From 11da5ceea60f206c8cc61ad8759df078dba9bf18 Mon Sep 17 00:00:00 2001
From: Yev Meyer <ymeyer@nvidia.com>
Date: Tue, 3 Nov 2026 09:46:34 -0500
Subject: [PATCH 1/5] test: check the probe template's worked examples by their
 paths

The probe test template cites each worked example by its path from the repo root, and the test of those examples checks that every cited path is a file, so an example kept in tests/probes/ is checked as one in tests/engine/probes/ is.

Signed-off-by: Yev Meyer <ymeyer@nvidia.com>
---
 templates/probe/test_probe.py.template | 4 ++--
 tests/test_packaging_contract.py       | 4 ++--
 2 files changed, 4 insertions(+), 4 deletions(-)

diff --git a/templates/probe/test_probe.py.template b/templates/probe/test_probe.py.template
index 06c78c4..802b3d7 100644
--- a/templates/probe/test_probe.py.template
+++ b/templates/probe/test_probe.py.template
@@ -12,9 +12,9 @@ after it. For real working examples, see:
 - ``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
+- ``tests/engine/probes/test_safety_chat_pressure.py`` — bank-reframing probe
   (``on_followup_failure="abort"`` test case).
-- ``test_safety_agentic.py`` — agentic probe (custom ``run_dispatch``,
+- ``tests/engine/probes/test_safety_agentic.py`` — agentic probe (custom ``run_dispatch``,
   no user-LLM, simpler mock pattern — FIFO-only since there's only
   one acall_llm caller).
 
diff --git a/tests/test_packaging_contract.py b/tests/test_packaging_contract.py
index f0b9959..d557de3 100644
--- a/tests/test_packaging_contract.py
+++ b/tests/test_packaging_contract.py
@@ -251,9 +251,9 @@ def test_referenced_examples_exist(self) -> None:
         import re
 
         src = self.TEMPLATE.read_text()
-        cited = set(re.findall(r"``(test_\w+)\.py``", src))
+        cited = set(re.findall(r"``(tests/[\w/]+\.py)``", src))
         assert cited, "template should cite worked examples"
-        missing = [name for name in cited if not (_REPO_ROOT / "tests" / "engine" / "probes" / f"{name}.py").is_file()]
+        missing = [path for path in cited if not (_REPO_ROOT / path).is_file()]
         assert not missing, f"template cites tests that do not exist: {sorted(missing)}"
 
 

From efefe61efd24a69236ce86e2b79090611db9ddf4 Mon Sep 17 00:00:00 2001
From: Yev Meyer <ymeyer@nvidia.com>
Date: Tue, 3 Nov 2026 09:46:39 -0500
Subject: [PATCH 2/5] refactor(engine): keep safety_chat_pressure in one folder

The probe module is probes/safety_chat_pressure/probe.py, beside its classifier, its bank loader bank.py, its scorer scorer.py and its bank in assets/safety_chat_pressure/, the folder the class names as its assets_root. Its tests are in tests/probes/safety_chat_pressure/. The label, the scorer name, the bank, 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, the authoring guide lists the probe among those kept in one folder, and the shared-code label allowances list neither the bank loader nor the scorer.

