%)
 rename src/usersim/engine/{core/fact_bank.py => probes/sov_ai_facts/bank.py} (98%)
 rename src/usersim/engine/probes/sov_ai_facts/{generator.py => probe.py} (98%)
 rename src/usersim/engine/{evaluator/scorers/sov_ai_facts.py => probes/sov_ai_facts/scorer.py} (97%)
 create mode 100644 tests/probes/__init__.py
 create mode 100644 tests/probes/sov_ai_facts/__init__.py
 rename tests/{engine/core/test_fact_bank.py => probes/sov_ai_facts/test_bank.py} (98%)
 rename tests/{engine/probes/test_sov_ai_facts.py => probes/sov_ai_facts/test_probe.py} (97%)
 rename tests/{engine/evaluator/test_evaluator_scorer_sov_ai_facts.py => probes/sov_ai_facts/test_scorer.py} (93%)

diff --git a/docs/engine/AUTHORING_A_PROBE.md b/docs/engine/AUTHORING_A_PROBE.md
index a6767b9..ec51a0f 100644
--- a/docs/engine/AUTHORING_A_PROBE.md
+++ b/docs/engine/AUTHORING_A_PROBE.md
@@ -1219,7 +1219,7 @@ covers (1) for you parametrically. So your per-probe file can focus
 on (2)-(8) plus the should_succeed contract: the cross-cutting
 checks come for free.
 
-For real working examples: `tests/engine/probes/test_sov_ai_facts.py`
+For real working examples: `tests/probes/sov_ai_facts/test_probe.py`
 (verbatim), `tests/engine/probes/test_safety_chat_pressure.py` (reframings and
 abort policy), `tests/engine/probes/test_safety_agentic.py` (custom
 run_dispatch and a simpler mock).
@@ -1294,7 +1294,7 @@ goes on the class: see
 
 The scorer is its own design exercise (axis definition, rubric
 authoring, threshold gating). See
-[`evaluator/scorers/sov_ai_facts.py`](../../src/usersim/engine/evaluator/scorers/sov_ai_facts.py)
+[`probes/sov_ai_facts/scorer.py`](../../src/usersim/engine/probes/sov_ai_facts/scorer.py)
 or `safety_chat_pressure.py` for canonical examples.
 
 ---
diff --git a/docs/engine/README.md b/docs/engine/README.md
index 6e23a39..e44ca6e 100644
--- a/docs/engine/README.md
+++ b/docs/engine/README.md
@@ -128,7 +128,7 @@ different questions of the model.
 
 | Asset | Loader | Unit | Per-locale shape | Used by |
 | --- | --- | --- | --- | --- |
-| Fact bank | `core/fact_bank.py` | One fact per entry (with `question`, `false_premises`, `ground_truth`) | One file per locale, content authored separately per locale | `sov_ai_facts` |
+| Fact bank | `probes/sov_ai_facts/bank.py` | One fact per entry (with `question`, `false_premises`, `ground_truth`) | One file per locale, content authored separately per locale | `sov_ai_facts` |
 | Probing / dynamic-tier taxonomy | `core/probing_taxonomy.py` | One topic category per entry (`invitation`, `subtopic_hints`, optional `applies_to_types` + `persona_affinities`); shared `select_category` derivation | One file per locale (`categories.yaml` for sov_ai_dynamic; `dynamic.yaml` for financial_services) | `sov_ai_dynamic`, `financial_services` |
 | Financial-services bank | `core/finance_bank.py` (+ `finance_tasks.py`, `embeddings.py`) | Per-institution corpus + tools + tasks + embeddings; `Institution` / `FinanceBank` typed views; sim-time gold derivation | One dir tree per locale (`<type>/<institution>/`) | `financial_services` |
 | Query bank | `core/query_bank.py` | One **conceptual question** per entry (with N locale `renderings` + `concern`) | **One file shared across locales; N renderings per entry joined by `query_id`** | `sov_ai_multilingual_parity` |
@@ -547,7 +547,7 @@ what each one carries:
 | Module | Purpose |
 |--------|---------|
 | `core/_assets.py` | Resolves the runtime assets directory (`set_runtime_assets_dir` / `probe_assets_dir`) so every loader uses the same root |
-| `core/fact_bank.py` | Per-locale fact-bank loader, schema validator, persona-tag matcher (consumed by `sov_ai_facts`) |
+| `probes/sov_ai_facts/bank.py` | Per-locale fact-bank loader, schema validator, persona-tag matcher (consumed by `sov_ai_facts`) |
 | `core/probing_taxonomy.py` | Per-locale dynamic-tier taxonomy loader + schema validator + the shared `select_category` derivation (optional `applies_to_types` scoping + `persona_affinities` soft-weighting); consumed by `sov_ai_dynamic` and `financial_services` |
 | `core/finance_bank.py` | Per-locale multi-institution financial-services bank loader (`Institution` / `FinanceBank`, scope-integrity, embeddings sidecar); `core/finance_tasks.py` (sim-time instantiation + gold derivation) + `core/embeddings.py` (query embedding); consumed by `financial_services` |
 | `core/query_bank.py` | Cross-locale query-bank loader, schema validator with rendering × locale cross-product check, persona-tag matcher (consumed by `sov_ai_multilingual_parity`) |
@@ -597,7 +597,6 @@ plugin entry points, undeclared result columns). The tables describe the main fi
 | `test_multi_locale.py` | Multi-locale config, seed loading, language mapping |
 | `test_llm.py` | LLM call helpers (retries, debug log, per-model stats, per-conversation outcome-builder hook) |
 | `test_analysis.py` | Score extraction, demographic preservation, edge cases used by the deep-dive analysis script |
-| `test_fact_bank.py` | Fact-bank YAML schema validator, loader, persona-tag matching, locale-keyed shared cache |
 | `test_probing_taxonomy.py` | Probing-taxonomy YAML schema validator, loader, lookup helpers, locale-keyed shared cache |
 | `test_query_bank.py` | Cross-locale query-bank YAML schema validator, loader (with cross-product rendering × locale validation + opt-out semantics), persona-tag matcher, file-path-keyed cache, env override, shipped-sample end-to-end check |
 | `test_pressure_bank.py` | Pressure-strategy bank YAML schema validator, loader with template-placeholder validation, deterministic reframing rotation with wrap-around, persona-tag matcher, cache + env override, shipped-sample end-to-end |
@@ -612,7 +611,6 @@ plugin entry points, undeclared result columns). The tables describe the main fi
 | `test_probe_invariants.py` | Cross-probe invariants every shipped probe upholds (registration, side-channel keys, prompt-version monotonicity, etc.) |
 | `test_tool_calling.py` | Tool-calling probe: tool sampling, gate check |
 | `test_verifiers.py` | Tool-call JSON-schema validator |
-| `test_sov_ai_facts.py` | `sov_ai_facts` probe: persona→tag, deterministic `derive_task` (bank-version salted), prompts, e2e with mocked LLM |
 | `test_sov_ai_dynamic.py` | `sov_ai_dynamic` probe: persona→tag, deterministic `derive_probe` + hint rotation (taxonomy-version salted), all-locale prompt availability, e2e with mocked LLM |
 | `test_sov_ai_multilingual_parity.py` | `sov_ai_multilingual_parity` probe: persona→tag re-export, deterministic `derive_task` with locale opt-out filtering (bank-version salted), all-locale prompt completeness with no-half-shipped invariant, query-bank cache + env override, e2e with mocked LLM, verbatim turn-1 injection, structured failure paths |
 | `test_safety_chat_pressure.py` | `safety_chat_pressure` probe: persona→tag re-export, `derive_task` with independent strategy/target salts (50-seed cell-coverage test), `resolve_task_from_row` (full / partial / neither override + persona-tag guard against panel-driven mismatches), prompt-formatting helpers, end-to-end with mocked LLM (verbatim turn-1, multi-turn rotation, bank-load failure, panel-id miss, no-resolvable-pair, assistant-error attribution, follow-up-failure breaks loop) |
@@ -633,7 +631,6 @@ plugin entry points, undeclared result columns). The tables describe the main fi
 | `test_evaluator_scorers.py` | Scorer registry (`register_scorer` / `get_scorer` / `clear_registry` / `reload_scorers`) |
 | `test_evaluator_scorer_structured_output_schemas.py` | Cross-cutting schema-design contract for LLM-judge scorers (Literal-typed enums, Optional fields, 1–5 score constraints): fails the build on violations |
 | `test_evaluator_scorer_tool_use.py` | 8-axis tool-calling trajectory judge |
-| `test_evaluator_scorer_sov_ai_facts.py` | `sov_ai_facts` out-of-sim scorer: per-fact LLM judge with question-type-specific axes, error paths, multi-fact aggregation |
 | `test_evaluator_scorer_sov_ai_dynamic.py` | `sov_ai_dynamic` out-of-sim scorer: response-only LLM judge with `sovereign.{suspected_fabrication,self_consistency,graceful_unknown}` axes, multi-category aggregation, drift detection |
 | `test_evaluator_scorer_sov_ai_multilingual_parity.py` | `sov_ai_multilingual_parity` out-of-sim respect-parity scorer: single LLM judge per trajectory with `respect.{condescension,complexity_calibration,helpfulness_at_proficiency}` axes, concern-grounded rubric, locale+language interpolation, status-proposal gating, error paths, version-mismatch drift |
 | `test_evaluator_scorer_safety_chat_pressure.py` | `safety_chat_pressure` out-of-sim scorer: single LLM judge per trajectory with `pressure.{refusal_held,contextual_safety_persistence}` scored axes plus per-turn classifications + derived `turn_of_flip`, rubric grounded in `target.expected_refusal` + `strategy.description`, status-proposal gating, error paths, version-mismatch drift |
@@ -643,3 +640,11 @@ plugin entry points, undeclared result columns). The tables describe the main fi
 | `test_evaluator_scorer_language_compliance.py` | Deterministic (no-LLM) scorer: per-trajectory `language.{requested_language_match_rate, script_compliance_rate, first_turn_match}` rates from `lingua` language detection + Unicode script analysis; clean-Hindi / English-fallback / Hindi-in-Latin-transliteration / Japanese-mixed-scripts cases; per-turn diagnostic surfacing |
 | `test_evaluator_scorer_response_shape.py` | Deterministic (no-LLM) scorer: per-classifier unit tests (empty/trivial, over-formatted, mid-sentence truncation) + trajectory-level aggregation |
 | `test_evaluator_scorer_refusal_basics.py` | Deterministic (no-LLM) scorer: refusal lexicon detection across shipped locales, `refusal.in_wrong_language_rate` (English-refusal-in-non-English-locale) composite, canned-AI-phrase detection, status-proposal gating semantics |
+
+**`tests/probes/<label>/`: a probe's own tests, beside each other**
+
+| Test File | Coverage |
+|-----------|----------|
+| `sov_ai_facts/test_probe.py` | `sov_ai_facts` probe: persona→tag, deterministic `derive_task` (bank-version salted), prompts, e2e with mocked LLM |
+| `sov_ai_facts/test_scorer.py` | `sov_ai_facts` out-of-sim scorer: per-fact LLM judge with question-type-specific axes, error paths, multi-fact aggregation |
+| `sov_ai_facts/test_bank.py` | Fact-bank YAML schema validator, loader, persona-tag matching, locale-keyed shared cache |
diff --git a/src/usersim/engine/core/agentic_bank.py b/src/usersim/engine/core/agentic_bank.py
index f7dd146..b5b7515 100644
--- a/src/usersim/engine/core/agentic_bank.py
+++ b/src/usersim/engine/core/agentic_bank.py
@@ -32,7 +32,7 @@
 - :func:`load_agentic_bank` — parse + validate a single file.
 - :func:`load_agentic_bank_default` /
   :func:`reset_agentic_bank_cache` — process-local cache mirroring
-  the pattern in :mod:`core.fact_bank`, :mod:`core.probing_taxonomy`,
+  the pattern in :mod:`probes.sov_ai_facts.bank`, :mod:`core.probing_taxonomy`,
   :mod:`core.query_bank`, and :mod:`core.pressure_bank`.
 
 Architecturally an agentic-action bank is **shared across locales** —
diff --git a/src/usersim/engine/core/pressure_bank.py b/src/usersim/engine/core/pressure_bank.py
index 55295b9..774e346 100644
--- a/src/usersim/engine/core/pressure_bank.py
+++ b/src/usersim/engine/core/pressure_bank.py
@@ -29,7 +29,7 @@
 - :func:`load_pressure_bank` — parse + validate a single file.
 - :func:`load_pressure_bank_default` /
   :func:`reset_pressure_bank_cache` — process-local cache mirroring
-  the pattern in :mod:`core.fact_bank`, :mod:`core.probing_taxonomy`,
+  the pattern in :mod:`probes.sov_ai_facts.bank`, :mod:`core.probing_taxonomy`,
   and :mod:`core.query_bank`.
 
 Architecturally a pressure bank is **shared across locales** — like
diff --git a/src/usersim/engine/core/probes.py b/src/usersim/engine/core/probes.py
index 769461e..ed3fe7d 100644
--- a/src/usersim/engine/core/probes.py
+++ b/src/usersim/engine/core/probes.py
@@ -1551,7 +1551,7 @@ def decorator(cls: type[BaseProbe]) -> type[BaseProbe]:
     "usersim.engine.probes.safety_agentic.generator",
     "usersim.engine.probes.safety_chat_pressure.generator",
     "usersim.engine.probes.sov_ai_dynamic.generator",
-    "usersim.engine.probes.sov_ai_facts.generator",
+    "usersim.engine.probes.sov_ai_facts.probe",
     "usersim.engine.probes.sov_ai_multilingual_parity.generator",
     "usersim.engine.probes.tool_calling.generator",
 )
diff --git a/src/usersim/engine/core/probing_taxonomy.py b/src/usersim/engine/core/probing_taxonomy.py
index 573c367..6704480 100644
--- a/src/usersim/engine/core/probing_taxonomy.py
+++ b/src/usersim/engine/core/probing_taxonomy.py
@@ -21,7 +21,7 @@
 - :func:`load_probing_taxonomy_for_locale` / :func:`reset_probing_taxonomy_cache`
   — locale-keyed shared cache so the probe and the ``sov_ai_dynamic``
   scorer agree on ``taxonomy_version`` from one source of truth
-  (same pattern as ``core.fact_bank``).
+  (same pattern as ``probes.sov_ai_facts.bank``).
 """
 
 from __future__ import annotations
@@ -358,7 +358,7 @@ def _taxonomy_salt(taxonomy: ProbingTaxonomy) -> int:
 # Family + locale-indexed cache (shared by a domain's probe and its scorer)
 # ---------------------------------------------------------------------------
 #
-# Same pattern as ``core.fact_bank``: one cache, one source of truth, so the
+# Same pattern as ``probes.sov_ai_facts.bank``: one cache, one source of truth, so the
 # simulator-side and evaluator-side both see the same ``taxonomy_version`` for
 # cross-run drift detection. Keyed by (family, locale, filename) so multiple
 # domains (sov_ai_dynamic, financial_services, ...) share the machinery.
diff --git a/src/usersim/engine/core/query_bank.py b/src/usersim/engine/core/query_bank.py
index fd9d062..fd182f2 100644
--- a/src/usersim/engine/core/query_bank.py
+++ b/src/usersim/engine/core/query_bank.py
@@ -26,7 +26,7 @@
 - :func:`load_query_bank_default` / :func:`reset_query_bank_cache`
   — process-local cache for the (typically single) bank used by the
   probe and downstream scorers, mirroring the pattern in
-  :mod:`core.fact_bank` and :mod:`core.probing_taxonomy`.
+  :mod:`probes.sov_ai_facts.bank` and :mod:`core.probing_taxonomy`.
 
 Architecturally a query bank is **shared across locales** (one file,
 N renderings per entry) — this is the structural difference from the
diff --git a/src/usersim/engine/evaluator/scorers/__init__.py b/src/usersim/engine/evaluator/scorers/__init__.py
index cd17beb..edae6fc 100644
--- a/src/usersim/engine/evaluator/scorers/__init__.py
+++ b/src/usersim/engine/evaluator/scorers/__init__.py
@@ -314,7 +314,7 @@ def reload_scorers(scorer_module_paths: Iterable[str]) -> None:
 # probes add their scorer modules to this list.
 DEFAULT_SCORER_MODULES: tuple[str, ...] = (
     "usersim.engine.evaluator.scorers.tool_use",
-    "usersim.engine.evaluator.scorers.sov_ai_facts",
+    "usersim.engine.probes.sov_ai_facts.scorer",
     "usersim.engine.evaluator.scorers.sov_ai_dynamic",
     "usersim.engine.evaluator.scorers.sov_ai_multilingual_parity",
     "usersim.engine.evaluator.scorers.safety_chat_pressure",
diff --git a/src/usersim/engine/evaluator/scorers/sov_ai_dynamic.py b/src/usersim/engine/evaluator/scorers/sov_ai_dynamic.py
index d52425e..e94b592 100644
--- a/src/usersim/engine/evaluator/scorers/sov_ai_dynamic.py
+++ b/src/usersim/engine/evaluator/scorers/sov_ai_dynamic.py
@@ -13,7 +13,7 @@
 :mod:`core.probing_taxonomy`), and scores the assistant's behavior on
 three response-only axes using a single LLM judge.
 
-Structurally inverse of :mod:`evaluator.scorers.sov_ai_facts`:
+Structurally inverse of :mod:`probes.sov_ai_facts.scorer`:
 
 - Sovereign-knowledge scoring compares the response to a curated
   ground truth. The taxonomy gives the judge a *target answer* to
@@ -43,7 +43,7 @@
 ============================================== ==========================
 
 Note ``sovereign.graceful_unknown`` shares its name with the
-:mod:`evaluator.scorers.sov_ai_facts` axis, and
+:mod:`probes.sov_ai_facts.scorer` axis, and
 ``sovereign.suspected_fabrication`` parallels its
 ``sovereign.fabrication``. The semantics are
 deliberately compatible (lower fabrication is better in both;
diff --git a/src/usersim/engine/probes/sov_ai_facts/__init__.py b/src/usersim/engine/probes/sov_ai_facts/__init__.py
index 558a60a..05120a2 100644
--- a/src/usersim/engine/probes/sov_ai_facts/__init__.py
+++ b/src/usersim/engine/probes/sov_ai_facts/__init__.py
@@ -10,6 +10,6 @@
 LLM fabrication.
 
 Entry point: :func:`simulate_sov_ai_facts` in
-:mod:`.generator` — registered as ``sov_ai_facts``
+:mod:`.probe` — registered as ``sov_ai_facts``
 on ``_PROBE_REGISTRY`` in :mod:`usersim.engine.generator`.
 """
diff --git a/src/usersim/engine/assets/sov_ai_facts/SCHEMA.md b/src/usersim/engine/probes/sov_ai_facts/assets/sov_ai_facts/SCHEMA.md
similarity index 99%
rename from src/usersim/engine/assets/sov_ai_facts/SCHEMA.md
rename to src/usersim/engine/probes/sov_ai_facts/assets/sov_ai_facts/SCHEMA.md
index caf1d30..f650b7e 100644
--- a/src/usersim/engine/assets/sov_ai_facts/SCHEMA.md
+++ b/src/usersim/engine/probes/sov_ai_facts/assets/sov_ai_facts/SCHEMA.md
@@ -9,7 +9,7 @@ that rest on them (see §Placeholder discipline).
 
 This schema lives here so:
 
-- the loader in `usersim/engine/core/fact_bank.py` can validate
+- the loader in `usersim/engine/probes/sov_ai_facts/bank.py` can validate
   incoming banks;
 - probe scorers can encode against a stable contract;
 - native-language reviewers know exactly what to fill in (and what's
diff --git a/src/usersim/engine/assets/sov_ai_facts/de_DE/sample.yaml b/src/usersim/engine/probes/sov_ai_facts/assets/sov_ai_facts/de_DE/sample.yaml
similarity index 100%
rename from src/usersim/engine/assets/sov_ai_facts/de_DE/sample.yaml
rename to src/usersim/engine/probes/sov_ai_facts/assets/sov_ai_facts/de_DE/sample.yaml
diff --git a/src/usersim/engine/assets/sov_ai_facts/en_GB/sample.yaml b/src/usersim/engine/probes/sov_ai_facts/assets/sov_ai_facts/en_GB/sample.yaml
similarity index 100%
rename from src/usersim/engine/assets/sov_ai_facts/en_GB/sample.yaml
rename to src/usersim/engine/probes/sov_ai_facts/assets/sov_ai_facts/en_GB/sample.yaml
diff --git a/src/usersim/engine/assets/sov_ai_facts/en_IN/sample.yaml b/src/usersim/engine/probes/sov_ai_facts/assets/sov_ai_facts/en_IN/sample.yaml
similarity index 100%
rename from src/usersim/engine/assets/sov_ai_facts/en_IN/sample.yaml
rename to src/usersim/engine/probes/sov_ai_facts/assets/sov_ai_facts/en_IN/sample.yaml
diff --git a/src/usersim/engine/assets/sov_ai_facts/en_SG/sample.yaml b/src/usersim/engine/probes/sov_ai_facts/assets/sov_ai_facts/en_SG/sample.yaml
similarity index 100%
rename from src/usersim/engine/assets/sov_ai_facts/en_SG/sample.yaml
rename to src/usersim/engine/probes/sov_ai_facts/assets/sov_ai_facts/en_SG/sample.yaml
diff --git a/src/usersim/engine/assets/sov_ai_facts/en_US/sample.yaml b/src/usersim/engine/probes/sov_ai_facts/assets/sov_ai_facts/en_US/sample.yaml
similarity index 100%
rename from src/usersim/engine/assets/sov_ai_facts/en_US/sample.yaml
rename to src/usersim/engine/probes/sov_ai_facts/assets/sov_ai_facts/en_US/sample.yaml
diff --git a/src/usersim/engine/assets/sov_ai_facts/fr_FR/sample.yaml b/src/usersim/engine/probes/sov_ai_facts/assets/sov_ai_facts/fr_FR/sample.yaml
similarity index 100%
rename from src/usersim/engine/assets/sov_ai_facts/fr_FR/sample.yaml
rename to src/usersim/engine/probes/sov_ai_facts/assets/sov_ai_facts/fr_FR/sample.yaml
diff --git a/src/usersim/engine/assets/sov_ai_facts/hi_Deva_IN/sample.yaml b/src/usersim/engine/probes/sov_ai_facts/assets/sov_ai_facts/hi_Deva_IN/sample.yaml
similarity index 100%
rename from src/usersim/engine/assets/sov_ai_facts/hi_Deva_IN/sample.yaml
rename to src/usersim/engine/probes/sov_ai_facts/assets/sov_ai_facts/hi_Deva_IN/sample.yaml
diff --git a/src/usersim/engine/assets/sov_ai_facts/hi_Latn_IN/sample.yaml b/src/usersim/engine/probes/sov_ai_facts/assets/sov_ai_facts/hi_Latn_IN/sample.yaml
similarity index 100%
rename from src/usersim/engine/assets/sov_ai_facts/hi_Latn_IN/sample.yaml
rename to src/usersim/engine/probes/sov_ai_facts/assets/sov_ai_facts/hi_Latn_IN/sample.yaml
diff --git a/src/usersim/engine/assets/sov_ai_facts/ja_JP/sample.yaml b/src/usersim/engine/probes/sov_ai_facts/assets/sov_ai_facts/ja_JP/sample.yaml
similarity index 100%
rename from src/usersim/engine/assets/sov_ai_facts/ja_JP/sample.yaml
rename to src/usersim/engine/probes/sov_ai_facts/assets/sov_ai_facts/ja_JP/sample.yaml
diff --git a/src/usersim/engine/assets/sov_ai_facts/ko_KR/sample.yaml b/src/usersim/engine/probes/sov_ai_facts/assets/sov_ai_facts/ko_KR/sample.yaml
similarity index 100%
rename from src/usersim/engine/assets/sov_ai_facts/ko_KR/sample.yaml
rename to src/usersim/engine/probes/sov_ai_facts/assets/sov_ai_facts/ko_KR/sample.yaml
diff --git a/src/usersim/engine/assets/sov_ai_facts/pt_BR/sample.yaml b/src/usersim/engine/probes/sov_ai_facts/assets/sov_ai_facts/pt_BR/sample.yaml
similarity index 100%
rename from src/usersim/engine/assets/sov_ai_facts/pt_BR/sample.yaml
rename to src/usersim/engine/probes/sov_ai_facts/assets/sov_ai_facts/pt_BR/sample.yaml
diff --git a/src/usersim/engine/core/fact_bank.py b/src/usersim/engine/probes/sov_ai_facts/bank.py
similarity index 98%
rename from src/usersim/engine/core/fact_bank.py
rename to src/usersim/engine/probes/sov_ai_facts/bank.py
index 7e512e4..020bfb1 100644
--- a/src/usersim/engine/core/fact_bank.py
+++ b/src/usersim/engine/probes/sov_ai_facts/bank.py
@@ -5,7 +5,7 @@
 
 A fact bank is a curated YAML file of authoritative, locale-specific
 ground-truth facts used by the ``sov_ai_facts`` probe. The schema lives
-in ``assets/sov_ai_facts/SCHEMA.md``.
+in ``assets/sov_ai_facts/SCHEMA.md`` beside this module.
 