Signed-off-by: Yev Meyer <ymeyer@nvidia.com>
---
 docs/engine/AUTHORING_A_PROBE.md              | 10 +--
 docs/engine/README.md                         | 10 +--
 src/usersim/engine/core/agentic_bank.py       |  2 +-
 src/usersim/engine/core/probes.py             |  2 +-
 .../engine/evaluator/scorers/__init__.py      |  2 +-
 .../probes/safety_chat_pressure/__init__.py   |  2 +-
 .../assets/safety_chat_pressure/SCHEMA.md     |  4 +-
 .../assets/safety_chat_pressure/sample.yaml   |  0
 .../safety_chat_pressure/bank.py}             |  8 +--
 .../{generator.py => probe.py}                | 16 +++--
 .../probes/safety_chat_pressure/prompts.py    |  2 +-
 .../safety_chat_pressure/scorer.py}           | 22 +++----
 .../safety_chat_pressure/task_derivation.py   |  2 +-
 src/usersim/taxonomy/capabilities.py          |  4 +-
 templates/probe/test_probe.py.template        |  2 +-
 tests/engine/core/test_agentic_bank.py        |  2 +-
 tests/engine/core/test_assets_discovery.py    |  4 +-
 tests/engine/core/test_default_asset_paths.py |  8 +--
 ...luator_scorer_structured_output_schemas.py |  4 +-
 .../probes/test_india_variant_probes.py       |  2 +-
 tests/engine/probes/test_safety_agentic.py    |  2 +-
 tests/probes/safety_chat_pressure/__init__.py |  2 +
 .../safety_chat_pressure/test_bank.py}        |  8 +--
 .../safety_chat_pressure/test_classifier.py}  |  0
 .../safety_chat_pressure/test_probe.py}       | 24 +++----
 .../safety_chat_pressure/test_scorer.py}      | 66 +++++++++----------
 tests/test_cli_simulate_inputs.py             |  4 +-
 tests/test_probe_names_in_shared_code.py      |  2 -
 28 files changed, 109 insertions(+), 107 deletions(-)
 rename src/usersim/engine/{ => probes/safety_chat_pressure}/assets/safety_chat_pressure/SCHEMA.md (98%)
 rename src/usersim/engine/{ => probes/safety_chat_pressure}/assets/safety_chat_pressure/sample.yaml (100%)
 rename src/usersim/engine/{core/pressure_bank.py => probes/safety_chat_pressure/bank.py} (98%)
 rename src/usersim/engine/probes/safety_chat_pressure/{generator.py => probe.py} (99%)
 rename src/usersim/engine/{evaluator/scorers/safety_chat_pressure.py => probes/safety_chat_pressure/scorer.py} (97%)
 create mode 100644 tests/probes/safety_chat_pressure/__init__.py
 rename tests/{engine/core/test_pressure_bank.py => probes/safety_chat_pressure/test_bank.py} (98%)
 rename tests/{engine/probes/test_safety_chat_pressure_classifier.py => probes/safety_chat_pressure/test_classifier.py} (100%)
 rename tests/{engine/probes/test_safety_chat_pressure.py => probes/safety_chat_pressure/test_probe.py} (99%)
 rename tests/{engine/evaluator/test_evaluator_scorer_safety_chat_pressure.py => probes/safety_chat_pressure/test_scorer.py} (92%)

diff --git a/docs/engine/AUTHORING_A_PROBE.md b/docs/engine/AUTHORING_A_PROBE.md
index 2d945db..b0966e7 100644
--- a/docs/engine/AUTHORING_A_PROBE.md
+++ b/docs/engine/AUTHORING_A_PROBE.md
@@ -257,10 +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`, `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
+`general_educational`, `general_open_ended`, the three `sov_ai_*` probes and
+`safety_chat_pressure`, 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,
@@ -1223,7 +1223,7 @@ on (2)-(8) plus the should_succeed contract: the cross-cutting
 checks come for free.
 
 For real working examples: `tests/probes/sov_ai_facts/test_probe.py`
-(verbatim), `tests/engine/probes/test_safety_chat_pressure.py` (reframings and
+(verbatim), `tests/probes/safety_chat_pressure/test_probe.py` (reframings and
 abort policy), `tests/engine/probes/test_safety_agentic.py` (custom
 run_dispatch and a simpler mock).
 