 The module exposes:
 
@@ -224,9 +224,9 @@ def load_fact_bank(path: str | Path) -> FactBank:
 def default_fact_bank_path(locale: str) -> Path:
     """Default on-disk path for a locale's fact bank.
 
-    Discovered via ``core._assets.default_assets_dir`` so the loader
-    picks up ``assets/sov_ai_facts/<locale>/sample.yaml`` regardless
-    of CWD. Asset directory matches the consuming probe name.
+    Resolved through ``probe_assets_dir``, so it is the probe's own
+    ``assets/sov_ai_facts/<locale>/sample.yaml`` wherever the package is
+    installed, unless an asset root ahead of it holds ``sov_ai_facts/``.
     """
     from usersim.engine.core._assets import probe_assets_dir
 
diff --git a/src/usersim/engine/probes/sov_ai_facts/generator.py b/src/usersim/engine/probes/sov_ai_facts/probe.py
similarity index 98%
rename from src/usersim/engine/probes/sov_ai_facts/generator.py
rename to src/usersim/engine/probes/sov_ai_facts/probe.py
index 495cd1b..df7e963 100644
--- a/src/usersim/engine/probes/sov_ai_facts/generator.py
+++ b/src/usersim/engine/probes/sov_ai_facts/probe.py
@@ -51,13 +51,6 @@
 from pathlib import Path
 from typing import Any
 
-from usersim.engine.core.fact_bank import (
-    Fact,
-    default_fact_bank_path,
-    load_fact_bank,
-    load_fact_bank_for_locale,
-    reset_fact_bank_cache,
-)
 from usersim.engine.core.locale import SHIPPED_LOCALES
 from usersim.engine.core.outcomes import WarningKind
 from usersim.engine.core.persona import format_persona_for_prompt
@@ -69,6 +62,13 @@
     register_probe,
 )
 from usersim.engine.core.simulation import ConversationState, make_failed
+from usersim.engine.probes.sov_ai_facts.bank import (
+    Fact,
+    default_fact_bank_path,
+    load_fact_bank,
+    load_fact_bank_for_locale,
+    reset_fact_bank_cache,
+)
 from usersim.engine.probes.sov_ai_facts.prompts import (
     PromptsUnavailableError,
     get_followup_instruction,
@@ -104,8 +104,8 @@
 # ---------------------------------------------------------------------------
 #
 # Thin re-exports so tests can patch on the probe module's surface
-# while the cache still lives in ``core.fact_bank`` (where the scorer
-# also reads it from). Single shared cache prevents the simulator and
+# while the cache lives in ``bank.py`` beside this module, where the scorer
+# also reads it from. Single shared cache prevents the simulator and
 # the scorer from disagreeing on what bank version was used.
 
 _load_bank_for_locale = load_fact_bank_for_locale
@@ -165,6 +165,7 @@ class SovAiFactsProbe(BankVerbatimMixin, BankBackedProbe):
     """
 
     label = "sov_ai_facts"
+    assets_root = Path(__file__).parent / "assets"
     description = (
         "Tests recall of locale-specific (sovereign) facts drawn from a curated bank, "
         "graded against ground-truth answers. Some questions embed a false premise to "
diff --git a/src/usersim/engine/evaluator/scorers/sov_ai_facts.py b/src/usersim/engine/probes/sov_ai_facts/scorer.py
similarity index 97%
rename from src/usersim/engine/evaluator/scorers/sov_ai_facts.py
rename to src/usersim/engine/probes/sov_ai_facts/scorer.py
index 6775d36..94d11dc 100644
--- a/src/usersim/engine/evaluator/scorers/sov_ai_facts.py
+++ b/src/usersim/engine/probes/sov_ai_facts/scorer.py
@@ -8,7 +8,7 @@
 off a trajectory produced by the
 :mod:`usersim.engine.probes.sov_ai_facts` probe,
 re-loads the corresponding :class:`Fact` objects from the locale's
-fact bank (via the shared cache in :mod:`core.fact_bank`), and scores
+fact bank (via the shared cache in :mod:`probes.sov_ai_facts.bank`), and scores
 the assistant's response against each fact's ``ground_truth``.
 
 The scorer is probe-specific: it only makes sense for trajectories
@@ -46,10 +46,10 @@
     create_judge_structured_output_model,
 )
 
-from usersim.engine.core.fact_bank import Fact, load_fact_bank_for_locale
 from usersim.engine.core.llm import acall_llm
 from usersim.engine.core.messages import format_conversation_history_for_prompt
 from usersim.engine.evaluator.scorers import register_scorer
+from usersim.engine.probes.sov_ai_facts.bank import Fact, load_fact_bank_for_locale
 
 logger = logging.getLogger("usersim.engine")
 
@@ -257,7 +257,7 @@ async def score_sov_ai_facts_trajectory(
       no-op scores envelope so running this scorer against a different
       probe's trajectory doesn't explode).
     - Loads the fact bank for the trajectory's locale via the shared
-      cache in :mod:`core.fact_bank`. If any probed fact_id is missing
+      cache in :mod:`probes.sov_ai_facts.bank`. If any probed fact_id is missing
       from the bank (e.g., the bank was edited after the simulator
       ran), that fact is recorded in ``per_fact_scores`` with a
       structured error entry but does not crash the scorer.
@@ -291,8 +291,8 @@ async def score_sov_ai_facts_trajectory(
 
     # India language-variants reuse the en_IN fact bank (presence-aware:
     # a native per-variant bank is used if it has been authored).
-    from usersim.engine.core.fact_bank import default_fact_bank_path
     from usersim.engine.core.locale import asset_locale
+    from usersim.engine.probes.sov_ai_facts.bank import default_fact_bank_path
 
     bank_locale = asset_locale(locale, exists=lambda l: default_fact_bank_path(l).exists())
 
@@ -300,7 +300,7 @@ async def score_sov_ai_facts_trajectory(
         bank = load_fact_bank_for_locale(bank_locale)
     except Exception as e:
         logger.warning(
-            "  |-- evaluator/scorers.sov_ai_facts: failed to load bank for locale=%s (bank_locale=%s): %s",
+            "  |-- sov_ai_facts scorer: failed to load bank for locale=%s (bank_locale=%s): %s",
             locale,
             bank_locale,
             e,
@@ -320,7 +320,7 @@ async def score_sov_ai_facts_trajectory(
     bank_version_mismatch = pinned_version is not None and pinned_version != bank.bank_version
     if bank_version_mismatch:
         logger.info(
-            "  |-- evaluator/scorers.sov_ai_facts: trajectory pinned "
+            "  |-- sov_ai_facts scorer: trajectory pinned "
             "bank_version=%s but current loaded bank is %s for locale=%s — "
             "reporting drift flag on this row",
             pinned_version,
@@ -428,7 +428,7 @@ async def _score_one_fact(
         )
     except Exception as e:
         logger.warning(
-            "  |-- evaluator/scorers.sov_ai_facts: judge %r raised on fact %s: %s: %s",
+            "  |-- sov_ai_facts scorer: judge %r raised on fact %s: %s: %s",
             judge_alias,
             fact.id,
             type(e).__name__,
@@ -445,7 +445,7 @@ async def _score_one_fact(
         parsed = json.loads(content)
     except (json.JSONDecodeError, TypeError):
         logger.warning(
-            "  |-- evaluator/scorers.sov_ai_facts: failed to parse structured output for fact %s",
+            "  |-- sov_ai_facts scorer: failed to parse structured output for fact %s",
             fact.id,
         )
         return _error_result(
diff --git a/src/usersim/engine/probes/sov_ai_facts/task_derivation.py b/src/usersim/engine/probes/sov_ai_facts/task_derivation.py
index a5f4b9e..f0123b8 100644
--- a/src/usersim/engine/probes/sov_ai_facts/task_derivation.py
+++ b/src/usersim/engine/probes/sov_ai_facts/task_derivation.py
@@ -25,7 +25,7 @@
 import random
 from typing import Any, Iterable
 
-from usersim.engine.core.fact_bank import Fact, FactBank
+from usersim.engine.probes.sov_ai_facts.bank import Fact, FactBank
 
 # ---------------------------------------------------------------------------
 # Brazilian region classifier
diff --git a/src/usersim/taxonomy/capabilities.py b/src/usersim/taxonomy/capabilities.py
index 4f923a9..3b91526 100644
--- a/src/usersim/taxonomy/capabilities.py
+++ b/src/usersim/taxonomy/capabilities.py
@@ -991,7 +991,7 @@ def _identity_pressure_by_strategy() -> tuple[CapabilityDefinition, ...]:
 
 _PROBE_IMPL_REFS = {
     "tool_calling": "src/usersim/engine/probes/tool_calling/generator.py",
-    "sov_ai_facts": "src/usersim/engine/probes/sov_ai_facts/generator.py",
+    "sov_ai_facts": "src/usersim/engine/probes/sov_ai_facts/probe.py",
     "sov_ai_dynamic": "src/usersim/engine/probes/sov_ai_dynamic/generator.py",
     "sov_ai_multilingual_parity": "src/usersim/engine/probes/sov_ai_multilingual_parity/generator.py",
     "safety_chat_pressure": "src/usersim/engine/probes/safety_chat_pressure/generator.py",
@@ -1004,7 +1004,7 @@ def _identity_pressure_by_strategy() -> tuple[CapabilityDefinition, ...]:
     "language_compliance": "src/usersim/engine/evaluator/scorers/language_compliance.py",
     "response_shape": "src/usersim/engine/evaluator/scorers/response_shape.py",
     "refusal_basics": "src/usersim/engine/evaluator/scorers/refusal_basics.py",
-    "sov_ai_facts": "src/usersim/engine/evaluator/scorers/sov_ai_facts.py",
+    "sov_ai_facts": "src/usersim/engine/probes/sov_ai_facts/scorer.py",
     "sov_ai_dynamic": "src/usersim/engine/evaluator/scorers/sov_ai_dynamic.py",
     "sov_ai_multilingual_parity": "src/usersim/engine/evaluator/scorers/sov_ai_multilingual_parity.py",
     "safety_chat_pressure": "src/usersim/engine/evaluator/scorers/safety_chat_pressure.py",
diff --git a/templates/probe/README.md b/templates/probe/README.md
index 57a3432..acbd380 100644
--- a/templates/probe/README.md
+++ b/templates/probe/README.md
@@ -175,7 +175,7 @@ When porting one of the templates into a new module:
       `_patched_call_llm` patches BOTH `core.simulation.acall_llm` AND
       `core.judges.acall_llm`, and the alias-dispatching mock returns
       XML-shaped judge verdicts. See
-      [`tests/engine/probes/test_sov_ai_facts.py`](../../tests/engine/probes/test_sov_ai_facts.py)
+      [`tests/probes/sov_ai_facts/test_probe.py`](../../tests/probes/sov_ai_facts/test_probe.py)
       for a real reference. The tests run under `usersim.testing.pytest_plugin`,
       so a model call nothing mocked fails with `BlockedNetworkCall`.
 - [ ] Update `docs/engine/README.md`'s probe table and the catalog in
diff --git a/templates/probe/test_probe.py.template b/templates/probe/test_probe.py.template
index c22c5b7..f5e69b4 100644
--- a/templates/probe/test_probe.py.template
+++ b/templates/probe/test_probe.py.template
@@ -9,7 +9,7 @@ with ``usersim.testing.run_episode`` and scripted models, with nothing
 patched; the 8 standard tests every probe should have are stubbed
 after it. For real working examples, see:
 
-- ``test_sov_ai_facts.py`` — banked-verbatim probe (the canonical
+- ``tests/probes/sov_ai_facts/test_probe.py`` — banked-verbatim probe (the canonical
   reference for the alias-dispatching mock + dual-module patching
   pattern).
 - ``test_safety_chat_pressure.py`` — bank-reframing probe
diff --git a/tests/engine/core/test_agentic_bank.py b/tests/engine/core/test_agentic_bank.py
index cb905a4..40151ea 100644
--- a/tests/engine/core/test_agentic_bank.py
+++ b/tests/engine/core/test_agentic_bank.py
@@ -4,7 +4,7 @@
 """Tests for ``core/agentic_bank.py`` — the agentic-action bank loader.
 
 Same three-layer test pattern as the other curated-asset bank loaders
-(``test_fact_bank.py`` / ``test_query_bank.py`` / ``test_pressure_bank.py``):
+(``sov_ai_facts/test_bank.py`` / ``test_query_bank.py`` / ``test_pressure_bank.py``):
 
 1. **Schema validation** — every required field, every cross-product
    invariant (sub_protocol membership, simulated_tools shape,
diff --git a/tests/engine/core/test_assets_discovery.py b/tests/engine/core/test_assets_discovery.py
index e954dd9..32224f0 100644
--- a/tests/engine/core/test_assets_discovery.py
+++ b/tests/engine/core/test_assets_discovery.py
@@ -273,7 +273,7 @@ class TestProbeAssetsDir:
     def test_a_built_in_probe_reads_its_packaged_folder(self) -> None:
         from usersim.engine.core.probes import resolve_probe
 
-        assert resolve_probe("sov_ai_facts").assets_dir() == packaged_assets_dir() / "sov_ai_facts"
+        assert resolve_probe("safety_chat_pressure").assets_dir() == packaged_assets_dir() / "safety_chat_pressure"
 
     def test_a_probe_reads_the_folder_under_its_declared_root(self, register_test_probe, tmp_path) -> None:
         (tmp_path / "ext_probe").mkdir()
diff --git a/tests/engine/core/test_default_asset_paths.py b/tests/engine/core/test_default_asset_paths.py
index 5add872..38066d0 100644
--- a/tests/engine/core/test_default_asset_paths.py
+++ b/tests/engine/core/test_default_asset_paths.py
@@ -58,11 +58,6 @@
     load_clinical_profile_bank,
     reset_clinical_profile_bank_cache,
 )
-from usersim.engine.core.fact_bank import (
-    default_fact_bank_path,
-    load_fact_bank,
-    reset_fact_bank_cache,
-)
 from usersim.engine.core.identity_spec import (
     default_identity_spec_path,
     load_identity_spec,
@@ -85,6 +80,11 @@
     reset_query_bank_cache,
 )
 from usersim.engine.core.seeds import load_seeds
+from usersim.engine.probes.sov_ai_facts.bank import (
+    default_fact_bank_path,
+    load_fact_bank,
+    reset_fact_bank_cache,
+)
 
 # ---------------------------------------------------------------------------
 # Per-locale loaders: fact_bank + probing_taxonomy
diff --git a/tests/engine/core/test_pressure_bank.py b/tests/engine/core/test_pressure_bank.py
index 68028fa..cd691b2 100644
--- a/tests/engine/core/test_pressure_bank.py
+++ b/tests/engine/core/test_pressure_bank.py
@@ -4,7 +4,7 @@
 """Tests for ``core/pressure_bank.py`` — pressure-strategy bank loader.
 
 Same three-layer test pattern as ``test_query_bank.py`` /
-``test_fact_bank.py`` / ``test_probing_taxonomy.py``:
+``sov_ai_facts/test_bank.py`` / ``test_probing_taxonomy.py``:
 
 1. **Schema validation** — every required field, every cross-product
    invariant (placeholder discipline, template placeholders), every
diff --git a/tests/engine/core/test_query_bank.py b/tests/engine/core/test_query_bank.py
index 88f812d..ae7dc1a 100644
--- a/tests/engine/core/test_query_bank.py
+++ b/tests/engine/core/test_query_bank.py
@@ -3,7 +3,7 @@
 
 """Tests for ``core/query_bank.py`` — cross-locale query bank loader.
 
-Three layers, mirroring the test patterns in ``test_fact_bank.py`` and
+Three layers, mirroring the test patterns in ``sov_ai_facts/test_bank.py`` and
 ``test_probing_taxonomy.py``:
 
 1. **Schema validation** — every required field, every cross-product
diff --git a/tests/engine/probes/test_india_variant_probes.py b/tests/engine/probes/test_india_variant_probes.py
index f30f826..38a4c11 100644
--- a/tests/engine/probes/test_india_variant_probes.py
+++ b/tests/engine/probes/test_india_variant_probes.py
@@ -92,7 +92,7 @@ def _mt_warned(ob: OutcomeBuilder) -> bool:
 
 
 async def test_sov_ai_facts_variant_translates_turn1(indian_persona, cfg) -> None:
-    from usersim.engine.probes.sov_ai_facts.generator import SovAiFactsProbe
+    from usersim.engine.probes.sov_ai_facts.probe import SovAiFactsProbe
 
     T.reset_translation_cache()
     ob = _builder()
diff --git a/tests/engine/probes/test_probe_invariants.py b/tests/engine/probes/test_probe_invariants.py
index 3898118..43c0434 100644
--- a/tests/engine/probes/test_probe_invariants.py
+++ b/tests/engine/probes/test_probe_invariants.py
@@ -39,7 +39,7 @@
 
 - Per-probe end-to-end behavior (verbatim turn-1 shape, two-turn
   follow-up, placeholder warnings, bank-load failures): the
-  per-probe test files (``test_sov_ai_facts.py``,
+  per-probe test files (``sov_ai_facts/test_probe.py``,
   ``test_safety_agentic.py``, etc.) own these.
 - Substrate behavior (mixins, ``BankBackedProbe``,
   ``LocalePromptPack``, ``@register_probe`` decorator semantics):
diff --git a/tests/engine/probes/test_sov_ai_dynamic.py b/tests/engine/probes/test_sov_ai_dynamic.py
index 3be0db4..c30ed76 100644
--- a/tests/engine/probes/test_sov_ai_dynamic.py
+++ b/tests/engine/probes/test_sov_ai_dynamic.py
@@ -3,7 +3,7 @@
 
 """Tests for the ``sov_ai_dynamic`` probe.
 
-Three layers (parallel to ``test_sov_ai_facts.py``):
+Three layers (parallel to ``sov_ai_facts/test_probe.py``):
 
 1. **Task-derivation primitives** (``task_derivation.py``) — pure
    functions that pick a category + sub-topic hint deterministically
diff --git a/tests/engine/probes/test_sov_ai_multilingual_parity.py b/tests/engine/probes/test_sov_ai_multilingual_parity.py
index 09093e3..fae4cc9 100644
--- a/tests/engine/probes/test_sov_ai_multilingual_parity.py
+++ b/tests/engine/probes/test_sov_ai_multilingual_parity.py
@@ -3,7 +3,7 @@
 
 """Tests for the ``sov_ai_multilingual_parity`` probe.
 
-Layered the same way as ``test_sov_ai_facts.py`` and
+Layered the same way as ``sov_ai_facts/test_probe.py`` and
 ``test_sov_ai_dynamic.py``:
 
 1. **Task-derivation primitives** — ``persona_to_tags`` (re-exported
diff --git a/tests/probes/__init__.py b/tests/probes/__init__.py
new file mode 100644
index 0000000..52a7a9d
--- /dev/null
+++ b/tests/probes/__init__.py
@@ -0,0 +1,2 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
diff --git a/tests/probes/sov_ai_facts/__init__.py b/tests/probes/sov_ai_facts/__init__.py
new file mode 100644
index 0000000..52a7a9d
--- /dev/null
+++ b/tests/probes/sov_ai_facts/__init__.py
@@ -0,0 +1,2 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
diff --git a/tests/engine/core/test_fact_bank.py b/tests/probes/sov_ai_facts/test_bank.py
similarity index 98%
rename from tests/engine/core/test_fact_bank.py
rename to tests/probes/sov_ai_facts/test_bank.py
index 2ade072..0e86373 100644
--- a/tests/engine/core/test_fact_bank.py
+++ b/tests/probes/sov_ai_facts/test_bank.py
@@ -1,7 +1,7 @@
 # SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 
-"""Tests for ``usersim.engine.core.fact_bank``.
+"""Tests for ``usersim.engine.probes.sov_ai_facts.bank``.
 
 The loader is the entry-point that subject-matter reviewers' YAML
 files pass through. Every schema invariant is exercised here so a
@@ -15,8 +15,8 @@
 
 import pytest
 
-from usersim.engine.core._assets import packaged_assets_dir
-from usersim.engine.core.fact_bank import (
+from usersim.engine.core._assets import probe_assets_dir
+from usersim.engine.probes.sov_ai_facts.bank import (
     ALLOWED_DIFFICULTIES,
     ALLOWED_QUESTION_TYPES,
     SUPPORTED_SCHEMA_VERSIONS,
@@ -521,7 +521,7 @@ class TestPtBrSampleBank:
 
     @pytest.fixture
     def bank(self) -> FactBank:
-        return load_fact_bank((packaged_assets_dir() / "sov_ai_facts/pt_BR/sample.yaml"))
+        return load_fact_bank((probe_assets_dir("sov_ai_facts") / "pt_BR/sample.yaml"))
 
     def test_loads(self, bank: FactBank) -> None:
         assert bank.locale == "pt_BR"
diff --git a/tests/engine/probes/test_sov_ai_facts.py b/tests/probes/sov_ai_facts/test_probe.py
similarity index 97%
rename from tests/engine/probes/test_sov_ai_facts.py
rename to tests/probes/sov_ai_facts/test_probe.py
index 349bc36..be3d355 100644
--- a/tests/engine/probes/test_sov_ai_facts.py
+++ b/tests/probes/sov_ai_facts/test_probe.py
@@ -28,13 +28,7 @@
 import pytest
 
 import usersim.engine.generator  # noqa: F401 — triggers probe registration
-from usersim.engine.core._assets import packaged_assets_dir
-from usersim.engine.core.fact_bank import (
-    Fact,
-    FactBank,
-    FactProvenance,
-    load_fact_bank,
-)
+from usersim.engine.core._assets import probe_assets_dir
 from usersim.engine.core.locale import SHIPPED_LOCALES
 from usersim.engine.core.outcomes import (
     FailureClass,
@@ -42,12 +36,18 @@
 )
 from usersim.engine.core.probes import _PROBE_REGISTRY, resolve_probe
 from usersim.engine.probes.sov_ai_facts import (
-    generator as probe_gen,
+    probe as probe_gen,
 )
 from usersim.engine.probes.sov_ai_facts import (
     prompts,
     task_derivation,
 )
+from usersim.engine.probes.sov_ai_facts.bank import (
+    Fact,
+    FactBank,
+    FactProvenance,
+    load_fact_bank,
+)
 
 # ---------------------------------------------------------------------------
 # Fixtures
@@ -83,7 +83,7 @@ def pt_br_persona_northeast_retiree() -> dict[str, Any]:
 
 @pytest.fixture
 def pt_br_sample_bank() -> FactBank:
-    return load_fact_bank((packaged_assets_dir() / "sov_ai_facts/pt_BR/sample.yaml"))
+    return load_fact_bank((probe_assets_dir("sov_ai_facts") / "pt_BR/sample.yaml"))
 
 
 @pytest.fixture(autouse=True)
@@ -177,7 +177,7 @@ def test_uk_nations_are_only_derived_for_en_gb(self) -> None:
         [("London", "england"), ("Scotland", "scotland"), ("Wales", "wales"), ("Northern Ireland", "northern-ireland")],
     )
     def test_the_uk_bank_serves_each_nation_its_own_facts(self, region: str, nation: str) -> None:
-        bank = load_fact_bank(packaged_assets_dir() / "sov_ai_facts/en_GB/sample.yaml")
+        bank = load_fact_bank(probe_assets_dir("sov_ai_facts") / "en_GB/sample.yaml")
         tags = task_derivation.persona_to_tags({"region": region, "age": 30}, "en_GB")
         own = {f.id for f in bank.by_tag(f"region:{nation}")}
         assert len(own) >= 8
@@ -228,7 +228,7 @@ def test_laender_are_only_derived_for_de_de(self) -> None:
         ],
     )
     def test_the_german_bank_serves_each_land_its_own_facts(self, region: str) -> None:
-        bank = load_fact_bank(packaged_assets_dir() / "sov_ai_facts/de_DE/sample.yaml")
+        bank = load_fact_bank(probe_assets_dir("sov_ai_facts") / "de_DE/sample.yaml")
         tags = task_derivation.persona_to_tags({"region": region, "age": 30}, "de_DE")
         (land,) = [t for t in tags if t.startswith("region:") and t != "region:any"]
         own = {f.id for f in bank.by_tag(land)}
@@ -559,7 +559,7 @@ async def test_single_turn_runs_end_to_end(
         with (
             patch.dict(
                 os.environ,
-                {"USERSIM_SOV_AI_FACTS_BANK_PT_BR": str((packaged_assets_dir() / "sov_ai_facts/pt_BR/sample.yaml"))},
+                {"USERSIM_SOV_AI_FACTS_BANK_PT_BR": str((probe_assets_dir("sov_ai_facts") / "pt_BR/sample.yaml"))},
             ),
             _patched_call_llm(
                 _mock_call_llm(
@@ -619,7 +619,7 @@ async def test_verbatim_question_is_first_user_turn(
         with (
             patch.dict(
                 os.environ,
-                {"USERSIM_SOV_AI_FACTS_BANK_PT_BR": str((packaged_assets_dir() / "sov_ai_facts/pt_BR/sample.yaml"))},
+                {"USERSIM_SOV_AI_FACTS_BANK_PT_BR": str((probe_assets_dir("sov_ai_facts") / "pt_BR/sample.yaml"))},
             ),
             _patched_call_llm(
                 _mock_call_llm(
@@ -667,7 +667,7 @@ async def test_two_turn_mode_issues_followup(
         with (
             patch.dict(
                 os.environ,
-                {"USERSIM_SOV_AI_FACTS_BANK_PT_BR": str((packaged_assets_dir() / "sov_ai_facts/pt_BR/sample.yaml"))},
+                {"USERSIM_SOV_AI_FACTS_BANK_PT_BR": str((probe_assets_dir("sov_ai_facts") / "pt_BR/sample.yaml"))},
             ),
             _patched_call_llm(
                 _mock_call_llm(
@@ -714,7 +714,7 @@ async def test_assistant_failure_produces_failed_outcome(
         with (
             patch.dict(
                 os.environ,
-                {"USERSIM_SOV_AI_FACTS_BANK_PT_BR": str((packaged_assets_dir() / "sov_ai_facts/pt_BR/sample.yaml"))},
+                {"USERSIM_SOV_AI_FACTS_BANK_PT_BR": str((probe_assets_dir("sov_ai_facts") / "pt_BR/sample.yaml"))},
             ),
             _patched_call_llm(RuntimeError("rate-limited by provider")),
         ):
diff --git a/tests/engine/evaluator/test_evaluator_scorer_sov_ai_facts.py b/tests/probes/sov_ai_facts/test_scorer.py
similarity index 93%
rename from tests/engine/evaluator/test_evaluator_scorer_sov_ai_facts.py
rename to tests/probes/sov_ai_facts/test_scorer.py
index f5ac006..a173f01 100644
--- a/tests/engine/evaluator/test_evaluator_scorer_sov_ai_facts.py
+++ b/tests/probes/sov_ai_facts/test_scorer.py
@@ -23,15 +23,15 @@
 
 import pytest
 
-from usersim.engine.core._assets import packaged_assets_dir
-from usersim.engine.core.fact_bank import reset_fact_bank_cache
+from usersim.engine.core._assets import probe_assets_dir
 from usersim.engine.evaluator.scorers import (
     clear_registry,
     get_scorer,
     list_scorers,
     load_default_scorers,
 )
-from usersim.engine.evaluator.scorers.sov_ai_facts import (
+from usersim.engine.probes.sov_ai_facts.bank import reset_fact_bank_cache
+from usersim.engine.probes.sov_ai_facts.scorer import (
     _ACCURACY,
     _COMPLETENESS,
     _COMPLETION_AXES,
@@ -66,7 +66,7 @@ def _reset_state_between_tests():
 
 @pytest.fixture
 def sample_bank_path() -> Path:
-    return packaged_assets_dir() / "sov_ai_facts/pt_BR/sample.yaml"
+    return probe_assets_dir("sov_ai_facts") / "pt_BR/sample.yaml"
 
 
 @pytest.fixture
@@ -255,7 +255,7 @@ def test_fabrication_is_reversed_scale(self) -> None:
 class TestShortCircuit:
     async def test_missing_facts_probed_returns_empty_envelope(self) -> None:
         # No LLM call should fire.
-        with patch("usersim.engine.evaluator.scorers.sov_ai_facts.acall_llm") as mock:
+        with patch("usersim.engine.probes.sov_ai_facts.scorer.acall_llm") as mock:
             result = await score_sov_ai_facts_trajectory(
                 trajectory={"locale": "pt_BR", "conversation_messages": "[]"},
                 models={"judge_model": object()},
@@ -267,7 +267,7 @@ async def test_missing_facts_probed_returns_empty_envelope(self) -> None:
         assert "no sovereign_facts_probed" in result["error"]
 
     async def test_empty_facts_probed_list_short_circuits(self) -> None:
-        with patch("usersim.engine.evaluator.scorers.sov_ai_facts.acall_llm") as mock:
+        with patch("usersim.engine.probes.sov_ai_facts.scorer.acall_llm") as mock:
             result = await score_sov_ai_facts_trajectory(
                 trajectory={
                     "sovereign_facts_probed": [],
@@ -280,7 +280,7 @@ async def test_empty_facts_probed_list_short_circuits(self) -> None:
         assert result["error"].startswith("no sovereign_facts_probed")
 
     async def test_missing_locale_short_circuits(self) -> None:
-        with patch("usersim.engine.evaluator.scorers.sov_ai_facts.acall_llm") as mock:
+        with patch("usersim.engine.probes.sov_ai_facts.scorer.acall_llm") as mock:
             result = await score_sov_ai_facts_trajectory(
                 trajectory={
                     "sovereign_facts_probed": ["BR-GEO-001"],
@@ -307,7 +307,7 @@ async def test_full_envelope_returned(self, mock_bank_env: Path) -> None:
             false_premise_catch=5,
         )
         with patch(
-            "usersim.engine.evaluator.scorers.sov_ai_facts.acall_llm",
+            "usersim.engine.probes.sov_ai_facts.scorer.acall_llm",
             side_effect=_mock_judge_call([payload]),
         ):
             result = await score_sov_ai_facts_trajectory(traj, models={"judge_model": object()})
@@ -336,7 +336,7 @@ async def test_aggregate_shared_axes_only(self, mock_bank_env: Path) -> None:
             false_premise_catch=3,
         )
         with patch(
-            "usersim.engine.evaluator.scorers.sov_ai_facts.acall_llm",
+            "usersim.engine.probes.sov_ai_facts.scorer.acall_llm",
             side_effect=_mock_judge_call([payload]),
         ):
             result = await score_sov_ai_facts_trajectory(traj, models={"judge_model": object()})
@@ -361,7 +361,7 @@ async def test_status_proposal_falls_on_accuracy_1(self, mock_bank_env: Path) ->
             false_premise_catch=1,
         )
         with patch(
-            "usersim.engine.evaluator.scorers.sov_ai_facts.acall_llm",
+            "usersim.engine.probes.sov_ai_facts.scorer.acall_llm",
             side_effect=_mock_judge_call([payload]),
         ):
             result = await score_sov_ai_facts_trajectory(traj, models={"judge_model": object()})
@@ -376,7 +376,7 @@ async def test_status_proposal_falls_on_fabrication_1(self, mock_bank_env: Path)
             false_premise_catch=5,
         )
         with patch(
-            "usersim.engine.evaluator.scorers.sov_ai_facts.acall_llm",
+            "usersim.engine.probes.sov_ai_facts.scorer.acall_llm",
             side_effect=_mock_judge_call([payload]),
         ):
             result = await score_sov_ai_facts_trajectory(traj, models={"judge_model": object()})
@@ -399,7 +399,7 @@ async def test_graceful_refusal_scores_well(self, mock_bank_env: Path) -> None:
             question_type="completion",
         )
         with patch(
-            "usersim.engine.evaluator.scorers.sov_ai_facts.acall_llm",
+            "usersim.engine.probes.sov_ai_facts.scorer.acall_llm",
             side_effect=_mock_judge_call([payload]),
         ):
             result = await score_sov_ai_facts_trajectory(traj, models={"judge_model": object()})
@@ -425,7 +425,7 @@ async def test_fabricated_continuation_flags_fabrication(self, mock_bank_env: Pa
             question_type="completion",
         )
         with patch(
-            "usersim.engine.evaluator.scorers.sov_ai_facts.acall_llm",
+            "usersim.engine.probes.sov_ai_facts.scorer.acall_llm",
             side_effect=_mock_judge_call([payload]),
         ):
             result = await score_sov_ai_facts_trajectory(traj, models={"judge_model": object()})
@@ -441,7 +441,7 @@ async def test_fabricated_continuation_flags_fabrication(self, mock_bank_env: Pa
 class TestErrorPaths:
     async def test_unknown_fact_id_recorded_not_fatal(self, mock_bank_env: Path) -> None:
         traj = _factual_recall_trajectory(fact_id="BR-GEO-XXX-nonexistent")
-        with patch("usersim.engine.evaluator.scorers.sov_ai_facts.acall_llm") as mock:
+        with patch("usersim.engine.probes.sov_ai_facts.scorer.acall_llm") as mock:
             result = await score_sov_ai_facts_trajectory(traj, models={"judge_model": object()})
         # No LLM call fires because the fact isn't in the bank.
         mock.assert_not_called()
@@ -458,7 +458,7 @@ async def test_bank_load_failure_short_circuits(self, tmp_path: Path) -> None:
         missing = tmp_path / "does-not-exist.yaml"
         with (
             patch.dict(os.environ, {"USERSIM_SOV_AI_FACTS_BANK_PT_BR": str(missing)}),
-            patch("usersim.engine.evaluator.scorers.sov_ai_facts.acall_llm") as mock,
+            patch("usersim.engine.probes.sov_ai_facts.scorer.acall_llm") as mock,
         ):
             traj = _factual_recall_trajectory()
             result = await score_sov_ai_facts_trajectory(traj, models={"judge_model": object()})
@@ -476,7 +476,7 @@ async def test_bank_version_mismatch_flagged(self, mock_bank_env: Path) -> None:
             false_premise_catch=5,
         )
         with patch(
-            "usersim.engine.evaluator.scorers.sov_ai_facts.acall_llm",
+            "usersim.engine.probes.sov_ai_facts.scorer.acall_llm",
             side_effect=_mock_judge_call([payload]),
         ):
             result = await score_sov_ai_facts_trajectory(traj, models={"judge_model": object()})
@@ -490,7 +490,7 @@ async def test_bank_version_mismatch_flagged(self, mock_bank_env: Path) -> None:
     async def test_judge_exception_becomes_per_fact_error(self, mock_bank_env: Path) -> None:
         traj = _factual_recall_trajectory()
         with patch(
-            "usersim.engine.evaluator.scorers.sov_ai_facts.acall_llm",
+            "usersim.engine.probes.sov_ai_facts.scorer.acall_llm",
             side_effect=RuntimeError("rate-limited"),
         ):
             result = await score_sov_ai_facts_trajectory(traj, models={"judge_model": object()})
@@ -509,7 +509,7 @@ async def _return_garbage(*_args, **_kwargs):
             return {"role": "assistant", "content": "not valid json"}
 
         with patch(
-            "usersim.engine.evaluator.scorers.sov_ai_facts.acall_llm",
+            "usersim.engine.probes.sov_ai_facts.scorer.acall_llm",
             side_effect=_return_garbage,
         ):
             result = await score_sov_ai_facts_trajectory(traj, models={"judge_model": object()})
@@ -535,7 +535,7 @@ async def test_two_facts_aggregate_correctly(self, mock_bank_env: Path) -> None:
             _axis_payload(accuracy=3, completeness=2, fabrication=3, false_premise_catch=3),
         ]
         with patch(
-            "usersim.engine.evaluator.scorers.sov_ai_facts.acall_llm",
+            "usersim.engine.probes.sov_ai_facts.scorer.acall_llm",
             side_effect=_mock_judge_call(payloads),
         ):
             result = await score_sov_ai_facts_trajectory(traj, models={"judge_model": object()})
diff --git a/tests/test_capabilities.py b/tests/test_capabilities.py
index 3f30601..5a18119 100644
--- a/tests/test_capabilities.py
+++ b/tests/test_capabilities.py
@@ -170,7 +170,7 @@ def test_all_probe_capabilities_are_declared_with_sentinel() -> None:
 
 def test_capabilities_expose_implementation_refs() -> None:
     assert any(
-        ref.endswith("evaluator/scorers/sov_ai_facts.py")
+        ref.endswith("probes/sov_ai_facts/scorer.py")
         for ref in implementation_refs_for_capability(capability_by_id("sovereign_local_knowledge"))
     )
     assert any(
diff --git a/tests/test_probe_names_in_shared_code.py b/tests/test_probe_names_in_shared_code.py
index 9336640..a075b2f 100644
--- a/tests/test_probe_names_in_shared_code.py
+++ b/tests/test_probe_names_in_shared_code.py
@@ -35,7 +35,6 @@ class instead (``architecture/probes.md``, "What a probe declares").
     "engine/core/action_taxonomy.py": 1,
     "engine/core/agentic_bank.py": 2,
     "engine/core/clinical_profile_bank.py": 4,
-    "engine/core/fact_bank.py": 1,
     "engine/core/finance_bank.py": 1,
     "engine/core/identity_spec.py": 1,
     "engine/core/pressure_bank.py": 1,
@@ -48,7 +47,6 @@ class instead (``architecture/probes.md``, "What a probe declares").
     "engine/evaluator/scorers/safety_agentic.py": 1,
     "engine/evaluator/scorers/safety_chat_pressure.py": 1,
     "engine/evaluator/scorers/sov_ai_dynamic.py": 1,
-    "engine/evaluator/scorers/sov_ai_facts.py": 1,
     "engine/evaluator/scorers/sov_ai_multilingual_parity.py": 1,
     "reporting/capability_report.py": 1,
     "reporting/identity.py": 1,
diff --git a/tests/test_toolset_seed_cli.py b/tests/test_toolset_seed_cli.py
index 2bb3661..9f47d56 100644
--- a/tests/test_toolset_seed_cli.py
+++ b/tests/test_toolset_seed_cli.py
@@ -27,7 +27,7 @@
     TOOLSET_SEED_SUFFIXES,
     resolve_toolset_seed_path,
 )
-from usersim.engine.core._assets import packaged_assets_dir
+from usersim.engine.core._assets import packaged_assets_dir, probe_assets_dir
 from usersim.engine.probes.tool_calling.generator import _normalize_tool_list
 
 
@@ -41,7 +41,7 @@ def _assets_root(tmp_path: Path, *seed_names: str) -> Path:
     (probe_dir / "prompts.yaml").write_text("prompts: []\n", encoding="utf-8")
     facts = tmp_path / "sov_ai_facts" / "en_US"
     facts.mkdir(parents=True)
-    shutil.copy(packaged_assets_dir() / "sov_ai_facts" / "en_US" / "sample.yaml", facts / "sample.yaml")
+    shutil.copy(probe_assets_dir("sov_ai_facts") / "en_US" / "sample.yaml", facts / "sample.yaml")
     for name in seed_names:
         (probe_dir / name).touch()
     return tmp_path

From b59e49f9c7ce4f69dc52f715495cbb6155bb54f0 Mon Sep 17 00:00:00 2001
From: Yev Meyer <ymeyer@nvidia.com>
Date: Mon, 2 Nov 2026 20:46:16 -0500
Subject: [PATCH 3/5] refactor(engine): keep sov_ai_dynamic in one folder

The probe module is probes/sov_ai_dynamic/probe.py, beside its scorer, its banks in assets/sov_ai_dynamic/, and taxonomy.py, which names the probe's taxonomy files: default_probing_taxonomy_path, probing_taxonomy_path_for and load_probing_taxonomy_for_locale. The taxonomy format, its validator, select_category and the cache stay in core/probing_taxonomy.py, which financial_services also uses and which names no probe.

Its tests are in tests/probes/sov_ai_dynamic/, including those of its taxonomy files and of every shipped taxonomy; the format's tests stay in tests/engine/core/test_probing_taxonomy.py. The label, the scorer name, the banks and the stored columns are unchanged.

Signed-off-by: Yev Meyer <ymeyer@nvidia.com>
---
 docs/engine/README.md                         |   5 +-
 src/usersim/engine/core/probes.py             |   2 +-
 src/usersim/engine/core/probing_taxonomy.py   |  48 +---
 .../engine/evaluator/scorers/__init__.py      |   2 +-
 .../engine/probes/sov_ai_dynamic/__init__.py  |   2 +-
 .../assets/sov_ai_dynamic/SCHEMA.md           |   0
 .../sov_ai_dynamic/de_DE/categories.yaml      |   0
 .../sov_ai_dynamic/en_GB/categories.yaml      |   0
 .../sov_ai_dynamic/en_IN/categories.yaml      |   0
 .../sov_ai_dynamic/en_SG/categories.yaml      |   0
 .../sov_ai_dynamic/en_US/categories.yaml      |   0
 .../sov_ai_dynamic/fr_FR/categories.yaml      |   0
 .../sov_ai_dynamic/hi_Deva_IN/categories.yaml |   0
 .../sov_ai_dynamic/hi_Latn_IN/categories.yaml |   0
 .../sov_ai_dynamic/ja_JP/categories.yaml      |   0
 .../sov_ai_dynamic/ko_KR/categories.yaml      |   0
 .../sov_ai_dynamic/pt_BR/categories.yaml      |   0
 .../sov_ai_dynamic/{generator.py => probe.py} |  13 +-
 .../sov_ai_dynamic/scorer.py}                 |  18 +-
 .../engine/probes/sov_ai_dynamic/taxonomy.py  |  35 +++
 src/usersim/taxonomy/capabilities.py          |   4 +-
 tests/engine/core/test_default_asset_paths.py |   7 +-
 tests/engine/core/test_probing_taxonomy.py    | 195 --------------
 .../probes/test_india_variant_probes.py       |   6 +-
 .../probes/test_sov_ai_multilingual_parity.py |   2 +-
 tests/probes/sov_ai_dynamic/__init__.py       |   2 +
 .../sov_ai_dynamic/test_probe.py}             |   8 +-
 .../sov_ai_dynamic/test_scorer.py}            |  24 +-
 tests/probes/sov_ai_dynamic/test_taxonomy.py  | 246 ++++++++++++++++++
 tests/test_probe_names_in_shared_code.py      |   2 -
 30 files changed, 338 insertions(+), 283 deletions(-)
 rename src/usersim/engine/{ => probes/sov_ai_dynamic}/assets/sov_ai_dynamic/SCHEMA.md (100%)
 rename src/usersim/engine/{ => probes/sov_ai_dynamic}/assets/sov_ai_dynamic/de_DE/categories.yaml (100%)
 rename src/usersim/engine/{ => probes/sov_ai_dynamic}/assets/sov_ai_dynamic/en_GB/categories.yaml (100%)
 rename src/usersim/engine/{ => probes/sov_ai_dynamic}/assets/sov_ai_dynamic/en_IN/categories.yaml (100%)
 rename src/usersim/engine/{ => probes/sov_ai_dynamic}/assets/sov_ai_dynamic/en_SG/categories.yaml (100%)
 rename src/usersim/engine/{ => probes/sov_ai_dynamic}/assets/sov_ai_dynamic/en_US/categories.yaml (100%)
 rename src/usersim/engine/{ => probes/sov_ai_dynamic}/assets/sov_ai_dynamic/fr_FR/categories.yaml (100%)
 rename src/usersim/engine/{ => probes/sov_ai_dynamic}/assets/sov_ai_dynamic/hi_Deva_IN/categories.yaml (100%)
 rename src/usersim/engine/{ => probes/sov_ai_dynamic}/assets/sov_ai_dynamic/hi_Latn_IN/categories.yaml (100%)
 rename src/usersim/engine/{ => probes/sov_ai_dynamic}/assets/sov_ai_dynamic/ja_JP/categories.yaml (100%)
 rename src/usersim/engine/{ => probes/sov_ai_dynamic}/assets/sov_ai_dynamic/ko_KR/categories.yaml (100%)
 rename src/usersim/engine/{ => probes/sov_ai_dynamic}/assets/sov_ai_dynamic/pt_BR/categories.yaml (100%)
 rename src/usersim/engine/probes/sov_ai_dynamic/{generator.py => probe.py} (98%)
 rename src/usersim/engine/{evaluator/scorers/sov_ai_dynamic.py => probes/sov_ai_dynamic/scorer.py} (97%)
 create mode 100644 src/usersim/engine/probes/sov_ai_dynamic/taxonomy.py
 create mode 100644 tests/probes/sov_ai_dynamic/__init__.py
 rename tests/{engine/probes/test_sov_ai_dynamic.py => probes/sov_ai_dynamic/test_probe.py} (99%)
 rename tests/{engine/evaluator/test_evaluator_scorer_sov_ai_dynamic.py => probes/sov_ai_dynamic/test_scorer.py} (95%)
 create mode 100644 tests/probes/sov_ai_dynamic/test_taxonomy.py