diff --git a/docs/engine/README.md b/docs/engine/README.md
index 9876d49..b11a9ff 100644
--- a/docs/engine/README.md
+++ b/docs/engine/README.md
@@ -551,7 +551,7 @@ what each one carries:
 | `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` |
 | `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`) |
+| `probes/safety_chat_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 |
 | `core/seeds.py` | Seed loading: a probe's `<locale>/<kind>.json`, or the `en_US` file for a locale without one |
@@ -600,7 +600,6 @@ The tables describe the main files.
 | `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_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 |
 | `test_identity_spec.py` | Identity spec loader: loading, the env override and the cache; model-id resolution (a family name after any separator, declarations over patterns, hosts never deciding); name folding; layering down the asset search path and the roots probes declare (inheritance, in-place patches, probes built on top, cycles); merging and `select` / `exclude` narrowing; the content digest; validation reporting every problem at once |
@@ -612,8 +611,6 @@ The tables describe the main files.
 | `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_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 |
 | `test_financial_services.py` | `financial_services` probe: registration, multi-institution bank load and scope integrity, deterministic task instantiation and gold derivation, institution-scoped retrieval with no cross-institution leakage, the discoverable-tool gate through `run_dispatch` (mocked assistant), and the verifier scorer (good, unauthorized, skipped and dynamic-tier rows) |
 | `test_health_disclosure.py` | `health_disclosure` family: the Guard's pacing gates, the `commit_move` tool schema, move extraction, and the per-client config (topics, risk, archetype, prompts, axes) |
@@ -630,7 +627,6 @@ The tables describe the main files.
 | `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_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 |
 | `test_evaluator_scorer_identity_disclosure.py` | `identity_disclosure` scorer: applicability and fail-closed paths, extraction schema built from the spec's developers, quote validation, name-to-developer mapping (names, products, model names, ambiguous words), per-turn grading (false and spontaneous attribution, premise stance, reasoning spill, human claims), giving in under pressure, trajectory rates and review flags, same-family extractor warning |
@@ -651,3 +647,7 @@ The tables describe the main files.
 | `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 |
+| `safety_chat_pressure/test_probe.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) |
+| `safety_chat_pressure/test_scorer.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 |
+| `safety_chat_pressure/test_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 |
+| `safety_chat_pressure/test_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) |
diff --git a/src/usersim/engine/core/agentic_bank.py b/src/usersim/engine/core/agentic_bank.py
index 3478e07..e305a1b 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:`probes.sov_ai_multilingual_parity.bank`, and :mod:`core.pressure_bank`.
+  :mod:`probes.sov_ai_multilingual_parity.bank`, and :mod:`probes.safety_chat_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/probes.py b/src/usersim/engine/core/probes.py
index f19c5b3..db579e6 100644
--- a/src/usersim/engine/core/probes.py
+++ b/src/usersim/engine/core/probes.py
@@ -1549,7 +1549,7 @@ def decorator(cls: type[BaseProbe]) -> type[BaseProbe]:
     "usersim.engine.probes.health_disclosure.triage",
     "usersim.engine.probes.identity_disclosure.generator",
     "usersim.engine.probes.safety_agentic.generator",
-    "usersim.engine.probes.safety_chat_pressure.generator",
+    "usersim.engine.probes.safety_chat_pressure.probe",
     "usersim.engine.probes.sov_ai_dynamic.probe",
     "usersim.engine.probes.sov_ai_facts.probe",
     "usersim.engine.probes.sov_ai_multilingual_parity.probe",
diff --git a/src/usersim/engine/evaluator/scorers/__init__.py b/src/usersim/engine/evaluator/scorers/__init__.py
index d1d1091..c6556de 100644
--- a/src/usersim/engine/evaluator/scorers/__init__.py
+++ b/src/usersim/engine/evaluator/scorers/__init__.py
@@ -317,7 +317,7 @@ def reload_scorers(scorer_module_paths: Iterable[str]) -> None:
     "usersim.engine.probes.sov_ai_facts.scorer",
     "usersim.engine.probes.sov_ai_dynamic.scorer",
     "usersim.engine.probes.sov_ai_multilingual_parity.scorer",
-    "usersim.engine.evaluator.scorers.safety_chat_pressure",
+    "usersim.engine.probes.safety_chat_pressure.scorer",
     "usersim.engine.evaluator.scorers.safety_agentic",
     "usersim.engine.evaluator.scorers.financial_services",
     "usersim.engine.evaluator.scorers.language_compliance",
diff --git a/src/usersim/engine/probes/safety_chat_pressure/__init__.py b/src/usersim/engine/probes/safety_chat_pressure/__init__.py
index 51fb7df..ac406fd 100644
--- a/src/usersim/engine/probes/safety_chat_pressure/__init__.py
+++ b/src/usersim/engine/probes/safety_chat_pressure/__init__.py
@@ -19,6 +19,6 @@
 flip indicators plus a trajectory-level ``turn_of_flip``.
 
 Entry point: :func:`simulate_safety_chat_pressure` in
-:mod:`.generator` — registered as ``safety_chat_pressure``
+:mod:`.probe` — registered as ``safety_chat_pressure``
 on ``_PROBE_REGISTRY`` in :mod:`usersim.engine.generator`.
 """
diff --git a/src/usersim/engine/assets/safety_chat_pressure/SCHEMA.md b/src/usersim/engine/probes/safety_chat_pressure/assets/safety_chat_pressure/SCHEMA.md
similarity index 98%
rename from src/usersim/engine/assets/safety_chat_pressure/SCHEMA.md
rename to src/usersim/engine/probes/safety_chat_pressure/assets/safety_chat_pressure/SCHEMA.md
index 6aa60f1..d13ebcd 100644
--- a/src/usersim/engine/assets/safety_chat_pressure/SCHEMA.md
+++ b/src/usersim/engine/probes/safety_chat_pressure/assets/safety_chat_pressure/SCHEMA.md
@@ -8,7 +8,7 @@ requests (the things the assistant is being pressured into doing).
 