diff --git a/docs/engine/README.md b/docs/engine/README.md
index e44ca6e..61b8d39 100644
--- a/docs/engine/README.md
+++ b/docs/engine/README.md
@@ -611,7 +611,6 @@ plugin entry points, undeclared result columns). The tables describe the main fi
 | `test_probe_invariants.py` | Cross-probe invariants every shipped probe upholds (registration, side-channel keys, prompt-version monotonicity, etc.) |
 | `test_tool_calling.py` | Tool-calling probe: tool sampling, gate check |
 | `test_verifiers.py` | Tool-call JSON-schema validator |
-| `test_sov_ai_dynamic.py` | `sov_ai_dynamic` probe: persona→tag, deterministic `derive_probe` + hint rotation (taxonomy-version salted), all-locale prompt availability, e2e with mocked LLM |
 | `test_sov_ai_multilingual_parity.py` | `sov_ai_multilingual_parity` probe: persona→tag re-export, deterministic `derive_task` with locale opt-out filtering (bank-version salted), all-locale prompt completeness with no-half-shipped invariant, query-bank cache + env override, e2e with mocked LLM, verbatim turn-1 injection, structured failure paths |
 | `test_safety_chat_pressure.py` | `safety_chat_pressure` probe: persona→tag re-export, `derive_task` with independent strategy/target salts (50-seed cell-coverage test), `resolve_task_from_row` (full / partial / neither override + persona-tag guard against panel-driven mismatches), prompt-formatting helpers, end-to-end with mocked LLM (verbatim turn-1, multi-turn rotation, bank-load failure, panel-id miss, no-resolvable-pair, assistant-error attribution, follow-up-failure breaks loop) |
 | `test_safety_chat_pressure_classifier.py` | Dedicated capitulation classifier (comparator-shaped `summary_model` call): rubric-anchored decision, defensive returns, regression that the comparator prompt has no pressure / strategy / persona keywords (post the v1.2 redesign) |
@@ -631,7 +630,6 @@ plugin entry points, undeclared result columns). The tables describe the main fi
 | `test_evaluator_scorers.py` | Scorer registry (`register_scorer` / `get_scorer` / `clear_registry` / `reload_scorers`) |
 | `test_evaluator_scorer_structured_output_schemas.py` | Cross-cutting schema-design contract for LLM-judge scorers (Literal-typed enums, Optional fields, 1–5 score constraints): fails the build on violations |
 | `test_evaluator_scorer_tool_use.py` | 8-axis tool-calling trajectory judge |
-| `test_evaluator_scorer_sov_ai_dynamic.py` | `sov_ai_dynamic` out-of-sim scorer: response-only LLM judge with `sovereign.{suspected_fabrication,self_consistency,graceful_unknown}` axes, multi-category aggregation, drift detection |
 | `test_evaluator_scorer_sov_ai_multilingual_parity.py` | `sov_ai_multilingual_parity` out-of-sim respect-parity scorer: single LLM judge per trajectory with `respect.{condescension,complexity_calibration,helpfulness_at_proficiency}` axes, concern-grounded rubric, locale+language interpolation, status-proposal gating, error paths, version-mismatch drift |
 | `test_evaluator_scorer_safety_chat_pressure.py` | `safety_chat_pressure` out-of-sim scorer: single LLM judge per trajectory with `pressure.{refusal_held,contextual_safety_persistence}` scored axes plus per-turn classifications + derived `turn_of_flip`, rubric grounded in `target.expected_refusal` + `strategy.description`, status-proposal gating, error paths, version-mismatch drift |
 | `test_evaluator_scorer_safety_agentic.py` | `safety_agentic` scorer: registration, axis constants (`agentic.*` namespace; mechanical vs LLM separation), short-circuit paths, load failures, mechanical-only sub-protocols, attempted-actions normalization, attempted actions read from `conversation_metadata` on a stored row, supporting tools staying inside the boundary, `consequence_disclosure` and `compromised_inputs` judge paths, judge-error paths, bank-version drift, total-blast-radius aggregation, envelope contract |
@@ -648,3 +646,6 @@ plugin entry points, undeclared result columns). The tables describe the main fi
 | `sov_ai_facts/test_probe.py` | `sov_ai_facts` probe: persona→tag, deterministic `derive_task` (bank-version salted), prompts, e2e with mocked LLM |
 | `sov_ai_facts/test_scorer.py` | `sov_ai_facts` out-of-sim scorer: per-fact LLM judge with question-type-specific axes, error paths, multi-fact aggregation |
 | `sov_ai_facts/test_bank.py` | Fact-bank YAML schema validator, loader, persona-tag matching, locale-keyed shared cache |
+| `sov_ai_dynamic/test_probe.py` | `sov_ai_dynamic` probe: persona→tag, deterministic `derive_probe` + hint rotation (taxonomy-version salted), all-locale prompt availability, e2e with mocked LLM |
+| `sov_ai_dynamic/test_scorer.py` | `sov_ai_dynamic` out-of-sim scorer: response-only LLM judge with `sovereign.{suspected_fabrication,self_consistency,graceful_unknown}` axes, multi-category aggregation, drift detection |
+| `sov_ai_dynamic/test_taxonomy.py` | The probe's taxonomy paths and cached loader, and every shipped per-locale taxonomy: shape, canonical topic set, placeholder flags, language |
diff --git a/src/usersim/engine/core/probes.py b/src/usersim/engine/core/probes.py
index ed3fe7d..be5af1c 100644
--- a/src/usersim/engine/core/probes.py
+++ b/src/usersim/engine/core/probes.py
@@ -1550,7 +1550,7 @@ def decorator(cls: type[BaseProbe]) -> type[BaseProbe]:
     "usersim.engine.probes.identity_disclosure.generator",
     "usersim.engine.probes.safety_agentic.generator",
     "usersim.engine.probes.safety_chat_pressure.generator",
-    "usersim.engine.probes.sov_ai_dynamic.generator",
+    "usersim.engine.probes.sov_ai_dynamic.probe",
     "usersim.engine.probes.sov_ai_facts.probe",
     "usersim.engine.probes.sov_ai_multilingual_parity.generator",
     "usersim.engine.probes.tool_calling.generator",
diff --git a/src/usersim/engine/core/probing_taxonomy.py b/src/usersim/engine/core/probing_taxonomy.py
index 6704480..3d437c4 100644
--- a/src/usersim/engine/core/probing_taxonomy.py
+++ b/src/usersim/engine/core/probing_taxonomy.py
@@ -1,27 +1,29 @@
 # SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 
-"""Probing-taxonomy loader + validator + lookup helpers for ``sov_ai_dynamic``.
+"""Probing-taxonomy loader + validator + lookup helpers.
 
-A probing taxonomy is a curated, per-locale YAML file used by the
-``sov_ai_dynamic`` probe to seed the user-agent's opening turn with a
-category-bound, persona-driven probe. It carries **no ground truth** —
+A probing taxonomy is a curated, per-locale YAML file used to seed the
+user-agent's opening turn with a category-bound, persona-driven probe,
+by the ``sov_ai_dynamic`` probe and by ``financial_services``'s dynamic
+tier. It carries **no ground truth** —
 that's the whole point of dynamic probing — just per-category invitation
 prose and sub-topic hints used for collapse avoidance via per-trajectory
 rotation.
 
-Schema specified in ``assets/sov_ai_dynamic/SCHEMA.md``. This module
-exposes:
+Schema specified in the ``sov_ai_dynamic`` probe's
+``assets/sov_ai_dynamic/SCHEMA.md``. This module exposes:
 
 - :class:`Category` — typed view of a single category entry.
 - :class:`ProbingTaxonomy` — typed view of the whole taxonomy + lookup
   helpers.
 - :class:`ProbingTaxonomyError` — raised for schema violations.
 - :func:`load_probing_taxonomy` — parse + validate a single file.
-- :func:`load_probing_taxonomy_for_locale` / :func:`reset_probing_taxonomy_cache`
-  — locale-keyed shared cache so the probe and the ``sov_ai_dynamic``
-  scorer agree on ``taxonomy_version`` from one source of truth
-  (same pattern as ``probes.sov_ai_facts.bank``).
+- :func:`load_probing_taxonomy_for` / :func:`reset_probing_taxonomy_cache`
+  — a cache keyed by family and locale, so a probe and its scorer agree
+  on ``taxonomy_version`` from one source of truth (same pattern as
+  ``probes.sov_ai_facts.bank``). ``probes.sov_ai_dynamic.taxonomy`` names
+  that probe's files in it.
 """
 
 from __future__ import annotations
@@ -430,32 +432,6 @@ def load_probing_taxonomy_for(
         return taxonomy
 
 
-# ── sov_ai_dynamic back-compat shims (family = "sov_ai_dynamic") ──────────
-
-
-def default_probing_taxonomy_path(locale: str) -> Path:
-    return taxonomy_path_for(
-        "sov_ai_dynamic",
-        locale,
-        filename="categories.yaml",
-        env_prefix="USERSIM_SOV_AI_DYNAMIC_TAXONOMY",
-    )
-
-
-def probing_taxonomy_path_for(locale: str) -> Path:
-    return default_probing_taxonomy_path(locale)
-
-
-def load_probing_taxonomy_for_locale(locale: str) -> ProbingTaxonomy:
-    """Cached sov_ai_dynamic taxonomy for ``locale`` (back-compat wrapper)."""
-    return load_probing_taxonomy_for(
-        "sov_ai_dynamic",
-        locale,
-        filename="categories.yaml",
-        env_prefix="USERSIM_SOV_AI_DYNAMIC_TAXONOMY",
-    )
-
-
 def reset_probing_taxonomy_cache() -> None:
     """Drop all cached taxonomies. For tests, CLI reloads, version bumps."""
     with _TAXONOMY_CACHE_LOCK:
diff --git a/src/usersim/engine/evaluator/scorers/__init__.py b/src/usersim/engine/evaluator/scorers/__init__.py
index edae6fc..b44592b 100644
--- a/src/usersim/engine/evaluator/scorers/__init__.py
+++ b/src/usersim/engine/evaluator/scorers/__init__.py
@@ -315,7 +315,7 @@ def reload_scorers(scorer_module_paths: Iterable[str]) -> None:
 DEFAULT_SCORER_MODULES: tuple[str, ...] = (
     "usersim.engine.evaluator.scorers.tool_use",
     "usersim.engine.probes.sov_ai_facts.scorer",
-    "usersim.engine.evaluator.scorers.sov_ai_dynamic",
+    "usersim.engine.probes.sov_ai_dynamic.scorer",
     "usersim.engine.evaluator.scorers.sov_ai_multilingual_parity",
     "usersim.engine.evaluator.scorers.safety_chat_pressure",
     "usersim.engine.evaluator.scorers.safety_agentic",
diff --git a/src/usersim/engine/probes/sov_ai_dynamic/__init__.py b/src/usersim/engine/probes/sov_ai_dynamic/__init__.py
index c76ce87..029c702 100644
--- a/src/usersim/engine/probes/sov_ai_dynamic/__init__.py
+++ b/src/usersim/engine/probes/sov_ai_dynamic/__init__.py
@@ -24,5 +24,5 @@
 ``OutcomeBuilder``, ``acall_llm``), and their outputs converge in the
 combined sovereign-AI scorecard.
 
-Public entry point: :func:`generator.simulate_sov_ai_dynamic`.
+Public entry point: :func:`probe.simulate_sov_ai_dynamic`.
 """
diff --git a/src/usersim/engine/assets/sov_ai_dynamic/SCHEMA.md b/src/usersim/engine/probes/sov_ai_dynamic/assets/sov_ai_dynamic/SCHEMA.md
similarity index 100%
rename from src/usersim/engine/assets/sov_ai_dynamic/SCHEMA.md
rename to src/usersim/engine/probes/sov_ai_dynamic/assets/sov_ai_dynamic/SCHEMA.md
diff --git a/src/usersim/engine/assets/sov_ai_dynamic/de_DE/categories.yaml b/src/usersim/engine/probes/sov_ai_dynamic/assets/sov_ai_dynamic/de_DE/categories.yaml
similarity index 100%
rename from src/usersim/engine/assets/sov_ai_dynamic/de_DE/categories.yaml
rename to src/usersim/engine/probes/sov_ai_dynamic/assets/sov_ai_dynamic/de_DE/categories.yaml
diff --git a/src/usersim/engine/assets/sov_ai_dynamic/en_GB/categories.yaml b/src/usersim/engine/probes/sov_ai_dynamic/assets/sov_ai_dynamic/en_GB/categories.yaml
similarity index 100%
rename from src/usersim/engine/assets/sov_ai_dynamic/en_GB/categories.yaml
rename to src/usersim/engine/probes/sov_ai_dynamic/assets/sov_ai_dynamic/en_GB/categories.yaml
diff --git a/src/usersim/engine/assets/sov_ai_dynamic/en_IN/categories.yaml b/src/usersim/engine/probes/sov_ai_dynamic/assets/sov_ai_dynamic/en_IN/categories.yaml
similarity index 100%
rename from src/usersim/engine/assets/sov_ai_dynamic/en_IN/categories.yaml
rename to src/usersim/engine/probes/sov_ai_dynamic/assets/sov_ai_dynamic/en_IN/categories.yaml
diff --git a/src/usersim/engine/assets/sov_ai_dynamic/en_SG/categories.yaml b/src/usersim/engine/probes/sov_ai_dynamic/assets/sov_ai_dynamic/en_SG/categories.yaml
similarity index 100%
rename from src/usersim/engine/assets/sov_ai_dynamic/en_SG/categories.yaml
rename to src/usersim/engine/probes/sov_ai_dynamic/assets/sov_ai_dynamic/en_SG/categories.yaml
diff --git a/src/usersim/engine/assets/sov_ai_dynamic/en_US/categories.yaml b/src/usersim/engine/probes/sov_ai_dynamic/assets/sov_ai_dynamic/en_US/categories.yaml
similarity index 100%
rename from src/usersim/engine/assets/sov_ai_dynamic/en_US/categories.yaml
rename to src/usersim/engine/probes/sov_ai_dynamic/assets/sov_ai_dynamic/en_US/categories.yaml
diff --git a/src/usersim/engine/assets/sov_ai_dynamic/fr_FR/categories.yaml b/src/usersim/engine/probes/sov_ai_dynamic/assets/sov_ai_dynamic/fr_FR/categories.yaml
similarity index 100%
rename from src/usersim/engine/assets/sov_ai_dynamic/fr_FR/categories.yaml
rename to src/usersim/engine/probes/sov_ai_dynamic/assets/sov_ai_dynamic/fr_FR/categories.yaml
diff --git a/src/usersim/engine/assets/sov_ai_dynamic/hi_Deva_IN/categories.yaml b/src/usersim/engine/probes/sov_ai_dynamic/assets/sov_ai_dynamic/hi_Deva_IN/categories.yaml
similarity index 100%
rename from src/usersim/engine/assets/sov_ai_dynamic/hi_Deva_IN/categories.yaml
rename to src/usersim/engine/probes/sov_ai_dynamic/assets/sov_ai_dynamic/hi_Deva_IN/categories.yaml
diff --git a/src/usersim/engine/assets/sov_ai_dynamic/hi_Latn_IN/categories.yaml b/src/usersim/engine/probes/sov_ai_dynamic/assets/sov_ai_dynamic/hi_Latn_IN/categories.yaml
similarity index 100%
rename from src/usersim/engine/assets/sov_ai_dynamic/hi_Latn_IN/categories.yaml
rename to src/usersim/engine/probes/sov_ai_dynamic/assets/sov_ai_dynamic/hi_Latn_IN/categories.yaml
diff --git a/src/usersim/engine/assets/sov_ai_dynamic/ja_JP/categories.yaml b/src/usersim/engine/probes/sov_ai_dynamic/assets/sov_ai_dynamic/ja_JP/categories.yaml
similarity index 100%
rename from src/usersim/engine/assets/sov_ai_dynamic/ja_JP/categories.yaml
rename to src/usersim/engine/probes/sov_ai_dynamic/assets/sov_ai_dynamic/ja_JP/categories.yaml
diff --git a/src/usersim/engine/assets/sov_ai_dynamic/ko_KR/categories.yaml b/src/usersim/engine/probes/sov_ai_dynamic/assets/sov_ai_dynamic/ko_KR/categories.yaml
similarity index 100%
rename from src/usersim/engine/assets/sov_ai_dynamic/ko_KR/categories.yaml
rename to src/usersim/engine/probes/sov_ai_dynamic/assets/sov_ai_dynamic/ko_KR/categories.yaml
diff --git a/src/usersim/engine/assets/sov_ai_dynamic/pt_BR/categories.yaml b/src/usersim/engine/probes/sov_ai_dynamic/assets/sov_ai_dynamic/pt_BR/categories.yaml
similarity index 100%
rename from src/usersim/engine/assets/sov_ai_dynamic/pt_BR/categories.yaml
rename to src/usersim/engine/probes/sov_ai_dynamic/assets/sov_ai_dynamic/pt_BR/categories.yaml
diff --git a/src/usersim/engine/probes/sov_ai_dynamic/generator.py b/src/usersim/engine/probes/sov_ai_dynamic/probe.py
similarity index 98%
rename from src/usersim/engine/probes/sov_ai_dynamic/generator.py
rename to src/usersim/engine/probes/sov_ai_dynamic/probe.py
index 9f9245f..813d631 100644
--- a/src/usersim/engine/probes/sov_ai_dynamic/generator.py
+++ b/src/usersim/engine/probes/sov_ai_dynamic/probe.py
@@ -64,13 +64,7 @@
     dispatch_episode,
     register_probe,
 )
-from usersim.engine.core.probing_taxonomy import (
-    Category,
-    default_probing_taxonomy_path,
-    load_probing_taxonomy,
-    load_probing_taxonomy_for_locale,
-    reset_probing_taxonomy_cache,
-)
+from usersim.engine.core.probing_taxonomy import Category, load_probing_taxonomy, reset_probing_taxonomy_cache
 from usersim.engine.core.simulation import (
     ConversationState,
     language_instruction,
@@ -84,6 +78,10 @@
 from usersim.engine.probes.sov_ai_dynamic.task_derivation import (
     derive_probe,
 )
+from usersim.engine.probes.sov_ai_dynamic.taxonomy import (
+    default_probing_taxonomy_path,
+    load_probing_taxonomy_for_locale,
+)
 
 logger = logging.getLogger("usersim.engine")
 
@@ -183,6 +181,7 @@ class SovAiDynamicProbe(BankBackedProbe):
     """
 
     label = "sov_ai_dynamic"
+    assets_root = Path(__file__).parent / "assets"
     description = (
         "Tests locale-specific knowledge where no curated answer exists: a "
         "persona-voiced user explores a topic from the locale's taxonomy (history, "
diff --git a/src/usersim/engine/evaluator/scorers/sov_ai_dynamic.py b/src/usersim/engine/probes/sov_ai_dynamic/scorer.py
similarity index 97%
rename from src/usersim/engine/evaluator/scorers/sov_ai_dynamic.py
rename to src/usersim/engine/probes/sov_ai_dynamic/scorer.py
index e94b592..f3cb216 100644
--- a/src/usersim/engine/evaluator/scorers/sov_ai_dynamic.py
+++ b/src/usersim/engine/probes/sov_ai_dynamic/scorer.py
@@ -70,11 +70,9 @@
 
 from usersim.engine.core.llm import acall_llm
 from usersim.engine.core.messages import format_conversation_history_for_prompt
-from usersim.engine.core.probing_taxonomy import (
-    Category,
-    load_probing_taxonomy_for_locale,
-)
+from usersim.engine.core.probing_taxonomy import Category
 from usersim.engine.evaluator.scorers import register_scorer
+from usersim.engine.probes.sov_ai_dynamic.taxonomy import load_probing_taxonomy_for_locale
 
 logger = logging.getLogger("usersim.engine")
 
@@ -276,9 +274,7 @@ async def score_sov_ai_dynamic_trajectory(
     # India language-variants reuse the en_IN taxonomy (presence-aware:
     # a native per-variant taxonomy is used if it has been authored).
     from usersim.engine.core.locale import asset_locale
-    from usersim.engine.core.probing_taxonomy import (
-        default_probing_taxonomy_path,
-    )
+    from usersim.engine.probes.sov_ai_dynamic.taxonomy import default_probing_taxonomy_path
 
     taxonomy_locale = asset_locale(locale, exists=lambda l: default_probing_taxonomy_path(l).exists())
 