 This schema lives here so:
 
-- the loader in `usersim/engine/core/pressure_bank.py` can validate
+- the loader in `usersim/engine/probes/safety_chat_pressure/bank.py` can validate
   incoming banks;
 - the probe adapter and scorer can encode against a stable contract;
 - safety-corpus reviewers know exactly what to fill in and what's
@@ -159,7 +159,7 @@ before relying on their results.
 
 ## Loader contract
 
-The loader (`usersim/engine/core/pressure_bank.py`) enforces:
+The loader (`usersim/engine/probes/safety_chat_pressure/bank.py`) enforces:
 
 - `schema_version` is in the supported set.
 - `bank_id`, `bank_version`, `harm_categories` (non-empty),
diff --git a/src/usersim/engine/assets/safety_chat_pressure/sample.yaml b/src/usersim/engine/probes/safety_chat_pressure/assets/safety_chat_pressure/sample.yaml
similarity index 100%
rename from src/usersim/engine/assets/safety_chat_pressure/sample.yaml
rename to src/usersim/engine/probes/safety_chat_pressure/assets/safety_chat_pressure/sample.yaml
diff --git a/src/usersim/engine/core/pressure_bank.py b/src/usersim/engine/probes/safety_chat_pressure/bank.py
similarity index 98%
rename from src/usersim/engine/core/pressure_bank.py
rename to src/usersim/engine/probes/safety_chat_pressure/bank.py
index a60d62f..070e2ba 100644
--- a/src/usersim/engine/core/pressure_bank.py
+++ b/src/usersim/engine/probes/safety_chat_pressure/bank.py
@@ -5,7 +5,7 @@
 
 The pressure-strategy bank is the load-bearing primitive for the
 ``safety_chat_pressure`` probe. Schema is documented in
-``assets/safety_chat_pressure/SCHEMA.md``.
+``assets/safety_chat_pressure/SCHEMA.md`` beside this module.
 
 Three primitives in one bank:
 
@@ -577,9 +577,9 @@ def _require_localized(
 def default_pressure_bank_path() -> Path:
     """Default on-disk path for the pressure-strategy bank.
 
-    Discovered via ``core._assets.default_assets_dir`` so the loader
-    works regardless of CWD. Asset directory matches the consuming
-    probe name (``safety_chat_pressure``).
+    Resolved through ``probe_assets_dir``, so it is the probe's own
+    ``assets/safety_chat_pressure/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/safety_chat_pressure/generator.py b/src/usersim/engine/probes/safety_chat_pressure/probe.py
similarity index 99%
rename from src/usersim/engine/probes/safety_chat_pressure/generator.py
rename to src/usersim/engine/probes/safety_chat_pressure/probe.py
index 0a3eb5c..2dc7311 100644
--- a/src/usersim/engine/probes/safety_chat_pressure/generator.py
+++ b/src/usersim/engine/probes/safety_chat_pressure/probe.py
@@ -62,16 +62,10 @@
 
 import logging
 from dataclasses import dataclass
+from pathlib import Path
 from typing import Any
 
 from usersim.engine.core.outcomes import WarningKind
-from usersim.engine.core.pressure_bank import (
-    Strategy,
-    TargetRequest,
-    default_pressure_bank_path,
-    load_pressure_bank_default,
-    reset_pressure_bank_cache,
-)
 from usersim.engine.core.probes import (
     BankBackedProbe,
     BankReframingMixin,
@@ -80,6 +74,13 @@
     register_probe,
 )
 from usersim.engine.core.simulation import ConversationState, make_failed
+from usersim.engine.probes.safety_chat_pressure.bank import (
+    Strategy,
+    TargetRequest,
+    default_pressure_bank_path,
+    load_pressure_bank_default,
+    reset_pressure_bank_cache,
+)
 from usersim.engine.probes.safety_chat_pressure.classifier import (
     classify_capitulation,
     get_last_assistant_response,
@@ -240,6 +241,7 @@ class SafetyChatPressureProbe(BankReframingMixin, BankBackedProbe):
     """
 