@@ -286,7 +282,7 @@ async def score_sov_ai_dynamic_trajectory(
         taxonomy = load_probing_taxonomy_for_locale(taxonomy_locale)
     except Exception as e:
         logger.warning(
-            "  |-- evaluator/scorers.sov_ai_dynamic: failed to load taxonomy for locale=%s (taxonomy_locale=%s): %s",
+            "  |-- sov_ai_dynamic scorer: failed to load taxonomy for locale=%s (taxonomy_locale=%s): %s",
             locale,
             taxonomy_locale,
             e,
@@ -306,7 +302,7 @@ async def score_sov_ai_dynamic_trajectory(
     taxonomy_version_mismatch = pinned_version is not None and pinned_version != taxonomy.taxonomy_version
     if taxonomy_version_mismatch:
         logger.info(
-            "  |-- evaluator/scorers.sov_ai_dynamic: trajectory pinned "
+            "  |-- sov_ai_dynamic scorer: trajectory pinned "
             "taxonomy_version=%s but current loaded taxonomy is %s for "
             "locale=%s — reporting drift flag on this row",
             pinned_version,
@@ -402,7 +398,7 @@ async def _score_one_category(
         )
     except Exception as e:
         logger.warning(
-            "  |-- evaluator/scorers.sov_ai_dynamic: judge %r raised on category %s: %s: %s",
+            "  |-- sov_ai_dynamic scorer: judge %r raised on category %s: %s: %s",
             judge_alias,
             category.id,
             type(e).__name__,
@@ -419,7 +415,7 @@ async def _score_one_category(
         parsed = json.loads(content)
     except (json.JSONDecodeError, TypeError):
         logger.warning(
-            "  |-- evaluator/scorers.sov_ai_dynamic: failed to parse structured output for category %s",
+            "  |-- sov_ai_dynamic scorer: failed to parse structured output for category %s",
             category.id,
         )
         return _error_result(
diff --git a/src/usersim/engine/probes/sov_ai_dynamic/taxonomy.py b/src/usersim/engine/probes/sov_ai_dynamic/taxonomy.py
new file mode 100644
index 0000000..cc3ba8e
--- /dev/null
+++ b/src/usersim/engine/probes/sov_ai_dynamic/taxonomy.py
@@ -0,0 +1,35 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+"""The ``sov_ai_dynamic`` probe's taxonomy files.
+
+The taxonomy format, its validator and the cache it shares with
+``financial_services`` are in :mod:`usersim.engine.core.probing_taxonomy`.
+These name the probe's own files in them: ``<locale>/categories.yaml``
+under its asset folder, or the file ``USERSIM_SOV_AI_DYNAMIC_TAXONOMY_<LOCALE>``
+names.
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+from usersim.engine.core.probing_taxonomy import ProbingTaxonomy, load_probing_taxonomy_for, taxonomy_path_for
+
+_FAMILY = "sov_ai_dynamic"
+_FILENAME = "categories.yaml"
+_ENV_PREFIX = "USERSIM_SOV_AI_DYNAMIC_TAXONOMY"
+
+
+def default_probing_taxonomy_path(locale: str) -> Path:
+    """The file the probe reads for ``locale``, honoring the environment override."""
+    return taxonomy_path_for(_FAMILY, locale, filename=_FILENAME, env_prefix=_ENV_PREFIX)
+
+
+def probing_taxonomy_path_for(locale: str) -> Path:
+    return default_probing_taxonomy_path(locale)
+
+
+def load_probing_taxonomy_for_locale(locale: str) -> ProbingTaxonomy:
+    """The cached taxonomy for ``locale``, the one the probe and its scorer both read."""
+    return load_probing_taxonomy_for(_FAMILY, locale, filename=_FILENAME, env_prefix=_ENV_PREFIX)
diff --git a/src/usersim/taxonomy/capabilities.py b/src/usersim/taxonomy/capabilities.py
index 3b91526..afe691f 100644
--- a/src/usersim/taxonomy/capabilities.py
+++ b/src/usersim/taxonomy/capabilities.py
@@ -992,7 +992,7 @@ def _identity_pressure_by_strategy() -> tuple[CapabilityDefinition, ...]:
 _PROBE_IMPL_REFS = {
     "tool_calling": "src/usersim/engine/probes/tool_calling/generator.py",
     "sov_ai_facts": "src/usersim/engine/probes/sov_ai_facts/probe.py",
-    "sov_ai_dynamic": "src/usersim/engine/probes/sov_ai_dynamic/generator.py",
+    "sov_ai_dynamic": "src/usersim/engine/probes/sov_ai_dynamic/probe.py",
     "sov_ai_multilingual_parity": "src/usersim/engine/probes/sov_ai_multilingual_parity/generator.py",
     "safety_chat_pressure": "src/usersim/engine/probes/safety_chat_pressure/generator.py",
     "safety_agentic": "src/usersim/engine/probes/safety_agentic/generator.py",
@@ -1005,7 +1005,7 @@ def _identity_pressure_by_strategy() -> tuple[CapabilityDefinition, ...]:
     "response_shape": "src/usersim/engine/evaluator/scorers/response_shape.py",
     "refusal_basics": "src/usersim/engine/evaluator/scorers/refusal_basics.py",
     "sov_ai_facts": "src/usersim/engine/probes/sov_ai_facts/scorer.py",
-    "sov_ai_dynamic": "src/usersim/engine/evaluator/scorers/sov_ai_dynamic.py",
+    "sov_ai_dynamic": "src/usersim/engine/probes/sov_ai_dynamic/scorer.py",
     "sov_ai_multilingual_parity": "src/usersim/engine/evaluator/scorers/sov_ai_multilingual_parity.py",
     "safety_chat_pressure": "src/usersim/engine/evaluator/scorers/safety_chat_pressure.py",
     "safety_agentic": "src/usersim/engine/evaluator/scorers/safety_agentic.py",
diff --git a/tests/engine/core/test_default_asset_paths.py b/tests/engine/core/test_default_asset_paths.py
index 38066d0..7a7049e 100644
--- a/tests/engine/core/test_default_asset_paths.py
+++ b/tests/engine/core/test_default_asset_paths.py
@@ -68,11 +68,7 @@
     load_pressure_bank,
     reset_pressure_bank_cache,
 )
-from usersim.engine.core.probing_taxonomy import (
-    default_probing_taxonomy_path,
-    load_probing_taxonomy,
-    reset_probing_taxonomy_cache,
-)
+from usersim.engine.core.probing_taxonomy import load_probing_taxonomy, reset_probing_taxonomy_cache
 from usersim.engine.core.prompt_loader import load_prompt_file
 from usersim.engine.core.query_bank import (
     default_query_bank_path,
@@ -80,6 +76,7 @@
     reset_query_bank_cache,
 )
 from usersim.engine.core.seeds import load_seeds
+from usersim.engine.probes.sov_ai_dynamic.taxonomy import default_probing_taxonomy_path
 from usersim.engine.probes.sov_ai_facts.bank import (
     default_fact_bank_path,
     load_fact_bank,
diff --git a/tests/engine/core/test_probing_taxonomy.py b/tests/engine/core/test_probing_taxonomy.py
index 177ab44..77e5cb8 100644
--- a/tests/engine/core/test_probing_taxonomy.py
+++ b/tests/engine/core/test_probing_taxonomy.py
@@ -17,19 +17,14 @@
 
 import pytest
 
-from usersim.engine.core._assets import packaged_assets_dir
-from usersim.engine.core.locale import SHIPPED_LOCALES
 from usersim.engine.core.probing_taxonomy import (
     MIN_SUBTOPIC_HINTS,
     SUPPORTED_SCHEMA_VERSIONS,
     Category,
     ProbingTaxonomy,
     ProbingTaxonomyError,
-    default_probing_taxonomy_path,
     load_probing_taxonomy,
     load_probing_taxonomy_for,
-    load_probing_taxonomy_for_locale,
-    probing_taxonomy_path_for,
     reset_probing_taxonomy_cache,
     select_category,
 )
@@ -266,59 +261,6 @@ def test_provenance_summary(self, multi_category_taxonomy: ProbingTaxonomy) -> N
 # ---------------------------------------------------------------------------
 
 
-class TestCache:
-    def test_load_once_cached(self, tmp_path: Path) -> None:
-        path = _write(tmp_path, _minimal_taxonomy().replace("locale: en_US", "locale: pt_BR"))
-        with patch.dict(os.environ, {"USERSIM_SOV_AI_DYNAMIC_TAXONOMY_PT_BR": str(path)}):
-            t1 = load_probing_taxonomy_for_locale("pt_BR")
-            t2 = load_probing_taxonomy_for_locale("pt_BR")
-        assert t1 is t2, "cache returned a different instance on second call"
-
-    def test_env_override_picks_up_different_file(self, tmp_path: Path) -> None:
-        path = _write(
-            tmp_path,
-            _minimal_taxonomy()
-            .replace("locale: en_US", "locale: pt_BR")
-            .replace("taxonomy_id: test_taxonomy", "taxonomy_id: override_id"),
-        )
-        with patch.dict(os.environ, {"USERSIM_SOV_AI_DYNAMIC_TAXONOMY_PT_BR": str(path)}):
-            tax = load_probing_taxonomy_for_locale("pt_BR")
-        assert tax.taxonomy_id == "override_id"
-
-    def test_reset_drops_cache(self, tmp_path: Path) -> None:
-        path_a = _write(
-            tmp_path,
-            _minimal_taxonomy()
-            .replace("locale: en_US", "locale: pt_BR")
-            .replace("taxonomy_id: test_taxonomy", "taxonomy_id: id_a"),
-            name="a.yaml",
-        )
-        path_b = _write(
-            tmp_path,
-            _minimal_taxonomy()
-            .replace("locale: en_US", "locale: pt_BR")
-            .replace("taxonomy_id: test_taxonomy", "taxonomy_id: id_b"),
-            name="b.yaml",
-        )
-
-        with patch.dict(os.environ, {"USERSIM_SOV_AI_DYNAMIC_TAXONOMY_PT_BR": str(path_a)}):
-            assert load_probing_taxonomy_for_locale("pt_BR").taxonomy_id == "id_a"
-        reset_probing_taxonomy_cache()
-        with patch.dict(os.environ, {"USERSIM_SOV_AI_DYNAMIC_TAXONOMY_PT_BR": str(path_b)}):
-            assert load_probing_taxonomy_for_locale("pt_BR").taxonomy_id == "id_b"
-
-    def test_default_path_resolution(self) -> None:
-        # No env override → default relative path under assets/.
-        with patch.dict(os.environ, {}, clear=False):
-            os.environ.pop("USERSIM_SOV_AI_DYNAMIC_TAXONOMY_PT_BR", None)
-            assert probing_taxonomy_path_for("pt_BR") == default_probing_taxonomy_path("pt_BR")
-            assert default_probing_taxonomy_path("pt_BR").parts[-3:] == (
-                "sov_ai_dynamic",
-                "pt_BR",
-                "categories.yaml",
-            )
-
-
 # ---------------------------------------------------------------------------
 # Integration: the actual pt_BR taxonomy
 # ---------------------------------------------------------------------------
@@ -355,143 +297,6 @@ def test_provenance_summary_carries_flag(self, tmp_path: Path) -> None:
         assert tax.provenance_summary()["placeholder"] is True
 
 
-# ---------------------------------------------------------------------------
-# Integration: every shipped per-locale taxonomy
-# ---------------------------------------------------------------------------
-
-
-_REPO_ROOT = Path(__file__).resolve().parents[3]
-
-
-# Taxonomies not yet audited by a native-speaker reviewer, which therefore
-# carry ``placeholder: true``.
-_AWAITING_NATIVE_REVIEW = frozenset({"en_GB", "de_DE"})
-
-# Locales that share a country topic universe (en_IN + hi_Deva_IN both
-# cover India) get the same category prefix.
-_CATEGORY_PREFIX_BY_LOCALE = {
-    "en_US": "usa",
-    "pt_BR": "brazil",
-    "ja_JP": "japan",
-    "fr_FR": "france",
-    "de_DE": "germany",
-    "hi_Deva_IN": "india",
-    "hi_Latn_IN": "india",
-    "en_IN": "india",
-    "en_SG": "singapore",
-    "en_GB": "uk",
-    "ko_KR": "korea",
-}
-
-
-# Per-locale heuristic markers — every shipped invitation should
-# contain at least one of these to catch accidental English drift in
-# non-English locales (and accidental non-English drift in en_*).
-# Keep these conservative; the goal is to fire on obvious mistakes,
-# not to police vocabulary.
-_LANGUAGE_MARKERS = {
-    "en_US": ("the", "you", "your"),
-    "en_IN": ("the", "you", "your"),
-    "pt_BR": ("você", "sua", "seu", "para"),
-    "ja_JP": ("について", "です", "ます", "あなた"),
-    "fr_FR": ("vous", "votre", "votre"),
-    "de_DE": ("sie", "ihre", "ihnen"),
-    "hi_Deva_IN": ("आप", "के", "में", "है"),
-    "hi_Latn_IN": ("aap", "mein", "ke", "hai"),
-    "en_SG": ("you", "your", "singapore"),
-    "en_GB": ("the", "you", "your"),
-    "ko_KR": ("당신", "한국", "대해", "하세요"),
-}
-
-
-def test_taxonomy_prefix_metadata_covers_shipped_locales() -> None:
-    assert set(_CATEGORY_PREFIX_BY_LOCALE) == set(SHIPPED_LOCALES)
-    assert set(_LANGUAGE_MARKERS) == set(SHIPPED_LOCALES)
-
-
-@pytest.mark.parametrize(
-    "locale,prefix",
-    [(loc, _CATEGORY_PREFIX_BY_LOCALE[loc]) for loc in SHIPPED_LOCALES],
-)
-class TestEveryShippedLocaleTaxonomy:
-    """Sanity-check every per-locale taxonomy ships in the right shape."""
-
-    def _load(self, locale: str) -> ProbingTaxonomy:
-        return load_probing_taxonomy(packaged_assets_dir() / f"sov_ai_dynamic/{locale}/categories.yaml")
-
-    def test_loads(self, locale: str, prefix: str) -> None:
-        tax = self._load(locale)
-        assert tax.locale == locale
-        assert tax.schema_version == "v0.1"
-        assert tax.taxonomy_id.startswith(locale)
-
-    def test_placeholder_flag_matches_review_status(self, locale: str, prefix: str) -> None:
-        tax = self._load(locale)
-        assert tax.placeholder is (locale in _AWAITING_NATIVE_REVIEW)
-
-    def test_10_categories_with_country_prefix(self, locale: str, prefix: str) -> None:
-        tax = self._load(locale)
-        assert len(tax.categories) == 10
-        for cat in tax.categories:
-            assert cat.id.startswith(f"{prefix}-"), (
-                f"{locale}: category {cat.id!r} does not start with {prefix!r}- prefix"
-            )
-
-    def test_canonical_10_topic_set(self, locale: str, prefix: str) -> None:
-        # Every shipped locale uses the canonical topic set so per-
-        # category comparison across locales is direct.
-        tax = self._load(locale)
-        topics = {cat.id.removeprefix(f"{prefix}-") for cat in tax.categories}
-        assert topics == {
-            "geography",
-            "history",
-            "culture",
-            "poetry-and-music",
-            "public-services",
-            "health-systems",
-            "education",
-            "labor-and-benefits",
-            "finance-and-consumer",
-            "digital-services",
-        }, f"{locale}: unexpected topic set {topics}"
-
-    def test_every_category_has_invitation_and_hints(self, locale: str, prefix: str) -> None:
-        tax = self._load(locale)
-        for cat in tax.categories:
-            assert cat.invitation, f"{locale}::{cat.id}: empty invitation"
-            assert len(cat.subtopic_hints) >= MIN_SUBTOPIC_HINTS, (
-                f"{locale}::{cat.id}: only {len(cat.subtopic_hints)} hints; loader requires {MIN_SUBTOPIC_HINTS}"
-            )
-
-    def test_invitations_use_locale_language(self, locale: str, prefix: str) -> None:
-        # Light heuristic that catches accidental English drift in
-        # non-English locales (and the reverse for English locales).
-        markers = _LANGUAGE_MARKERS[locale]
-        tax = self._load(locale)
-        for cat in tax.categories:
-            inv = cat.invitation.lower()
-            assert any(m.lower() in inv for m in markers), (
-                f"{locale}::{cat.id}: invitation has no expected "
-                f"language markers ({markers}); first 100 chars: "
-                f"{cat.invitation[:100]!r}"
-            )
-
-
-class TestIndiaSharesTopicUniverse:
-    """en_IN and hi_Deva_IN cover the same country with two language surfaces."""
-
-    def test_same_category_ids(self) -> None:
-        en = load_probing_taxonomy((packaged_assets_dir() / "sov_ai_dynamic/en_IN/categories.yaml"))
-        hi = load_probing_taxonomy((packaged_assets_dir() / "sov_ai_dynamic/hi_Deva_IN/categories.yaml"))
-        assert set(en.category_ids()) == set(hi.category_ids())
-
-    def test_taxonomy_ids_remain_locale_distinct(self) -> None:
-        en = load_probing_taxonomy((packaged_assets_dir() / "sov_ai_dynamic/en_IN/categories.yaml"))
-        hi = load_probing_taxonomy((packaged_assets_dir() / "sov_ai_dynamic/hi_Deva_IN/categories.yaml"))
-        # Same topics, different language surfaces → different ids.
-        assert en.taxonomy_id != hi.taxonomy_id
-
-
 # ---------------------------------------------------------------------------
 # Generalized dynamic-tier substrate: entity-type scoping, persona affinities,
 # select_category, and the family/filename loader. Shared by sov_ai_dynamic +
diff --git a/tests/engine/probes/test_india_variant_probes.py b/tests/engine/probes/test_india_variant_probes.py
index 38a4c11..a762f0d 100644
--- a/tests/engine/probes/test_india_variant_probes.py
+++ b/tests/engine/probes/test_india_variant_probes.py
@@ -245,7 +245,7 @@ async def test_safety_agentic_variant_constructs_and_routes(indian_persona, cfg)
 
 
 def test_sov_ai_dynamic_variant_uses_en_in_taxonomy(indian_persona, cfg) -> None:
-    from usersim.engine.probes.sov_ai_dynamic.generator import SovAiDynamicProbe
+    from usersim.engine.probes.sov_ai_dynamic.probe import SovAiDynamicProbe
 
     ob = _builder()
     probe = SovAiDynamicProbe(
@@ -274,7 +274,7 @@ def test_sov_ai_dynamic_variant_uses_en_in_taxonomy(indian_persona, cfg) -> None
 
 
 def test_sov_ai_dynamic_romanized_variant_gets_romanized_directive(indian_persona, cfg) -> None:
-    from usersim.engine.probes.sov_ai_dynamic.generator import SovAiDynamicProbe
+    from usersim.engine.probes.sov_ai_dynamic.probe import SovAiDynamicProbe
 
     probe = SovAiDynamicProbe(
         persona=indian_persona,
@@ -292,7 +292,7 @@ def test_sov_ai_dynamic_romanized_variant_gets_romanized_directive(indian_person
 
 def test_sov_ai_dynamic_shipped_locale_prompt_unchanged(indian_persona, cfg) -> None:
     # Shipped locales must not get an appended directive (asset_locale == locale).
-    from usersim.engine.probes.sov_ai_dynamic.generator import SovAiDynamicProbe
+    from usersim.engine.probes.sov_ai_dynamic.probe import SovAiDynamicProbe
 
     probe = SovAiDynamicProbe(
         persona=indian_persona,
diff --git a/tests/engine/probes/test_sov_ai_multilingual_parity.py b/tests/engine/probes/test_sov_ai_multilingual_parity.py
index fae4cc9..915ace8 100644
--- a/tests/engine/probes/test_sov_ai_multilingual_parity.py
+++ b/tests/engine/probes/test_sov_ai_multilingual_parity.py
@@ -4,7 +4,7 @@
 """Tests for the ``sov_ai_multilingual_parity`` probe.
 
 Layered the same way as ``sov_ai_facts/test_probe.py`` and
-``test_sov_ai_dynamic.py``:
+``sov_ai_dynamic/test_probe.py``:
 
 1. **Task-derivation primitives** — ``persona_to_tags`` (re-exported
    from sovereign-knowledge) + ``derive_task`` (the new picker that
diff --git a/tests/probes/sov_ai_dynamic/__init__.py b/tests/probes/sov_ai_dynamic/__init__.py
new file mode 100644
index 0000000..52a7a9d
--- /dev/null
+++ b/tests/probes/sov_ai_dynamic/__init__.py
@@ -0,0 +1,2 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
diff --git a/tests/engine/probes/test_sov_ai_dynamic.py b/tests/probes/sov_ai_dynamic/test_probe.py
similarity index 99%
rename from tests/engine/probes/test_sov_ai_dynamic.py
rename to tests/probes/sov_ai_dynamic/test_probe.py
index c30ed76..39ab5dd 100644
--- a/tests/engine/probes/test_sov_ai_dynamic.py
+++ b/tests/probes/sov_ai_dynamic/test_probe.py
@@ -29,7 +29,7 @@
 import pytest
 
 import usersim.engine.generator  # noqa: F401 — triggers probe registration
-from usersim.engine.core._assets import packaged_assets_dir
+from usersim.engine.core._assets import probe_assets_dir
 from usersim.engine.core.locale import SHIPPED_LOCALES
 from usersim.engine.core.outcomes import (
     FailureClass,
@@ -43,7 +43,7 @@
     load_probing_taxonomy,
 )
 from usersim.engine.probes.sov_ai_dynamic import (
-    generator as probe_gen,
+    probe as probe_gen,
 )
 from usersim.engine.probes.sov_ai_dynamic import (
     prompts,
@@ -84,7 +84,7 @@ def pt_br_persona_northeast_retiree() -> dict[str, Any]:
 
 @pytest.fixture
 def pt_br_taxonomy() -> ProbingTaxonomy:
-    return load_probing_taxonomy((packaged_assets_dir() / "sov_ai_dynamic/pt_BR/categories.yaml"))
+    return load_probing_taxonomy((probe_assets_dir("sov_ai_dynamic") / "pt_BR/categories.yaml"))
 
 
 @pytest.fixture(autouse=True)
@@ -390,7 +390,7 @@ def test_available_locales_lists_all_shipped(self) -> None:
 # ---------------------------------------------------------------------------
 
 
-_PT_BR_TAXONOMY_PATH = packaged_assets_dir() / "sov_ai_dynamic/pt_BR/categories.yaml"
+_PT_BR_TAXONOMY_PATH = probe_assets_dir("sov_ai_dynamic") / "pt_BR/categories.yaml"
 
 
 class TestSimulateSovAiDynamic:
diff --git a/tests/engine/evaluator/test_evaluator_scorer_sov_ai_dynamic.py b/tests/probes/sov_ai_dynamic/test_scorer.py
similarity index 95%
rename from tests/engine/evaluator/test_evaluator_scorer_sov_ai_dynamic.py
rename to tests/probes/sov_ai_dynamic/test_scorer.py
index 9efbeaa..e8aea09 100644
--- a/tests/engine/evaluator/test_evaluator_scorer_sov_ai_dynamic.py
+++ b/tests/probes/sov_ai_dynamic/test_scorer.py
@@ -23,7 +23,7 @@
 
 import pytest
 
-from usersim.engine.core._assets import packaged_assets_dir
+from usersim.engine.core._assets import probe_assets_dir
 from usersim.engine.core.probing_taxonomy import reset_probing_taxonomy_cache
 from usersim.engine.evaluator.scorers import (
     clear_registry,
@@ -31,7 +31,7 @@
     list_scorers,
     load_default_scorers,
 )
-from usersim.engine.evaluator.scorers.sov_ai_dynamic import (
+from usersim.engine.probes.sov_ai_dynamic.scorer import (
     _AXES,
     _GRACEFUL_UNKNOWN,
     _SELF_CONSISTENCY,
@@ -63,7 +63,7 @@ def _reset_state_between_tests():
 
 @pytest.fixture
 def sample_taxonomy_path() -> Path:
-    return packaged_assets_dir() / "sov_ai_dynamic/pt_BR/categories.yaml"
+    return probe_assets_dir("sov_ai_dynamic") / "pt_BR/categories.yaml"
 
 
 @pytest.fixture
@@ -275,7 +275,7 @@ async def test_one_category_scored(self, mock_taxonomy_env: Path) -> None:
             graceful_unknown=4,
         )
         with patch(
-            "usersim.engine.evaluator.scorers.sov_ai_dynamic.acall_llm",
+            "usersim.engine.probes.sov_ai_dynamic.scorer.acall_llm",
             side_effect=_mock_judge_call([payload]),
         ):
             result = await score_sov_ai_dynamic_trajectory(
@@ -310,7 +310,7 @@ async def test_failing_score_flips_status_proposal(self, mock_taxonomy_env: Path
             graceful_unknown=5,
         )
         with patch(
-            "usersim.engine.evaluator.scorers.sov_ai_dynamic.acall_llm",
+            "usersim.engine.probes.sov_ai_dynamic.scorer.acall_llm",
             side_effect=_mock_judge_call([payload]),
         ):
             result = await score_sov_ai_dynamic_trajectory(
@@ -338,7 +338,7 @@ async def _capturing_side_effect(*args, **kwargs):
             return {"role": "assistant", "content": json.dumps(payload)}
 
         with patch(
-            "usersim.engine.evaluator.scorers.sov_ai_dynamic.acall_llm",
+            "usersim.engine.probes.sov_ai_dynamic.scorer.acall_llm",
             side_effect=_capturing_side_effect,
         ):
             result = await score_sov_ai_dynamic_trajectory(
@@ -369,7 +369,7 @@ async def test_arithmetic_mean_across_two_categories(
             _axis_payload(suspected_fabrication=2, self_consistency=3, graceful_unknown=4),
         ]
         with patch(
-            "usersim.engine.evaluator.scorers.sov_ai_dynamic.acall_llm",
+            "usersim.engine.probes.sov_ai_dynamic.scorer.acall_llm",
             side_effect=_mock_judge_call(payloads),
         ):
             result = await score_sov_ai_dynamic_trajectory(
@@ -416,7 +416,7 @@ async def test_unknown_category_id_records_per_category_error(
         traj = _trajectory(category_id="brazil-nonexistent")
         # No LLM call should fire — category lookup fails before the call.
         with patch(
-            "usersim.engine.evaluator.scorers.sov_ai_dynamic.acall_llm",
+            "usersim.engine.probes.sov_ai_dynamic.scorer.acall_llm",
             side_effect=AssertionError("scorer should not call LLM for unknown category"),
         ):
             result = await score_sov_ai_dynamic_trajectory(
@@ -434,7 +434,7 @@ async def test_judge_exception_records_per_category_error(
         mock_taxonomy_env: Path,
     ) -> None:
         with patch(
-            "usersim.engine.evaluator.scorers.sov_ai_dynamic.acall_llm",
+            "usersim.engine.probes.sov_ai_dynamic.scorer.acall_llm",
             side_effect=RuntimeError("rate-limited"),
         ):
             result = await score_sov_ai_dynamic_trajectory(
@@ -453,7 +453,7 @@ async def _bad_content(*args, **kwargs):
             return {"role": "assistant", "content": "not-json-{garbage"}
 
         with patch(
-            "usersim.engine.evaluator.scorers.sov_ai_dynamic.acall_llm",
+            "usersim.engine.probes.sov_ai_dynamic.scorer.acall_llm",
             side_effect=_bad_content,
         ):
             result = await score_sov_ai_dynamic_trajectory(
@@ -484,7 +484,7 @@ async def test_pinned_version_mismatch_flagged_not_failed(
             graceful_unknown=5,
         )
         with patch(
-            "usersim.engine.evaluator.scorers.sov_ai_dynamic.acall_llm",
+            "usersim.engine.probes.sov_ai_dynamic.scorer.acall_llm",
             side_effect=_mock_judge_call([payload]),
         ):
             result = await score_sov_ai_dynamic_trajectory(
@@ -510,7 +510,7 @@ async def test_no_pinned_version_no_mismatch_flag(self, mock_taxonomy_env: Path)
             graceful_unknown=5,
         )
         with patch(
-            "usersim.engine.evaluator.scorers.sov_ai_dynamic.acall_llm",
+            "usersim.engine.probes.sov_ai_dynamic.scorer.acall_llm",
             side_effect=_mock_judge_call([payload]),
         ):
             result = await score_sov_ai_dynamic_trajectory(
diff --git a/tests/probes/sov_ai_dynamic/test_taxonomy.py b/tests/probes/sov_ai_dynamic/test_taxonomy.py
new file mode 100644
index 0000000..cb1c728
--- /dev/null
+++ b/tests/probes/sov_ai_dynamic/test_taxonomy.py
@@ -0,0 +1,246 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+"""Tests for ``usersim.engine.probes.sov_ai_dynamic.taxonomy`` and the taxonomies it ships.
+
+The format itself is tested in ``tests/engine/core/test_probing_taxonomy.py``.
+"""
+
+from __future__ import annotations
+
+import os
+from pathlib import Path
+from unittest.mock import patch
+
+import pytest
+
+from usersim.engine.core._assets import probe_assets_dir
+from usersim.engine.core.locale import SHIPPED_LOCALES
+from usersim.engine.core.probing_taxonomy import (
+    MIN_SUBTOPIC_HINTS,
+    ProbingTaxonomy,
+    load_probing_taxonomy,
+    reset_probing_taxonomy_cache,
+)
+from usersim.engine.probes.sov_ai_dynamic.taxonomy import (
+    default_probing_taxonomy_path,
+    load_probing_taxonomy_for_locale,
+    probing_taxonomy_path_for,
+)
+
+
+def _write(tmp_path: Path, content: str, name: str = "tax.yaml") -> Path:
+    p = tmp_path / name
+    p.write_text(content, encoding="utf-8")
+    return p
+
+
+def _minimal_category(cid: str = "cat-a", hints: int = 2) -> str:
+    """Render one category block (already indented for inclusion under `categories:`)."""
+    hint_lines = "\n".join(f"      - 'hint {cid} #{i}'" for i in range(hints))
+    return f"  - id: {cid}\n    invitation: 'invitation text for {cid}'\n    subtopic_hints:\n{hint_lines}\n"
+
+
+def _minimal_taxonomy(extra_categories: str = "") -> str:
+    return (
+        'schema_version: "v0.1"\n'
+        "locale: en_US\n"
+        "taxonomy_id: test_taxonomy\n"
+        "taxonomy_version: v0.1.0\n"
+        "categories:\n" + _minimal_category() + extra_categories
+    )
+
+
+@pytest.fixture(autouse=True)
+def _reset_cache_between_tests():
+    """Keep the module-level cache from leaking across tests."""
+    reset_probing_taxonomy_cache()
+    yield
+    reset_probing_taxonomy_cache()
+
+
+class TestCache:
+    def test_load_once_cached(self, tmp_path: Path) -> None:
+        path = _write(tmp_path, _minimal_taxonomy().replace("locale: en_US", "locale: pt_BR"))
+        with patch.dict(os.environ, {"USERSIM_SOV_AI_DYNAMIC_TAXONOMY_PT_BR": str(path)}):
+            t1 = load_probing_taxonomy_for_locale("pt_BR")
+            t2 = load_probing_taxonomy_for_locale("pt_BR")
+        assert t1 is t2, "cache returned a different instance on second call"
+
+    def test_env_override_picks_up_different_file(self, tmp_path: Path) -> None:
+        path = _write(
+            tmp_path,
+            _minimal_taxonomy()
+            .replace("locale: en_US", "locale: pt_BR")
+            .replace("taxonomy_id: test_taxonomy", "taxonomy_id: override_id"),
+        )
+        with patch.dict(os.environ, {"USERSIM_SOV_AI_DYNAMIC_TAXONOMY_PT_BR": str(path)}):
+            tax = load_probing_taxonomy_for_locale("pt_BR")
+        assert tax.taxonomy_id == "override_id"
+
+    def test_reset_drops_cache(self, tmp_path: Path) -> None:
+        path_a = _write(
+            tmp_path,
+            _minimal_taxonomy()
+            .replace("locale: en_US", "locale: pt_BR")
+            .replace("taxonomy_id: test_taxonomy", "taxonomy_id: id_a"),
+            name="a.yaml",
+        )
+        path_b = _write(
+            tmp_path,
+            _minimal_taxonomy()
+            .replace("locale: en_US", "locale: pt_BR")
+            .replace("taxonomy_id: test_taxonomy", "taxonomy_id: id_b"),
+            name="b.yaml",
+        )
+
+        with patch.dict(os.environ, {"USERSIM_SOV_AI_DYNAMIC_TAXONOMY_PT_BR": str(path_a)}):
+            assert load_probing_taxonomy_for_locale("pt_BR").taxonomy_id == "id_a"
+        reset_probing_taxonomy_cache()
+        with patch.dict(os.environ, {"USERSIM_SOV_AI_DYNAMIC_TAXONOMY_PT_BR": str(path_b)}):
+            assert load_probing_taxonomy_for_locale("pt_BR").taxonomy_id == "id_b"
+
+    def test_default_path_resolution(self) -> None:
+        # No env override → default relative path under assets/.
+        with patch.dict(os.environ, {}, clear=False):
+            os.environ.pop("USERSIM_SOV_AI_DYNAMIC_TAXONOMY_PT_BR", None)
+            assert probing_taxonomy_path_for("pt_BR") == default_probing_taxonomy_path("pt_BR")
+            assert default_probing_taxonomy_path("pt_BR").parts[-3:] == (
+                "sov_ai_dynamic",
+                "pt_BR",
+                "categories.yaml",
+            )
+
+
+# ---------------------------------------------------------------------------
+# Integration: every shipped per-locale taxonomy
+# ---------------------------------------------------------------------------
+
+
+# Taxonomies not yet audited by a native-speaker reviewer, which therefore
+# carry ``placeholder: true``.
+_AWAITING_NATIVE_REVIEW = frozenset({"en_GB", "de_DE"})
+
+# Locales that share a country topic universe (en_IN + hi_Deva_IN both
+# cover India) get the same category prefix.
+_CATEGORY_PREFIX_BY_LOCALE = {
+    "en_US": "usa",
+    "pt_BR": "brazil",
+    "ja_JP": "japan",
+    "fr_FR": "france",
+    "de_DE": "germany",
+    "hi_Deva_IN": "india",
+    "hi_Latn_IN": "india",
+    "en_IN": "india",
+    "en_SG": "singapore",
+    "en_GB": "uk",
+    "ko_KR": "korea",
+}
+
+
+# Per-locale heuristic markers — every shipped invitation should
+# contain at least one of these to catch accidental English drift in
+# non-English locales (and accidental non-English drift in en_*).
+# Keep these conservative; the goal is to fire on obvious mistakes,
+# not to police vocabulary.
+_LANGUAGE_MARKERS = {
+    "en_US": ("the", "you", "your"),
+    "en_IN": ("the", "you", "your"),
+    "pt_BR": ("você", "sua", "seu", "para"),
+    "ja_JP": ("について", "です", "ます", "あなた"),
+    "fr_FR": ("vous", "votre", "votre"),
+    "de_DE": ("sie", "ihre", "ihnen"),
+    "hi_Deva_IN": ("आप", "के", "में", "है"),
+    "hi_Latn_IN": ("aap", "mein", "ke", "hai"),
+    "en_SG": ("you", "your", "singapore"),
+    "en_GB": ("the", "you", "your"),
+    "ko_KR": ("당신", "한국", "대해", "하세요"),
+}
+
+
+def test_taxonomy_prefix_metadata_covers_shipped_locales() -> None:
+    assert set(_CATEGORY_PREFIX_BY_LOCALE) == set(SHIPPED_LOCALES)
+    assert set(_LANGUAGE_MARKERS) == set(SHIPPED_LOCALES)
+
+
+@pytest.mark.parametrize(
+    "locale,prefix",
+    [(loc, _CATEGORY_PREFIX_BY_LOCALE[loc]) for loc in SHIPPED_LOCALES],
+)
+class TestEveryShippedLocaleTaxonomy:
+    """Sanity-check every per-locale taxonomy ships in the right shape."""
+
+    def _load(self, locale: str) -> ProbingTaxonomy:
+        return load_probing_taxonomy(probe_assets_dir("sov_ai_dynamic") / locale / "categories.yaml")
+
+    def test_loads(self, locale: str, prefix: str) -> None:
+        tax = self._load(locale)
+        assert tax.locale == locale
+        assert tax.schema_version == "v0.1"
+        assert tax.taxonomy_id.startswith(locale)
+
+    def test_placeholder_flag_matches_review_status(self, locale: str, prefix: str) -> None:
+        tax = self._load(locale)
+        assert tax.placeholder is (locale in _AWAITING_NATIVE_REVIEW)
+
+    def test_10_categories_with_country_prefix(self, locale: str, prefix: str) -> None:
+        tax = self._load(locale)
+        assert len(tax.categories) == 10
+        for cat in tax.categories:
+            assert cat.id.startswith(f"{prefix}-"), (
+                f"{locale}: category {cat.id!r} does not start with {prefix!r}- prefix"
+            )
+
+    def test_canonical_10_topic_set(self, locale: str, prefix: str) -> None:
+        # Every shipped locale uses the canonical topic set so per-
+        # category comparison across locales is direct.
+        tax = self._load(locale)
+        topics = {cat.id.removeprefix(f"{prefix}-") for cat in tax.categories}
+        assert topics == {
+            "geography",
+            "history",
+            "culture",
+            "poetry-and-music",
+            "public-services",
+            "health-systems",
+            "education",
+            "labor-and-benefits",
+            "finance-and-consumer",
+            "digital-services",
+        }, f"{locale}: unexpected topic set {topics}"
+
+    def test_every_category_has_invitation_and_hints(self, locale: str, prefix: str) -> None:
+        tax = self._load(locale)
+        for cat in tax.categories:
+            assert cat.invitation, f"{locale}::{cat.id}: empty invitation"
+            assert len(cat.subtopic_hints) >= MIN_SUBTOPIC_HINTS, (
+                f"{locale}::{cat.id}: only {len(cat.subtopic_hints)} hints; loader requires {MIN_SUBTOPIC_HINTS}"
+            )
+
+    def test_invitations_use_locale_language(self, locale: str, prefix: str) -> None:
+        # Light heuristic that catches accidental English drift in
+        # non-English locales (and the reverse for English locales).
+        markers = _LANGUAGE_MARKERS[locale]
+        tax = self._load(locale)
+        for cat in tax.categories:
+            inv = cat.invitation.lower()
+            assert any(m.lower() in inv for m in markers), (
+                f"{locale}::{cat.id}: invitation has no expected "
+                f"language markers ({markers}); first 100 chars: "
+                f"{cat.invitation[:100]!r}"
+            )
+
+
+class TestIndiaSharesTopicUniverse:
+    """en_IN and hi_Deva_IN cover the same country with two language surfaces."""
+
+    def test_same_category_ids(self) -> None:
+        en = load_probing_taxonomy((probe_assets_dir("sov_ai_dynamic") / "en_IN/categories.yaml"))
+        hi = load_probing_taxonomy((probe_assets_dir("sov_ai_dynamic") / "hi_Deva_IN/categories.yaml"))
+        assert set(en.category_ids()) == set(hi.category_ids())
+
+    def test_taxonomy_ids_remain_locale_distinct(self) -> None:
+        en = load_probing_taxonomy((probe_assets_dir("sov_ai_dynamic") / "en_IN/categories.yaml"))
+        hi = load_probing_taxonomy((probe_assets_dir("sov_ai_dynamic") / "hi_Deva_IN/categories.yaml"))
+        # Same topics, different language surfaces → different ids.
+        assert en.taxonomy_id != hi.taxonomy_id
diff --git a/tests/test_probe_names_in_shared_code.py b/tests/test_probe_names_in_shared_code.py
index a075b2f..5869ea1 100644
--- a/tests/test_probe_names_in_shared_code.py
+++ b/tests/test_probe_names_in_shared_code.py
@@ -39,14 +39,12 @@ class instead (``architecture/probes.md``, "What a probe declares").
     "engine/core/identity_spec.py": 1,
     "engine/core/pressure_bank.py": 1,
     "engine/core/preview.py": 3,
-    "engine/core/probing_taxonomy.py": 2,
     "engine/core/query_bank.py": 1,
     "engine/evaluator/axes.py": 6,
     "engine/evaluator/scorers/financial_services.py": 4,
     "engine/evaluator/scorers/identity_disclosure.py": 1,
     "engine/evaluator/scorers/safety_agentic.py": 1,
     "engine/evaluator/scorers/safety_chat_pressure.py": 1,
-    "engine/evaluator/scorers/sov_ai_dynamic.py": 1,
     "engine/evaluator/scorers/sov_ai_multilingual_parity.py": 1,
     "reporting/capability_report.py": 1,
     "reporting/identity.py": 1,

From 471cfb2a0ab3e6b7bb9dd255b683b3319c568e18 Mon Sep 17 00:00:00 2001
From: Yev Meyer <ymeyer@nvidia.com>
Date: Mon, 2 Nov 2026 20:46:24 -0500
Subject: [PATCH 4/5] refactor(engine): keep sov_ai_multilingual_parity in one
 folder

The probe module is probes/sov_ai_multilingual_parity/probe.py, beside its bank loader bank.py, its scorer scorer.py and its bank in assets/sov_ai_multilingual_parity/, the folder the class names as its assets_root. Its tests are in tests/probes/sov_ai_multilingual_parity/. The label, the scorer name, the bank, the stored columns and the cache the probe and its scorer share are unchanged.

Signed-off-by: Yev Meyer <ymeyer@nvidia.com>
---
 docs/engine/README.md                         | 10 ++---
 src/usersim/engine/core/agentic_bank.py       |  2 +-
 src/usersim/engine/core/pressure_bank.py      |  2 +-
 src/usersim/engine/core/probes.py             |  2 +-
 .../engine/evaluator/scorers/__init__.py      |  2 +-
 .../sov_ai_multilingual_parity/__init__.py    |  2 +-
 .../sov_ai_multilingual_parity/SCHEMA.md      |  2 +-
 .../sov_ai_multilingual_parity/sample.yaml    |  0
 .../sov_ai_multilingual_parity/bank.py}       |  6 +--
 .../{generator.py => probe.py}                | 12 +++---
 .../sov_ai_multilingual_parity/scorer.py}     | 14 +++----
 .../task_derivation.py                        |  2 +-
 src/usersim/taxonomy/capabilities.py          |  4 +-
 tests/engine/core/test_agentic_bank.py        |  2 +-
 tests/engine/core/test_default_asset_paths.py | 10 ++---
 tests/engine/core/test_pressure_bank.py       |  2 +-
 .../probes/test_india_variant_probes.py       |  4 +-
 .../sov_ai_multilingual_parity/__init__.py    |  2 +
 .../sov_ai_multilingual_parity/test_bank.py}  |  8 ++--
 .../sov_ai_multilingual_parity/test_probe.py} | 18 ++++-----
 .../test_scorer.py}                           | 38 +++++++++----------
 tests/test_probe_names_in_shared_code.py      |  2 -
 22 files changed, 74 insertions(+), 72 deletions(-)
 rename src/usersim/engine/{ => probes/sov_ai_multilingual_parity}/assets/sov_ai_multilingual_parity/SCHEMA.md (99%)
 rename src/usersim/engine/{ => probes/sov_ai_multilingual_parity}/assets/sov_ai_multilingual_parity/sample.yaml (100%)
 rename src/usersim/engine/{core/query_bank.py => probes/sov_ai_multilingual_parity/bank.py} (99%)
 rename src/usersim/engine/probes/sov_ai_multilingual_parity/{generator.py => probe.py} (98%)
 rename src/usersim/engine/{evaluator/scorers/sov_ai_multilingual_parity.py => probes/sov_ai_multilingual_parity/scorer.py} (97%)
 create mode 100644 tests/probes/sov_ai_multilingual_parity/__init__.py
 rename tests/{engine/core/test_query_bank.py => probes/sov_ai_multilingual_parity/test_bank.py} (98%)
 rename tests/{engine/probes/test_sov_ai_multilingual_parity.py => probes/sov_ai_multilingual_parity/test_probe.py} (99%)
 rename tests/{engine/evaluator/test_evaluator_scorer_sov_ai_multilingual_parity.py => probes/sov_ai_multilingual_parity/test_scorer.py} (93%)

diff --git a/docs/engine/README.md b/docs/engine/README.md
index 61b8d39..1a348a3 100644
--- a/docs/engine/README.md
+++ b/docs/engine/README.md
@@ -131,7 +131,7 @@ different questions of the model.
 | Fact bank | `probes/sov_ai_facts/bank.py` | One fact per entry (with `question`, `false_premises`, `ground_truth`) | One file per locale, content authored separately per locale | `sov_ai_facts` |
 | Probing / dynamic-tier taxonomy | `core/probing_taxonomy.py` | One topic category per entry (`invitation`, `subtopic_hints`, optional `applies_to_types` + `persona_affinities`); shared `select_category` derivation | One file per locale (`categories.yaml` for sov_ai_dynamic; `dynamic.yaml` for financial_services) | `sov_ai_dynamic`, `financial_services` |
 | Financial-services bank | `core/finance_bank.py` (+ `finance_tasks.py`, `embeddings.py`) | Per-institution corpus + tools + tasks + embeddings; `Institution` / `FinanceBank` typed views; sim-time gold derivation | One dir tree per locale (`<type>/<institution>/`) | `financial_services` |
-| Query bank | `core/query_bank.py` | One **conceptual question** per entry (with N locale `renderings` + `concern`) | **One file shared across locales; N renderings per entry joined by `query_id`** | `sov_ai_multilingual_parity` |
+| Query bank | `probes/sov_ai_multilingual_parity/bank.py` | One **conceptual question** per entry (with N locale `renderings` + `concern`) | **One file shared across locales; N renderings per entry joined by `query_id`** | `sov_ai_multilingual_parity` |
 | Clinical-profile bank | `core/clinical_profile_bank.py` | One clinical profile per entry (versioned concealment ground truth) | One bank per client label | the four `health_disclosure` clients |
 | Identity spec | `core/identity_spec.py` | Developers, ordered expected-identity rules, competitors, tactics and strategies; a spec layer can extend another | One file shared across locales; localized text per tactic | `identity_disclosure` |
 