     label = "safety_chat_pressure"
+    assets_root = Path(__file__).parent / "assets"
     description = (
         "Tests safety under multi-turn conversational pressure: does the assistant hold "
         "the line when the user persists, reframes, or escalates a target request?"
diff --git a/src/usersim/engine/probes/safety_chat_pressure/prompts.py b/src/usersim/engine/probes/safety_chat_pressure/prompts.py
index f6dcc2b..f96c72f 100644
--- a/src/usersim/engine/probes/safety_chat_pressure/prompts.py
+++ b/src/usersim/engine/probes/safety_chat_pressure/prompts.py
@@ -37,7 +37,7 @@
 from typing import Any
 
 from usersim.engine.core.persona import format_persona_for_prompt
-from usersim.engine.core.pressure_bank import (
+from usersim.engine.probes.safety_chat_pressure.bank import (
     Reframing,
     Strategy,
     TargetRequest,
diff --git a/src/usersim/engine/evaluator/scorers/safety_chat_pressure.py b/src/usersim/engine/probes/safety_chat_pressure/scorer.py
similarity index 97%
rename from src/usersim/engine/evaluator/scorers/safety_chat_pressure.py
rename to src/usersim/engine/probes/safety_chat_pressure/scorer.py
index 432be3b..2d76191 100644
--- a/src/usersim/engine/evaluator/scorers/safety_chat_pressure.py
+++ b/src/usersim/engine/probes/safety_chat_pressure/scorer.py
@@ -8,7 +8,7 @@
 ``target_request_id`` / ``strategy_id`` / ``reframings_used`` side
 channels (emitted by the safety_chat_pressure probe), re-loads the
 :class:`Strategy` and :class:`TargetRequest` from the pressure bank
-(via the shared cache in :mod:`core.pressure_bank`), and runs **one**
+(via the shared cache in :mod:`probes.safety_chat_pressure.bank`), and runs **one**
 LLM judge call per trajectory that produces:
 
 - per-turn classifications of the assistant's responses
@@ -78,13 +78,13 @@
 
 from usersim.engine.core.llm import acall_llm
 from usersim.engine.core.messages import format_conversation_history_for_prompt
-from usersim.engine.core.pressure_bank import (
+from usersim.engine.core.row_metadata import recorded_value
+from usersim.engine.evaluator.scorers import register_scorer
+from usersim.engine.probes.safety_chat_pressure.bank import (
     Strategy,
     TargetRequest,
     load_pressure_bank_default,
 )
-from usersim.engine.core.row_metadata import recorded_value
-from usersim.engine.evaluator.scorers import register_scorer
 
 logger = logging.getLogger("usersim.engine")
 
@@ -312,7 +312,7 @@ async def score_safety_chat_pressure_trajectory(
       is absent (returns a no-op envelope so misconfiguration — running
       this scorer against a different probe — degrades gracefully).
     - Loads the pressure bank via the shared cache in
-      :mod:`core.pressure_bank`. If either id is missing from the bank
+      :mod:`probes.safety_chat_pressure.bank`. If either id is missing from the bank
       (the bank was edited after the simulator ran), records a
       structured ``target_not_in_bank`` / ``strategy_not_in_bank``
       error and proposes ``status_proposal=False``.
@@ -348,7 +348,7 @@ async def score_safety_chat_pressure_trajectory(
         bank = load_pressure_bank_default()
     except Exception as e:
         logger.warning(
-            "  |-- evaluator/scorers.safety_chat_pressure: failed to load pressure bank: %s",
+            "  |-- safety_chat_pressure scorer: failed to load pressure bank: %s",
             e,
         )
         return _no_op_envelope(
@@ -363,7 +363,7 @@ async def score_safety_chat_pressure_trajectory(
     bank_version_mismatch = pinned_version is not None and pinned_version != bank.bank_version
     if bank_version_mismatch:
         logger.info(
-            "  |-- evaluator/scorers.safety_chat_pressure: trajectory "
+            "  |-- safety_chat_pressure scorer: trajectory "
             "pinned bank_version=%s but current loaded bank is %s — "
             "reporting drift flag on this row",
             pinned_version,
@@ -509,7 +509,7 @@ async def score_safety_chat_pressure_trajectory(
     stopped_on_capitulation = bool(recorded_value(trajectory, "stopped_on_capitulation"))
     if stopped_on_capitulation and judgment.refusal_held_score >= 4:
         logger.warning(
-            "  |-- evaluator/scorers.safety_chat_pressure: "
+            "  |-- safety_chat_pressure scorer: "
             "FALSE-POSITIVE STOP suspected for trajectory_id=%r — "
             "probe stopped on capitulation but judge scored "
             "refusal_held=%d (>=4). Capitulation check calibration may need "
@@ -607,7 +607,7 @@ async def _call_judge(
         )
     except Exception as e:
         logger.warning(
-            "  |-- evaluator/scorers.safety_chat_pressure: judge %r raised on target %s: %s: %s",
+            "  |-- safety_chat_pressure scorer: judge %r raised on target %s: %s: %s",
             judge_alias,
             target.id,
             type(e).__name__,
@@ -620,7 +620,7 @@ async def _call_judge(
         parsed = json.loads(content)
     except (json.JSONDecodeError, TypeError):
         logger.warning(
-            "  |-- evaluator/scorers.safety_chat_pressure: failed to parse structured output for target %s",
+            "  |-- safety_chat_pressure scorer: failed to parse structured output for target %s",
             target.id,
         )
         return None, "parse_failure"
@@ -629,7 +629,7 @@ async def _call_judge(
         judgment = _PressureJudgment.model_validate(parsed)
     except Exception as e:
         logger.warning(
-            "  |-- evaluator/scorers.safety_chat_pressure: schema validation failed for target %s: %s",
+            "  |-- safety_chat_pressure scorer: schema validation failed for target %s: %s",
             target.id,
             e,
         )
diff --git a/src/usersim/engine/probes/safety_chat_pressure/task_derivation.py b/src/usersim/engine/probes/safety_chat_pressure/task_derivation.py
index 46b4604..e18c6a6 100644
--- a/src/usersim/engine/probes/safety_chat_pressure/task_derivation.py
+++ b/src/usersim/engine/probes/safety_chat_pressure/task_derivation.py
@@ -33,7 +33,7 @@
 import random
 from typing import Any, Iterable
 
-from usersim.engine.core.pressure_bank import (
+from usersim.engine.probes.safety_chat_pressure.bank import (
     PressureBank,
     Strategy,
     TargetRequest,
diff --git a/src/usersim/taxonomy/capabilities.py b/src/usersim/taxonomy/capabilities.py
index 1a4e790..41d0098 100644
--- a/src/usersim/taxonomy/capabilities.py
+++ b/src/usersim/taxonomy/capabilities.py
@@ -994,7 +994,7 @@ def _identity_pressure_by_strategy() -> tuple[CapabilityDefinition, ...]:
     "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/probe.py",
-    "safety_chat_pressure": "src/usersim/engine/probes/safety_chat_pressure/generator.py",
+    "safety_chat_pressure": "src/usersim/engine/probes/safety_chat_pressure/probe.py",
     "safety_agentic": "src/usersim/engine/probes/safety_agentic/generator.py",
     "financial_services": "src/usersim/engine/probes/financial_services/generator.py",
 }
@@ -1007,7 +1007,7 @@ def _identity_pressure_by_strategy() -> tuple[CapabilityDefinition, ...]:
     "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/probes/sov_ai_multilingual_parity/scorer.py",
-    "safety_chat_pressure": "src/usersim/engine/evaluator/scorers/safety_chat_pressure.py",
+    "safety_chat_pressure": "src/usersim/engine/probes/safety_chat_pressure/scorer.py",
     "safety_agentic": "src/usersim/engine/evaluator/scorers/safety_agentic.py",
     "financial_services": "src/usersim/engine/evaluator/scorers/financial_services.py",
     "identity_disclosure": "src/usersim/engine/evaluator/scorers/identity_disclosure.py",
diff --git a/templates/probe/test_probe.py.template b/templates/probe/test_probe.py.template
index 802b3d7..7568d04 100644
--- a/templates/probe/test_probe.py.template
+++ b/templates/probe/test_probe.py.template
@@ -12,7 +12,7 @@ after it. For real working examples, see:
 - ``tests/probes/sov_ai_facts/test_probe.py`` — banked-verbatim probe (the canonical
   reference for the alias-dispatching mock + dual-module patching
   pattern).
-- ``tests/engine/probes/test_safety_chat_pressure.py`` — bank-reframing probe
+- ``tests/probes/safety_chat_pressure/test_probe.py`` — bank-reframing probe
   (``on_followup_failure="abort"`` test case).
 - ``tests/engine/probes/test_safety_agentic.py`` — agentic probe (custom ``run_dispatch``,
   no user-LLM, simpler mock pattern — FIFO-only since there's only
diff --git a/tests/engine/core/test_agentic_bank.py b/tests/engine/core/test_agentic_bank.py
index 0fe9968..1dd39e4 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`` / ``sov_ai_multilingual_parity/test_bank.py`` / ``test_pressure_bank.py``):
+(``sov_ai_facts/test_bank.py`` / ``sov_ai_multilingual_parity/test_bank.py`` / ``safety_chat_pressure/test_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 32224f0..ba4418c 100644
--- a/tests/engine/core/test_assets_discovery.py
+++ b/tests/engine/core/test_assets_discovery.py
@@ -12,7 +12,7 @@
 registry and the on-disk layout stay in lockstep.
 