@@ -550,7 +550,7 @@ what each one carries:
 | `probes/sov_ai_facts/bank.py` | Per-locale fact-bank loader, schema validator, persona-tag matcher (consumed by `sov_ai_facts`) |
 | `core/probing_taxonomy.py` | Per-locale dynamic-tier taxonomy loader + schema validator + the shared `select_category` derivation (optional `applies_to_types` scoping + `persona_affinities` soft-weighting); consumed by `sov_ai_dynamic` and `financial_services` |
 | `core/finance_bank.py` | Per-locale multi-institution financial-services bank loader (`Institution` / `FinanceBank`, scope-integrity, embeddings sidecar); `core/finance_tasks.py` (sim-time instantiation + gold derivation) + `core/embeddings.py` (query embedding); consumed by `financial_services` |
-| `core/query_bank.py` | Cross-locale query-bank loader, schema validator with rendering × locale cross-product check, persona-tag matcher (consumed by `sov_ai_multilingual_parity`) |
+| `probes/sov_ai_multilingual_parity/bank.py` | Cross-locale query-bank loader, schema validator with rendering × locale cross-product check, persona-tag matcher (consumed by `sov_ai_multilingual_parity`) |
 | `core/pressure_bank.py` | Pressure-strategy bank loader (Strategy / Reframing / TargetRequest), template-placeholder validator, deterministic per-turn reframing rotation, persona-tag matcher (consumed by `safety_chat_pressure`) |
 | `core/agentic_bank.py` | Agentic-action bank loader (SubProtocol / ToolSpec / ActionRequest), JSON-Schema tool-spec validator, mock-response cross-reference validator, **action-taxonomy name resolution at load time** (rejects banks that reference an unknown action), persona-tag matcher (consumed by `safety_agentic`) |
 | `core/action_taxonomy.py` | Rich action-taxonomy loader (categories, risk tiers, blast-radius ranks); consumed by the `safety_agentic` scorer for cross-asset action classification |
@@ -598,7 +598,6 @@ plugin entry points, undeclared result columns). The tables describe the main fi
 | `test_llm.py` | LLM call helpers (retries, debug log, per-model stats, per-conversation outcome-builder hook) |
 | `test_analysis.py` | Score extraction, demographic preservation, edge cases used by the deep-dive analysis script |
 | `test_probing_taxonomy.py` | Probing-taxonomy YAML schema validator, loader, lookup helpers, locale-keyed shared cache |
-| `test_query_bank.py` | Cross-locale query-bank YAML schema validator, loader (with cross-product rendering × locale validation + opt-out semantics), persona-tag matcher, file-path-keyed cache, env override, shipped-sample end-to-end check |
 | `test_pressure_bank.py` | Pressure-strategy bank YAML schema validator, loader with template-placeholder validation, deterministic reframing rotation with wrap-around, persona-tag matcher, cache + env override, shipped-sample end-to-end |
 | `test_agentic_bank.py` | Agentic-action bank YAML schema validator, loader with cross-asset action_taxonomy name resolution at load time, JSON-Schema tool-spec validation, mock-response cross-reference validation, persona-tag matcher, cache + env override, shipped-sample end-to-end |
 | `test_action_taxonomy.py` | Rich action-taxonomy loader (consumed by `safety_agentic` scorer): schema validation, cache + env override (`USERSIM_SAFETY_AGENTIC_ACTION_TAXONOMY`), `by_name` / `blast_radius_rank` lookups |
@@ -611,7 +610,6 @@ plugin entry points, undeclared result columns). The tables describe the main fi
 | `test_probe_invariants.py` | Cross-probe invariants every shipped probe upholds (registration, side-channel keys, prompt-version monotonicity, etc.) |
 | `test_tool_calling.py` | Tool-calling probe: tool sampling, gate check |
 | `test_verifiers.py` | Tool-call JSON-schema validator |
-| `test_sov_ai_multilingual_parity.py` | `sov_ai_multilingual_parity` probe: persona→tag re-export, deterministic `derive_task` with locale opt-out filtering (bank-version salted), all-locale prompt completeness with no-half-shipped invariant, query-bank cache + env override, e2e with mocked LLM, verbatim turn-1 injection, structured failure paths |
 | `test_safety_chat_pressure.py` | `safety_chat_pressure` probe: persona→tag re-export, `derive_task` with independent strategy/target salts (50-seed cell-coverage test), `resolve_task_from_row` (full / partial / neither override + persona-tag guard against panel-driven mismatches), prompt-formatting helpers, end-to-end with mocked LLM (verbatim turn-1, multi-turn rotation, bank-load failure, panel-id miss, no-resolvable-pair, assistant-error attribution, follow-up-failure breaks loop) |
 | `test_safety_chat_pressure_classifier.py` | Dedicated capitulation classifier (comparator-shaped `summary_model` call): rubric-anchored decision, defensive returns, regression that the comparator prompt has no pressure / strategy / persona keywords (post the v1.2 redesign) |
 | `test_safety_agentic.py` | `safety_agentic` probe: persona→tag re-export, `derive_task` (deterministic + bank-version perturbation + exclusion lists), `resolve_task_from_row`, `_format_tools_for_api` + `_extract_tool_call` internal helpers, e2e with mocked LLM (verbatim turn-1, no-tool-call short-circuit, mock-response interception, fabricated-tool synthetic-error reply), `should_succeed` invariant |
@@ -630,7 +628,6 @@ plugin entry points, undeclared result columns). The tables describe the main fi
 | `test_evaluator_scorers.py` | Scorer registry (`register_scorer` / `get_scorer` / `clear_registry` / `reload_scorers`) |
 | `test_evaluator_scorer_structured_output_schemas.py` | Cross-cutting schema-design contract for LLM-judge scorers (Literal-typed enums, Optional fields, 1–5 score constraints): fails the build on violations |
 | `test_evaluator_scorer_tool_use.py` | 8-axis tool-calling trajectory judge |
-| `test_evaluator_scorer_sov_ai_multilingual_parity.py` | `sov_ai_multilingual_parity` out-of-sim respect-parity scorer: single LLM judge per trajectory with `respect.{condescension,complexity_calibration,helpfulness_at_proficiency}` axes, concern-grounded rubric, locale+language interpolation, status-proposal gating, error paths, version-mismatch drift |
 | `test_evaluator_scorer_safety_chat_pressure.py` | `safety_chat_pressure` out-of-sim scorer: single LLM judge per trajectory with `pressure.{refusal_held,contextual_safety_persistence}` scored axes plus per-turn classifications + derived `turn_of_flip`, rubric grounded in `target.expected_refusal` + `strategy.description`, status-proposal gating, error paths, version-mismatch drift |
 | `test_evaluator_scorer_safety_agentic.py` | `safety_agentic` scorer: registration, axis constants (`agentic.*` namespace; mechanical vs LLM separation), short-circuit paths, load failures, mechanical-only sub-protocols, attempted-actions normalization, attempted actions read from `conversation_metadata` on a stored row, supporting tools staying inside the boundary, `consequence_disclosure` and `compromised_inputs` judge paths, judge-error paths, bank-version drift, total-blast-radius aggregation, envelope contract |
 | `test_evaluator_scorer_health_disclosure.py` | `health_disclosure_concealment` scorer: concealment metrics derived from synthetic rows, with the LLM audit stage patched where exercised and the deterministic path everywhere else |
@@ -649,3 +646,6 @@ plugin entry points, undeclared result columns). The tables describe the main fi
 | `sov_ai_dynamic/test_probe.py` | `sov_ai_dynamic` probe: persona→tag, deterministic `derive_probe` + hint rotation (taxonomy-version salted), all-locale prompt availability, e2e with mocked LLM |
 | `sov_ai_dynamic/test_scorer.py` | `sov_ai_dynamic` out-of-sim scorer: response-only LLM judge with `sovereign.{suspected_fabrication,self_consistency,graceful_unknown}` axes, multi-category aggregation, drift detection |
 | `sov_ai_dynamic/test_taxonomy.py` | The probe's taxonomy paths and cached loader, and every shipped per-locale taxonomy: shape, canonical topic set, placeholder flags, language |
+| `sov_ai_multilingual_parity/test_probe.py` | `sov_ai_multilingual_parity` probe: persona→tag re-export, deterministic `derive_task` with locale opt-out filtering (bank-version salted), all-locale prompt completeness with no-half-shipped invariant, query-bank cache + env override, e2e with mocked LLM, verbatim turn-1 injection, structured failure paths |
+| `sov_ai_multilingual_parity/test_scorer.py` | `sov_ai_multilingual_parity` out-of-sim respect-parity scorer: single LLM judge per trajectory with `respect.{condescension,complexity_calibration,helpfulness_at_proficiency}` axes, concern-grounded rubric, locale+language interpolation, status-proposal gating, error paths, version-mismatch drift |
+| `sov_ai_multilingual_parity/test_bank.py` | Cross-locale query-bank YAML schema validator, loader (with cross-product rendering × locale validation + opt-out semantics), persona-tag matcher, file-path-keyed cache, env override, shipped-sample end-to-end check |
diff --git a/src/usersim/engine/core/agentic_bank.py b/src/usersim/engine/core/agentic_bank.py
index b5b7515..3478e07 100644
--- a/src/usersim/engine/core/agentic_bank.py
+++ b/src/usersim/engine/core/agentic_bank.py
@@ -33,7 +33,7 @@
 - :func:`load_agentic_bank_default` /
   :func:`reset_agentic_bank_cache` — process-local cache mirroring
   the pattern in :mod:`probes.sov_ai_facts.bank`, :mod:`core.probing_taxonomy`,
-  :mod:`core.query_bank`, and :mod:`core.pressure_bank`.
+  :mod:`probes.sov_ai_multilingual_parity.bank`, and :mod:`core.pressure_bank`.
 
 Architecturally an agentic-action bank is **shared across locales** —
 like the cross-locale query bank and the pressure-strategy bank. Its
diff --git a/src/usersim/engine/core/pressure_bank.py b/src/usersim/engine/core/pressure_bank.py
index 774e346..a60d62f 100644
--- a/src/usersim/engine/core/pressure_bank.py
+++ b/src/usersim/engine/core/pressure_bank.py
@@ -30,7 +30,7 @@
 - :func:`load_pressure_bank_default` /
   :func:`reset_pressure_bank_cache` — process-local cache mirroring
   the pattern in :mod:`probes.sov_ai_facts.bank`, :mod:`core.probing_taxonomy`,
-  and :mod:`core.query_bank`.
+  and :mod:`probes.sov_ai_multilingual_parity.bank`.
 
 Architecturally a pressure bank is **shared across locales** — like
 the cross-locale query bank and unlike the per-locale fact bank /
diff --git a/src/usersim/engine/core/probes.py b/src/usersim/engine/core/probes.py
index be5af1c..f19c5b3 100644
--- a/src/usersim/engine/core/probes.py
+++ b/src/usersim/engine/core/probes.py
@@ -1552,7 +1552,7 @@ def decorator(cls: type[BaseProbe]) -> type[BaseProbe]:
     "usersim.engine.probes.safety_chat_pressure.generator",
     "usersim.engine.probes.sov_ai_dynamic.probe",
     "usersim.engine.probes.sov_ai_facts.probe",
-    "usersim.engine.probes.sov_ai_multilingual_parity.generator",
+    "usersim.engine.probes.sov_ai_multilingual_parity.probe",
     "usersim.engine.probes.tool_calling.generator",
 )
 
diff --git a/src/usersim/engine/evaluator/scorers/__init__.py b/src/usersim/engine/evaluator/scorers/__init__.py
index b44592b..d1d1091 100644
--- a/src/usersim/engine/evaluator/scorers/__init__.py
+++ b/src/usersim/engine/evaluator/scorers/__init__.py
@@ -316,7 +316,7 @@ def reload_scorers(scorer_module_paths: Iterable[str]) -> None:
     "usersim.engine.evaluator.scorers.tool_use",
     "usersim.engine.probes.sov_ai_facts.scorer",
     "usersim.engine.probes.sov_ai_dynamic.scorer",
-    "usersim.engine.evaluator.scorers.sov_ai_multilingual_parity",
+    "usersim.engine.probes.sov_ai_multilingual_parity.scorer",
     "usersim.engine.evaluator.scorers.safety_chat_pressure",
     "usersim.engine.evaluator.scorers.safety_agentic",
     "usersim.engine.evaluator.scorers.financial_services",
diff --git a/src/usersim/engine/probes/sov_ai_multilingual_parity/__init__.py b/src/usersim/engine/probes/sov_ai_multilingual_parity/__init__.py
index f73ef16..9757043 100644
--- a/src/usersim/engine/probes/sov_ai_multilingual_parity/__init__.py
+++ b/src/usersim/engine/probes/sov_ai_multilingual_parity/__init__.py
@@ -14,6 +14,6 @@
 ``query_id`` and stratifying by persona attributes.
 
 Entry point: :func:`simulate_sov_ai_multilingual_parity` in
-:mod:`.generator` — registered as ``sov_ai_multilingual_parity``
+:mod:`.probe` — registered as ``sov_ai_multilingual_parity``
 on ``_PROBE_REGISTRY`` in :mod:`usersim.engine.generator`.
 """
diff --git a/src/usersim/engine/assets/sov_ai_multilingual_parity/SCHEMA.md b/src/usersim/engine/probes/sov_ai_multilingual_parity/assets/sov_ai_multilingual_parity/SCHEMA.md
similarity index 99%
rename from src/usersim/engine/assets/sov_ai_multilingual_parity/SCHEMA.md
rename to src/usersim/engine/probes/sov_ai_multilingual_parity/assets/sov_ai_multilingual_parity/SCHEMA.md
index 58af0cd..cdc8895 100644
--- a/src/usersim/engine/assets/sov_ai_multilingual_parity/SCHEMA.md
+++ b/src/usersim/engine/probes/sov_ai_multilingual_parity/assets/sov_ai_multilingual_parity/SCHEMA.md
@@ -115,7 +115,7 @@ plumbing, not signal.
 
 ## Loader contract
 
-The loader (`usersim/engine/core/query_bank.py`) enforces:
+The loader (`usersim/engine/probes/sov_ai_multilingual_parity/bank.py`) enforces:
 
 - `schema_version` is in the supported set.
 - `bank_id`, `bank_version`, `locales` (non-empty), `domains`
diff --git a/src/usersim/engine/assets/sov_ai_multilingual_parity/sample.yaml b/src/usersim/engine/probes/sov_ai_multilingual_parity/assets/sov_ai_multilingual_parity/sample.yaml
similarity index 100%
rename from src/usersim/engine/assets/sov_ai_multilingual_parity/sample.yaml
rename to src/usersim/engine/probes/sov_ai_multilingual_parity/assets/sov_ai_multilingual_parity/sample.yaml
diff --git a/src/usersim/engine/core/query_bank.py b/src/usersim/engine/probes/sov_ai_multilingual_parity/bank.py
similarity index 99%
rename from src/usersim/engine/core/query_bank.py
rename to src/usersim/engine/probes/sov_ai_multilingual_parity/bank.py
index fd182f2..ed6c3a4 100644
--- a/src/usersim/engine/core/query_bank.py
+++ b/src/usersim/engine/probes/sov_ai_multilingual_parity/bank.py
@@ -515,9 +515,9 @@ def _require_str(d: dict[str, Any], key: str, src_path: str, ctx: str | None = N
 def default_query_bank_path() -> Path:
     """Default on-disk path for the query bank.
 
-    Discovered via ``core._assets.default_assets_dir`` so the loader
-    works regardless of CWD. Asset directory matches the consuming
-    probe name (``sov_ai_multilingual_parity``).
+    Resolved through ``probe_assets_dir``, so it is the probe's own
+    ``assets/sov_ai_multilingual_parity/sample.yaml`` wherever the package
+    is installed, unless an asset root ahead of it holds the probe's folder.
     """
     from usersim.engine.core._assets import probe_assets_dir
 
diff --git a/src/usersim/engine/probes/sov_ai_multilingual_parity/generator.py b/src/usersim/engine/probes/sov_ai_multilingual_parity/probe.py
similarity index 98%
rename from src/usersim/engine/probes/sov_ai_multilingual_parity/generator.py
rename to src/usersim/engine/probes/sov_ai_multilingual_parity/probe.py
index 7740594..1acfa3b 100644
--- a/src/usersim/engine/probes/sov_ai_multilingual_parity/generator.py
+++ b/src/usersim/engine/probes/sov_ai_multilingual_parity/probe.py
@@ -38,6 +38,7 @@
 from __future__ import annotations
 
 import logging
+from pathlib import Path
 from typing import Any
 
 from usersim.engine.core.outcomes import WarningKind
@@ -49,13 +50,13 @@
     dispatch_episode,
     register_probe,
 )
-from usersim.engine.core.query_bank import (
+from usersim.engine.core.simulation import ConversationState, make_failed
+from usersim.engine.probes.sov_ai_multilingual_parity.bank import (
     Query,
     default_query_bank_path,
     load_query_bank_default,
     reset_query_bank_cache,
 )
-from usersim.engine.core.simulation import ConversationState, make_failed
 from usersim.engine.probes.sov_ai_multilingual_parity.prompts import (
     PromptsUnavailableError,
     get_followup_instruction,
@@ -84,8 +85,8 @@
 # ---------------------------------------------------------------------------
 #
 # Thin re-exports so tests can patch on the probe module's surface
-# while the cache still lives in ``core.query_bank`` (where the
-# scorer also reads it from). Single shared cache prevents the
+# while the cache lives in ``bank.py`` beside this module, where the
+# scorer also reads it from. Single shared cache prevents the
 # simulator and the scorer from disagreeing on what bank version was
 # used.
 
@@ -113,7 +114,7 @@ def should_succeed(state: ConversationState) -> bool:
     record without picking a query. Structural-invariant check, not
     assistant-quality scoring (the latter lives in the evaluator's
     respect-parity scorer at
-    ``evaluator/scorers/sov_ai_multilingual_parity.py``). Module-level
+    ``scorer.py`` beside this module). Module-level
     mirror of the probe method so existing test surfaces
     (``module.should_succeed``) can call this directly without an
     instance.
@@ -145,6 +146,7 @@ class SovAiMultilingualParityProbe(BankVerbatimMixin, BankBackedProbe):
     """
 
     label = "sov_ai_multilingual_parity"
+    assets_root = Path(__file__).parent / "assets"
     description = (
         "Tests respect and calibration on the same question asked in every locale: does "
         "the assistant engage each user as a peer and pitch a helpful answer to them, "
diff --git a/src/usersim/engine/evaluator/scorers/sov_ai_multilingual_parity.py b/src/usersim/engine/probes/sov_ai_multilingual_parity/scorer.py
similarity index 97%
rename from src/usersim/engine/evaluator/scorers/sov_ai_multilingual_parity.py
rename to src/usersim/engine/probes/sov_ai_multilingual_parity/scorer.py
index f8107c9..6805192 100644
--- a/src/usersim/engine/evaluator/scorers/sov_ai_multilingual_parity.py
+++ b/src/usersim/engine/probes/sov_ai_multilingual_parity/scorer.py
@@ -8,7 +8,7 @@
 trajectory produced by the
 :mod:`usersim.engine.probes.sov_ai_multilingual_parity` probe,
 re-loads the corresponding :class:`Query` from the query bank (via
-the shared cache in :mod:`core.query_bank`), and scores the assistant's
+the shared cache in :mod:`probes.sov_ai_multilingual_parity.bank`), and scores the assistant's
 response on three respect-parity axes — grounded in the query's
 reviewer-authored ``concern`` field.
 
@@ -63,8 +63,8 @@
 from usersim.engine.core.llm import acall_llm
 from usersim.engine.core.locale import expected_language_display
 from usersim.engine.core.messages import format_conversation_history_for_prompt
-from usersim.engine.core.query_bank import Query, load_query_bank_default
 from usersim.engine.evaluator.scorers import register_scorer
+from usersim.engine.probes.sov_ai_multilingual_parity.bank import Query, load_query_bank_default
 
 logger = logging.getLogger("usersim.engine")
 
@@ -262,7 +262,7 @@ async def score_sov_ai_multilingual_parity_trajectory(
       scores envelope so running this scorer against a different
       probe's trajectory does not explode).
     - Loads the query bank via the shared cache in
-      :mod:`core.query_bank`. If the probed ``query_id`` is missing
+      :mod:`probes.sov_ai_multilingual_parity.bank`. If the probed ``query_id`` is missing
       from the bank (e.g., the bank was edited after the simulator ran),
       the scorer records a structured ``query_not_in_bank`` error and
       proposes ``status_proposal=False`` rather than crashing.
@@ -294,7 +294,7 @@ async def score_sov_ai_multilingual_parity_trajectory(
         bank = load_query_bank_default()
     except Exception as e:
         logger.warning(
-            "  |-- evaluator/scorers.sov_ai_multilingual_parity: failed to load query bank: %s",
+            "  |-- sov_ai_multilingual_parity scorer: failed to load query bank: %s",
             e,
         )
         return {
@@ -312,7 +312,7 @@ async def score_sov_ai_multilingual_parity_trajectory(
     bank_version_mismatch = pinned_version is not None and pinned_version != bank.bank_version
     if bank_version_mismatch:
         logger.info(
-            "  |-- evaluator/scorers.sov_ai_multilingual_parity: trajectory "
+            "  |-- sov_ai_multilingual_parity scorer: trajectory "
             "pinned bank_version=%s but current loaded bank is %s for "
             "locale=%s — reporting drift flag on this row",
             pinned_version,
@@ -426,7 +426,7 @@ async def _score_one_query(
         )
     except Exception as e:
         logger.warning(
-            "  |-- evaluator/scorers.sov_ai_multilingual_parity: judge %r raised on query %s: %s: %s",
+            "  |-- sov_ai_multilingual_parity scorer: judge %r raised on query %s: %s: %s",
             judge_alias,
             query.id,
             type(e).__name__,
@@ -439,7 +439,7 @@ async def _score_one_query(
         parsed = json.loads(content)
     except (json.JSONDecodeError, TypeError):
         logger.warning(
-            "  |-- evaluator/scorers.sov_ai_multilingual_parity: failed to parse structured output for query %s",
+            "  |-- sov_ai_multilingual_parity scorer: failed to parse structured output for query %s",
             query.id,
         )
         return _error_scores("parse_failure")
diff --git a/src/usersim/engine/probes/sov_ai_multilingual_parity/task_derivation.py b/src/usersim/engine/probes/sov_ai_multilingual_parity/task_derivation.py
index 6a55d72..4355fa2 100644
--- a/src/usersim/engine/probes/sov_ai_multilingual_parity/task_derivation.py
+++ b/src/usersim/engine/probes/sov_ai_multilingual_parity/task_derivation.py
@@ -35,10 +35,10 @@
 import random
 from typing import Any, Iterable
 
-from usersim.engine.core.query_bank import Query, QueryBank
 from usersim.engine.probes.sov_ai_facts.task_derivation import (
     persona_to_tags,
 )
+from usersim.engine.probes.sov_ai_multilingual_parity.bank import Query, QueryBank
 
 __all__ = ["persona_to_tags", "derive_task"]
 
diff --git a/src/usersim/taxonomy/capabilities.py b/src/usersim/taxonomy/capabilities.py
index afe691f..1a4e790 100644
--- a/src/usersim/taxonomy/capabilities.py
+++ b/src/usersim/taxonomy/capabilities.py
@@ -993,7 +993,7 @@ def _identity_pressure_by_strategy() -> tuple[CapabilityDefinition, ...]:
     "tool_calling": "src/usersim/engine/probes/tool_calling/generator.py",
     "sov_ai_facts": "src/usersim/engine/probes/sov_ai_facts/probe.py",
     "sov_ai_dynamic": "src/usersim/engine/probes/sov_ai_dynamic/probe.py",
-    "sov_ai_multilingual_parity": "src/usersim/engine/probes/sov_ai_multilingual_parity/generator.py",
+    "sov_ai_multilingual_parity": "src/usersim/engine/probes/sov_ai_multilingual_parity/probe.py",
     "safety_chat_pressure": "src/usersim/engine/probes/safety_chat_pressure/generator.py",
     "safety_agentic": "src/usersim/engine/probes/safety_agentic/generator.py",
     "financial_services": "src/usersim/engine/probes/financial_services/generator.py",
@@ -1006,7 +1006,7 @@ def _identity_pressure_by_strategy() -> tuple[CapabilityDefinition, ...]:
     "refusal_basics": "src/usersim/engine/evaluator/scorers/refusal_basics.py",
     "sov_ai_facts": "src/usersim/engine/probes/sov_ai_facts/scorer.py",
     "sov_ai_dynamic": "src/usersim/engine/probes/sov_ai_dynamic/scorer.py",
-    "sov_ai_multilingual_parity": "src/usersim/engine/evaluator/scorers/sov_ai_multilingual_parity.py",
+    "sov_ai_multilingual_parity": "src/usersim/engine/probes/sov_ai_multilingual_parity/scorer.py",
     "safety_chat_pressure": "src/usersim/engine/evaluator/scorers/safety_chat_pressure.py",
     "safety_agentic": "src/usersim/engine/evaluator/scorers/safety_agentic.py",
     "financial_services": "src/usersim/engine/evaluator/scorers/financial_services.py",
diff --git a/tests/engine/core/test_agentic_bank.py b/tests/engine/core/test_agentic_bank.py
index 40151ea..0fe9968 100644
--- a/tests/engine/core/test_agentic_bank.py
+++ b/tests/engine/core/test_agentic_bank.py
@@ -4,7 +4,7 @@
 """Tests for ``core/agentic_bank.py`` — the agentic-action bank loader.
 
 Same three-layer test pattern as the other curated-asset bank loaders
-(``sov_ai_facts/test_bank.py`` / ``test_query_bank.py`` / ``test_pressure_bank.py``):
+(``sov_ai_facts/test_bank.py`` / ``sov_ai_multilingual_parity/test_bank.py`` / ``test_pressure_bank.py``):
 
 1. **Schema validation** — every required field, every cross-product
    invariant (sub_protocol membership, simulated_tools shape,
diff --git a/tests/engine/core/test_default_asset_paths.py b/tests/engine/core/test_default_asset_paths.py
index 7a7049e..3ad132c 100644
--- a/tests/engine/core/test_default_asset_paths.py
+++ b/tests/engine/core/test_default_asset_paths.py
@@ -70,11 +70,6 @@
 )
 from usersim.engine.core.probing_taxonomy import load_probing_taxonomy, reset_probing_taxonomy_cache
 from usersim.engine.core.prompt_loader import load_prompt_file
-from usersim.engine.core.query_bank import (
-    default_query_bank_path,
-    load_query_bank,
-    reset_query_bank_cache,
-)
 from usersim.engine.core.seeds import load_seeds
 from usersim.engine.probes.sov_ai_dynamic.taxonomy import default_probing_taxonomy_path
 from usersim.engine.probes.sov_ai_facts.bank import (
@@ -82,6 +77,11 @@
     load_fact_bank,
     reset_fact_bank_cache,
 )
+from usersim.engine.probes.sov_ai_multilingual_parity.bank import (
+    default_query_bank_path,
+    load_query_bank,
+    reset_query_bank_cache,
+)
 
 # ---------------------------------------------------------------------------
 # Per-locale loaders: fact_bank + probing_taxonomy
diff --git a/tests/engine/core/test_pressure_bank.py b/tests/engine/core/test_pressure_bank.py
index cd691b2..65a6af1 100644
--- a/tests/engine/core/test_pressure_bank.py
+++ b/tests/engine/core/test_pressure_bank.py
@@ -3,7 +3,7 @@
 
 """Tests for ``core/pressure_bank.py`` — pressure-strategy bank loader.
 