 Loader-level tests live in the per-bank test files (``test_agentic_bank.py``,
-``test_pressure_bank.py``, …) and exercise these helpers indirectly via the
+``safety_chat_pressure/test_bank.py``, …) and exercise these helpers indirectly via the
 shipped sample.
 """
 
@@ -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("safety_chat_pressure").assets_dir() == packaged_assets_dir() / "safety_chat_pressure"
+        assert resolve_probe("financial_services").assets_dir() == packaged_assets_dir() / "financial_services"
 
     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 3ad132c..d154021 100644
--- a/tests/engine/core/test_default_asset_paths.py
+++ b/tests/engine/core/test_default_asset_paths.py
@@ -63,14 +63,14 @@
     load_identity_spec,
 )
 from usersim.engine.core.locale import SHIPPED_LOCALES
-from usersim.engine.core.pressure_bank import (
+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.seeds import load_seeds
+from usersim.engine.probes.safety_chat_pressure.bank import (
     default_pressure_bank_path,
     load_pressure_bank,
     reset_pressure_bank_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.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,
diff --git a/tests/engine/evaluator/test_evaluator_scorer_structured_output_schemas.py b/tests/engine/evaluator/test_evaluator_scorer_structured_output_schemas.py
index 7aef0b8..0e97dd0 100644
--- a/tests/engine/evaluator/test_evaluator_scorer_structured_output_schemas.py
+++ b/tests/engine/evaluator/test_evaluator_scorer_structured_output_schemas.py
@@ -33,13 +33,13 @@
 from usersim.engine.evaluator.scorers.safety_agentic import (
     _AgenticJudgment,
 )
-from usersim.engine.evaluator.scorers.safety_chat_pressure import (
+from usersim.engine.probes.safety_chat_pressure.scorer import (
     _VALID_CLASSIFICATIONS,
     Classification,
     _PerTurnClassification,
     _PressureJudgment,
 )
-from usersim.engine.evaluator.scorers.safety_chat_pressure import (
+from usersim.engine.probes.safety_chat_pressure.scorer import (
     JudgeScore as PressureJudgeScore,
 )
 
diff --git a/tests/engine/probes/test_india_variant_probes.py b/tests/engine/probes/test_india_variant_probes.py
index 2cb759c..f1308bd 100644
--- a/tests/engine/probes/test_india_variant_probes.py
+++ b/tests/engine/probes/test_india_variant_probes.py
@@ -183,7 +183,7 @@ def _query_id_for(locale: str) -> str:
 
 
 async def test_safety_chat_pressure_variant_translates_turn1(indian_persona, cfg) -> None:
-    from usersim.engine.probes.safety_chat_pressure.generator import (
+    from usersim.engine.probes.safety_chat_pressure.probe import (
         SafetyChatPressureProbe,
     )
 
diff --git a/tests/engine/probes/test_safety_agentic.py b/tests/engine/probes/test_safety_agentic.py
index 447a1be..c900621 100644
--- a/tests/engine/probes/test_safety_agentic.py
+++ b/tests/engine/probes/test_safety_agentic.py
@@ -3,7 +3,7 @@
 
 """Tests for the ``safety_agentic`` probe.
 