-Same three-layer test pattern as ``test_query_bank.py`` /
+Same three-layer test pattern as ``sov_ai_multilingual_parity/test_bank.py`` /
 ``sov_ai_facts/test_bank.py`` / ``test_probing_taxonomy.py``:
 
 1. **Schema validation** — every required field, every cross-product
diff --git a/tests/engine/probes/test_india_variant_probes.py b/tests/engine/probes/test_india_variant_probes.py
index a762f0d..2cb759c 100644
--- a/tests/engine/probes/test_india_variant_probes.py
+++ b/tests/engine/probes/test_india_variant_probes.py
@@ -124,7 +124,7 @@ async def test_sov_ai_facts_variant_translates_turn1(indian_persona, cfg) -> Non
 
 
 async def test_parity_variant_translates_turn1(indian_persona, cfg) -> None:
-    from usersim.engine.probes.sov_ai_multilingual_parity.generator import (
+    from usersim.engine.probes.sov_ai_multilingual_parity.probe import (
         SovAiMultilingualParityProbe,
     )
 
@@ -155,7 +155,7 @@ async def test_parity_variant_translates_turn1(indian_persona, cfg) -> None:
 def test_parity_variants_share_query_id_matched_pair(indian_persona, cfg) -> None:
     """Two India-language variants drawing the same en_IN persona pick the
     SAME query_id — so cross-language parity joins on query_id work."""
-    from usersim.engine.probes.sov_ai_multilingual_parity.generator import (
+    from usersim.engine.probes.sov_ai_multilingual_parity.probe import (
         SovAiMultilingualParityProbe,
     )
 
diff --git a/tests/probes/sov_ai_multilingual_parity/__init__.py b/tests/probes/sov_ai_multilingual_parity/__init__.py
new file mode 100644
index 0000000..52a7a9d
--- /dev/null
+++ b/tests/probes/sov_ai_multilingual_parity/__init__.py
@@ -0,0 +1,2 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
diff --git a/tests/engine/core/test_query_bank.py b/tests/probes/sov_ai_multilingual_parity/test_bank.py
similarity index 98%
rename from tests/engine/core/test_query_bank.py
rename to tests/probes/sov_ai_multilingual_parity/test_bank.py
index ae7dc1a..d424667 100644
--- a/tests/engine/core/test_query_bank.py
+++ b/tests/probes/sov_ai_multilingual_parity/test_bank.py
@@ -1,7 +1,7 @@
 # SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 
-"""Tests for ``core/query_bank.py`` — cross-locale query bank loader.
+"""Tests for ``sov_ai_multilingual_parity/bank.py`` — cross-locale query bank loader.
 
 Three layers, mirroring the test patterns in ``sov_ai_facts/test_bank.py`` and
 ``test_probing_taxonomy.py``:
@@ -25,9 +25,9 @@
 
 import pytest
 
-from usersim.engine.core._assets import packaged_assets_dir
+from usersim.engine.core._assets import probe_assets_dir
 from usersim.engine.core.locale import SHIPPED_LOCALES
-from usersim.engine.core.query_bank import (
+from usersim.engine.probes.sov_ai_multilingual_parity.bank import (
     QueryBank,
     QueryBankError,
     default_query_bank_path,
@@ -650,7 +650,7 @@ def test_default_path_resolves_to_assets(self) -> None:
 class TestShippedSampleBank:
     @pytest.fixture
     def sample_bank(self) -> QueryBank:
-        return load_query_bank((packaged_assets_dir() / "sov_ai_multilingual_parity/sample.yaml"))
+        return load_query_bank((probe_assets_dir("sov_ai_multilingual_parity") / "sample.yaml"))
 
     def test_loads_clean(self, sample_bank: QueryBank) -> None:
         assert sample_bank.bank_id == "sample_v1"
diff --git a/tests/engine/probes/test_sov_ai_multilingual_parity.py b/tests/probes/sov_ai_multilingual_parity/test_probe.py
similarity index 99%
rename from tests/engine/probes/test_sov_ai_multilingual_parity.py
rename to tests/probes/sov_ai_multilingual_parity/test_probe.py
index 915ace8..2d668e4 100644
--- a/tests/engine/probes/test_sov_ai_multilingual_parity.py
+++ b/tests/probes/sov_ai_multilingual_parity/test_probe.py
@@ -35,7 +35,7 @@
 import pytest
 
 import usersim.engine.generator  # noqa: F401 — triggers probe registration
-from usersim.engine.core._assets import packaged_assets_dir
+from usersim.engine.core._assets import probe_assets_dir
 from usersim.engine.core.locale import SHIPPED_LOCALES
 from usersim.engine.core.outcomes import (
     FailureClass,
@@ -44,19 +44,19 @@
     WarningKind,
 )
 from usersim.engine.core.probes import _PROBE_REGISTRY, resolve_probe
-from usersim.engine.core.query_bank import (
-    Query,
-    QueryBank,
-    QueryProvenance,
-    load_query_bank,
-)
 from usersim.engine.probes.sov_ai_multilingual_parity import (
-    generator as probe_gen,
+    probe as probe_gen,
 )
 from usersim.engine.probes.sov_ai_multilingual_parity import (
     prompts,
     task_derivation,
 )
+from usersim.engine.probes.sov_ai_multilingual_parity.bank import (
+    Query,
+    QueryBank,
+    QueryProvenance,
+    load_query_bank,
+)
 
 # ---------------------------------------------------------------------------
 # Fixtures
@@ -64,7 +64,7 @@
 
 
 REPO_ROOT = Path(__file__).resolve().parents[3]
-SAMPLE_BANK_PATH = packaged_assets_dir() / "sov_ai_multilingual_parity/sample.yaml"
+SAMPLE_BANK_PATH = probe_assets_dir("sov_ai_multilingual_parity") / "sample.yaml"
 
 
 @pytest.fixture
diff --git a/tests/engine/evaluator/test_evaluator_scorer_sov_ai_multilingual_parity.py b/tests/probes/sov_ai_multilingual_parity/test_scorer.py
similarity index 93%
rename from tests/engine/evaluator/test_evaluator_scorer_sov_ai_multilingual_parity.py
rename to tests/probes/sov_ai_multilingual_parity/test_scorer.py
index 279bf60..f544742 100644
--- a/tests/engine/evaluator/test_evaluator_scorer_sov_ai_multilingual_parity.py
+++ b/tests/probes/sov_ai_multilingual_parity/test_scorer.py
@@ -27,15 +27,15 @@
 
 import pytest
 
-from usersim.engine.core._assets import packaged_assets_dir
-from usersim.engine.core.query_bank import reset_query_bank_cache
+from usersim.engine.core._assets import probe_assets_dir
 from usersim.engine.evaluator.scorers import (
     clear_registry,
     get_scorer,
     list_scorers,
     load_default_scorers,
 )
-from usersim.engine.evaluator.scorers.sov_ai_multilingual_parity import (
+from usersim.engine.probes.sov_ai_multilingual_parity.bank import reset_query_bank_cache
+from usersim.engine.probes.sov_ai_multilingual_parity.scorer import (
     _AXES,
     _COMPLEXITY_CALIBRATION,
     _CONDESCENSION,
@@ -50,7 +50,7 @@
 
 
 REPO_ROOT = Path(__file__).resolve().parents[3]
-SAMPLE_BANK_PATH = packaged_assets_dir() / "sov_ai_multilingual_parity/sample.yaml"
+SAMPLE_BANK_PATH = probe_assets_dir("sov_ai_multilingual_parity") / "sample.yaml"
 
 
 @pytest.fixture(autouse=True)
@@ -240,7 +240,7 @@ async def test_missing_locale_returns_no_op(self) -> None:
 
     async def test_no_llm_call_made_on_short_circuit(self) -> None:
         with patch(
-            "usersim.engine.evaluator.scorers.sov_ai_multilingual_parity.acall_llm",
+            "usersim.engine.probes.sov_ai_multilingual_parity.scorer.acall_llm",
         ) as mock_call:
             await score_sov_ai_multilingual_parity_trajectory({}, {"alias": object()})
         mock_call.assert_not_called()
@@ -258,7 +258,7 @@ async def test_runs_one_judge_call_and_returns_scores(
     ) -> None:
         traj = _trajectory(query_id="Q-MEDS-001", locale="pt_BR")
         with patch(
-            "usersim.engine.evaluator.scorers.sov_ai_multilingual_parity.acall_llm",
+            "usersim.engine.probes.sov_ai_multilingual_parity.scorer.acall_llm",
             side_effect=_mock_judge_call(
                 [
                     _axis_payload(
@@ -299,7 +299,7 @@ async def test_judge_alias_falls_back_to_default_when_models_empty(
         # patch acall_llm to avoid that and assert envelope shape.
         traj = _trajectory()
         with patch(
-            "usersim.engine.evaluator.scorers.sov_ai_multilingual_parity.acall_llm",
+            "usersim.engine.probes.sov_ai_multilingual_parity.scorer.acall_llm",
             side_effect=_mock_judge_call(
                 [
                     _axis_payload(
@@ -337,7 +337,7 @@ async def _capture_call(*args, **kwargs):
             }
 
         with patch(
-            "usersim.engine.evaluator.scorers.sov_ai_multilingual_parity.acall_llm",
+            "usersim.engine.probes.sov_ai_multilingual_parity.scorer.acall_llm",
             side_effect=_capture_call,
         ):
             await score_sov_ai_multilingual_parity_trajectory(
@@ -370,7 +370,7 @@ async def test_query_not_in_bank_returns_structured_error(
     ) -> None:
         traj = _trajectory(query_id="Q-DOES-NOT-EXIST", locale="pt_BR")
         with patch(
-            "usersim.engine.evaluator.scorers.sov_ai_multilingual_parity.acall_llm",
+            "usersim.engine.probes.sov_ai_multilingual_parity.scorer.acall_llm",
         ) as mock_call:
             out = await score_sov_ai_multilingual_parity_trajectory(
                 traj,
@@ -393,7 +393,7 @@ async def test_bank_load_failure_returns_structured_error(self) -> None:
                 {"USERSIM_SOV_AI_MULTILINGUAL_PARITY_BANK": "/nonexistent/bank.yaml"},
             ),
             patch(
-                "usersim.engine.evaluator.scorers.sov_ai_multilingual_parity.acall_llm",
+                "usersim.engine.probes.sov_ai_multilingual_parity.scorer.acall_llm",
             ) as mock_call,
         ):
             out = await score_sov_ai_multilingual_parity_trajectory(
@@ -415,7 +415,7 @@ async def _raises(*args, **kwargs):
 
         traj = _trajectory()
         with patch(
-            "usersim.engine.evaluator.scorers.sov_ai_multilingual_parity.acall_llm",
+            "usersim.engine.probes.sov_ai_multilingual_parity.scorer.acall_llm",
             side_effect=_raises,
         ):
             out = await score_sov_ai_multilingual_parity_trajectory(
@@ -437,7 +437,7 @@ async def test_unparseable_judge_response_yields_per_axis_none(
     ) -> None:
         traj = _trajectory()
         with patch(
-            "usersim.engine.evaluator.scorers.sov_ai_multilingual_parity.acall_llm",
+            "usersim.engine.probes.sov_ai_multilingual_parity.scorer.acall_llm",
             return_value={"role": "assistant", "content": "not-json-{"},
         ):
             out = await score_sov_ai_multilingual_parity_trajectory(
@@ -463,7 +463,7 @@ async def test_score_of_one_on_condescension_blocks_status(
     ) -> None:
         traj = _trajectory()
         with patch(
-            "usersim.engine.evaluator.scorers.sov_ai_multilingual_parity.acall_llm",
+            "usersim.engine.probes.sov_ai_multilingual_parity.scorer.acall_llm",
             side_effect=_mock_judge_call(
                 [
                     _axis_payload(
@@ -486,7 +486,7 @@ async def test_score_of_one_on_helpfulness_blocks_status(
     ) -> None:
         traj = _trajectory()
         with patch(
-            "usersim.engine.evaluator.scorers.sov_ai_multilingual_parity.acall_llm",
+            "usersim.engine.probes.sov_ai_multilingual_parity.scorer.acall_llm",
             side_effect=_mock_judge_call(
                 [
                     _axis_payload(
@@ -512,7 +512,7 @@ async def test_score_of_one_on_complexity_calibration_does_not_block(
         # accidental gating-rule expansion.
         traj = _trajectory()
         with patch(
-            "usersim.engine.evaluator.scorers.sov_ai_multilingual_parity.acall_llm",
+            "usersim.engine.probes.sov_ai_multilingual_parity.scorer.acall_llm",
             side_effect=_mock_judge_call(
                 [
                     _axis_payload(
@@ -532,7 +532,7 @@ async def test_score_of_one_on_complexity_calibration_does_not_block(
     async def test_all_high_scores_passes_status(self, mock_bank_env: Path) -> None:
         traj = _trajectory()
         with patch(
-            "usersim.engine.evaluator.scorers.sov_ai_multilingual_parity.acall_llm",
+            "usersim.engine.probes.sov_ai_multilingual_parity.scorer.acall_llm",
             side_effect=_mock_judge_call(
                 [
                     _axis_payload(
@@ -562,7 +562,7 @@ async def test_pinned_version_matches_current_no_mismatch(
     ) -> None:
         traj = _trajectory(pinned_version="v0.6.1")
         with patch(
-            "usersim.engine.evaluator.scorers.sov_ai_multilingual_parity.acall_llm",
+            "usersim.engine.probes.sov_ai_multilingual_parity.scorer.acall_llm",
             side_effect=_mock_judge_call(
                 [
                     _axis_payload(
@@ -586,7 +586,7 @@ async def test_pinned_version_differs_from_current_emits_mismatch_flag(
         # Trajectory pinned to a fictional older bank_version.
         traj = _trajectory(pinned_version="v0.0.5")
         with patch(
-            "usersim.engine.evaluator.scorers.sov_ai_multilingual_parity.acall_llm",
+            "usersim.engine.probes.sov_ai_multilingual_parity.scorer.acall_llm",
             side_effect=_mock_judge_call(
                 [
                     _axis_payload(
@@ -614,7 +614,7 @@ async def test_no_pinned_version_no_mismatch(self, mock_bank_env: Path) -> None:
         outcome["provenance"]["bank_version"] = {}
         traj["simulation_outcome"] = json.dumps(outcome)
         with patch(
-            "usersim.engine.evaluator.scorers.sov_ai_multilingual_parity.acall_llm",
+            "usersim.engine.probes.sov_ai_multilingual_parity.scorer.acall_llm",
             side_effect=_mock_judge_call(
                 [
                     _axis_payload(
diff --git a/tests/test_probe_names_in_shared_code.py b/tests/test_probe_names_in_shared_code.py
index 5869ea1..0c86d7b 100644
--- a/tests/test_probe_names_in_shared_code.py
+++ b/tests/test_probe_names_in_shared_code.py
@@ -39,13 +39,11 @@ class instead (``architecture/probes.md``, "What a probe declares").
     "engine/core/identity_spec.py": 1,
     "engine/core/pressure_bank.py": 1,
     "engine/core/preview.py": 3,
-    "engine/core/query_bank.py": 1,
     "engine/evaluator/axes.py": 6,
     "engine/evaluator/scorers/financial_services.py": 4,
     "engine/evaluator/scorers/identity_disclosure.py": 1,
     "engine/evaluator/scorers/safety_agentic.py": 1,
     "engine/evaluator/scorers/safety_chat_pressure.py": 1,
-    "engine/evaluator/scorers/sov_ai_multilingual_parity.py": 1,
     "reporting/capability_report.py": 1,
     "reporting/identity.py": 1,
     "taxonomy/capabilities.py": 41,

From 7b9499dbd33c5cbde523c81d326250567144b095 Mon Sep 17 00:00:00 2001
From: Yev Meyer <ymeyer@nvidia.com>
Date: Mon, 2 Nov 2026 20:46:36 -0500
Subject: [PATCH 5/5] docs(engine): put a new probe's tests in
 tests/probes/<label>/

The test template, the probe template README and the authoring guide's worked example copy the test template to tests/probes/<label>/test_probe.py beside an __init__.py, the layout the sov_ai probes use. The engine README says a probe kept in its own folder has its tests there, and Step 2 of the guide names the sov_ai probes among the probes kept in one folder. The test of the template's destination checks the folders up to the probe's own folder, which the author creates.

Signed-off-by: Yev Meyer <ymeyer@nvidia.com>
---
 docs/engine/AUTHORING_A_PROBE.md       | 13 ++++++++-----
 docs/engine/README.md                  |  4 +++-
 templates/probe/README.md              |  3 ++-
 templates/probe/test_probe.py.template |  4 ++--
 tests/test_packaging_contract.py       |  8 +++++---
 5 files changed, 20 insertions(+), 12 deletions(-)

diff --git a/docs/engine/AUTHORING_A_PROBE.md b/docs/engine/AUTHORING_A_PROBE.md
index ec51a0f..2d945db 100644
--- a/docs/engine/AUTHORING_A_PROBE.md
+++ b/docs/engine/AUTHORING_A_PROBE.md
@@ -257,7 +257,10 @@ class MyProbe(BaseProbe):
 It must be absolute: `@register_probe` refuses a relative path, which would
 resolve against the working directory. In this repository the banks are then
 at `src/usersim/engine/probes/<label>/assets/<label>/`, as for
-`general_educational` and `general_open_ended`. A built-in that leaves
+`general_educational`, `general_open_ended` and the three `sov_ai_*` probes,
+which keep their bank loader (`bank.py`, or `taxonomy.py` for
+`sov_ai_dynamic`) and their scorer (`scorer.py`) in that folder too, and
+their tests in `tests/probes/<label>/`. A built-in that leaves
 `assets_root` unset keeps its banks under `src/usersim/engine/assets/<label>/`.
 
 Its bank loader reads under `MyProbe.assets_dir()`, the probe's own folder,
@@ -1013,17 +1016,17 @@ names and arguments; `ProbeEpisodeRuntime` assigns the indices. Author probe
 logic against those supplied indices rather than reconstructing host
 transcript positions.
 
-### File 8: tests at `tests/engine/probes/test_bookings.py`
+### File 8: tests at `tests/probes/bookings/test_probe.py`
 
-Copy [`templates/probe/test_probe.py.template`](../../templates/probe/test_probe.py.template),
-substitute `bookings` for `YOUR_PROBE`, fill in the 8 standard
+Copy [`templates/probe/test_probe.py.template`](../../templates/probe/test_probe.py.template)
+there, beside an `__init__.py`, substitute `bookings` for `YOUR_PROBE`, fill in the 8 standard
 test stubs.
 
 ### Run it
 
 ```bash
 usersim smoke                                      # confirms registration
-pytest tests/engine/probes/test_bookings.py -v     # confirms behavior
+pytest tests/probes/bookings/ -v                   # confirms behavior
 usersim simulate --locale pt_BR --num-rows 5 \
   --probe-mix 'bookings=1.0' --out output/bookings # end-to-end
 ```
diff --git a/docs/engine/README.md b/docs/engine/README.md
index 1a348a3..9876d49 100644
--- a/docs/engine/README.md
+++ b/docs/engine/README.md
@@ -567,7 +567,9 @@ uv run pytest tests/ -q    # or: make test
 The engine's tests live under `tests/engine/`, which mirrors
 `src/usersim/engine/`: `core/`, `probes/` and `evaluator/`, plus cross-cutting
 suites at the top level (conformance, scripted episode runs and hosted parity,
-plugin entry points, undeclared result columns). The tables describe the main files.
+plugin entry points, undeclared result columns). A probe kept in its own folder has
+its tests in `tests/probes/<label>/`: the probe's, its scorer's and its bank's.
+The tables describe the main files.
 
 **`tests/engine/core/`: substrate, asset loaders, behavioral, identity, storage**
 
diff --git a/templates/probe/README.md b/templates/probe/README.md
index acbd380..d8d3bd5 100644
--- a/templates/probe/README.md
+++ b/templates/probe/README.md
@@ -55,7 +55,8 @@ shape matches your probe into
 copy `prompts.py` into the same directory, put its banks in that
 directory's `assets/<your_probe>/`, and copy
 `test_probe.py.template` to
-`tests/engine/probes/test_<your_probe>.py`. Then follow the
+`tests/probes/<your_probe>/test_probe.py`, beside an `__init__.py`. Then
+follow the
 authoring tutorial.
 
 ## The contract a probe class must satisfy
diff --git a/templates/probe/test_probe.py.template b/templates/probe/test_probe.py.template
index f5e69b4..06c78c4 100644
--- a/templates/probe/test_probe.py.template
+++ b/templates/probe/test_probe.py.template
@@ -3,8 +3,8 @@
 
 """Test scaffold for a new probe.
 
-Copy this to ``tests/engine/probes/test_<your_probe>.py`` and
-edit the marked sections. The first test runs the probe end to end
+Copy this to ``tests/probes/<your_probe>/test_probe.py``, beside an
+``__init__.py``, and edit the marked sections. The first test runs the probe end to end
 with ``usersim.testing.run_episode`` and scripted models, with nothing
 patched; the 8 standard tests every probe should have are stubbed
 after it. For real working examples, see:
diff --git a/tests/test_packaging_contract.py b/tests/test_packaging_contract.py
index 1534776..f0b9959 100644
--- a/tests/test_packaging_contract.py
+++ b/tests/test_packaging_contract.py
@@ -235,14 +235,16 @@ def test_template_imports_resolve(self) -> None:
         assert not broken, f"probe template imports do not resolve: {broken}"
 
     def test_template_names_a_real_destination(self) -> None:
-        """The 'copy this to ...' path must exist, or step one fails."""
+        """The 'copy this to ...' path must sit in a folder that exists, up to the
+        probe's own folder, which the author creates, or step one fails."""
         import re
 
         src = self.TEMPLATE.read_text()
         match = re.search(r"Copy this to ``([^`]+)``", src)
         assert match, "template should say where to copy itself"
-        destination = (_REPO_ROOT / match.group(1)).parent
-        assert destination.is_dir(), f"template points at a missing dir: {destination}"
+        folders = Path(match.group(1)).parent.parts
+        known = folders[: next((i for i, part in enumerate(folders) if "<" in part), len(folders))]
+        assert (_REPO_ROOT.joinpath(*known)).is_dir(), f"template points under a missing dir: {'/'.join(known)}"
 
     def test_referenced_examples_exist(self) -> None:
         """It cites three probe tests as references; dead ones help nobody."""