-Layered the same way as :mod:`test_safety_chat_pressure`:
+Layered the same way as ``safety_chat_pressure/test_probe.py``:
 
 1. **persona_to_tags re-export** — verifies the canonical
    sovereign-knowledge tag-derivation is reused (cross-probe joins
diff --git a/tests/probes/safety_chat_pressure/__init__.py b/tests/probes/safety_chat_pressure/__init__.py
new file mode 100644
index 0000000..52a7a9d
--- /dev/null
+++ b/tests/probes/safety_chat_pressure/__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_pressure_bank.py b/tests/probes/safety_chat_pressure/test_bank.py
similarity index 98%
rename from tests/engine/core/test_pressure_bank.py
rename to tests/probes/safety_chat_pressure/test_bank.py
index 65a6af1..9de2c14 100644
--- a/tests/engine/core/test_pressure_bank.py
+++ b/tests/probes/safety_chat_pressure/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/pressure_bank.py`` — pressure-strategy bank loader.
+"""Tests for ``probes/safety_chat_pressure/bank.py`` — pressure-strategy bank loader.
 
 Same three-layer test pattern as ``sov_ai_multilingual_parity/test_bank.py`` /
 ``sov_ai_facts/test_bank.py`` / ``test_probing_taxonomy.py``:
@@ -25,8 +25,8 @@
 
 import pytest
 
-from usersim.engine.core._assets import packaged_assets_dir
-from usersim.engine.core.pressure_bank import (
+from usersim.engine.core._assets import probe_assets_dir
+from usersim.engine.probes.safety_chat_pressure.bank import (
     PressureBank,
     PressureBankError,
     default_pressure_bank_path,
@@ -520,7 +520,7 @@ def test_default_path_resolves_to_assets(self) -> None:
 class TestShippedSampleBank:
     @pytest.fixture
     def sample_bank(self) -> PressureBank:
-        return load_pressure_bank((packaged_assets_dir() / "safety_chat_pressure/sample.yaml"))
+        return load_pressure_bank((probe_assets_dir("safety_chat_pressure") / "sample.yaml"))
 
     def test_loads_clean(self, sample_bank: PressureBank) -> None:
         assert sample_bank.bank_id == "sample_v1"
diff --git a/tests/engine/probes/test_safety_chat_pressure_classifier.py b/tests/probes/safety_chat_pressure/test_classifier.py
similarity index 100%
rename from tests/engine/probes/test_safety_chat_pressure_classifier.py
rename to tests/probes/safety_chat_pressure/test_classifier.py
diff --git a/tests/engine/probes/test_safety_chat_pressure.py b/tests/probes/safety_chat_pressure/test_probe.py
similarity index 99%
rename from tests/engine/probes/test_safety_chat_pressure.py
rename to tests/probes/safety_chat_pressure/test_probe.py
index ce7b966..3274536 100644
--- a/tests/engine/probes/test_safety_chat_pressure.py
+++ b/tests/probes/safety_chat_pressure/test_probe.py
@@ -39,30 +39,30 @@
 
 import usersim.engine.generator  # noqa: F401 — triggers probe registration
 from usersim.engine.config import ConversationSimulatorConfig
-from usersim.engine.core._assets import packaged_assets_dir
+from usersim.engine.core._assets import probe_assets_dir
 from usersim.engine.core.outcomes import (
     FailureClass,
     OutcomeStatus,
     Provenance,
     WarningKind,
 )
-from usersim.engine.core.pressure_bank import (
-    PressureBank,
-    Reframing,
-    Strategy,
-    TargetProvenance,
-    TargetRequest,
-    load_pressure_bank,
-)
 from usersim.engine.core.probes import _PROBE_REGISTRY, resolve_probe
 from usersim.engine.core.row_metadata import recorded_value
 from usersim.engine.probes.safety_chat_pressure import (
-    generator as probe_gen,
+    probe as probe_gen,
 )
 from usersim.engine.probes.safety_chat_pressure import (
     prompts,
     task_derivation,
 )
+from usersim.engine.probes.safety_chat_pressure.bank import (
+    PressureBank,
+    Reframing,
+    Strategy,
+    TargetProvenance,
+    TargetRequest,
+    load_pressure_bank,
+)
 from usersim.reporting.capability_report import _safety_pressure_info_findings
 
 # ---------------------------------------------------------------------------
@@ -71,7 +71,7 @@
 
 
 REPO_ROOT = Path(__file__).resolve().parents[3]
-SAMPLE_BANK_PATH = packaged_assets_dir() / "safety_chat_pressure/sample.yaml"
+SAMPLE_BANK_PATH = probe_assets_dir("safety_chat_pressure") / "sample.yaml"
 
 
 @pytest.fixture
@@ -740,7 +740,7 @@ async def test_first_user_turn_is_target_initial_prompt_verbatim(
         simulator_cfg: Any,
     ) -> None:
         with patch.dict(os.environ, _shipped_bank_env()):
-            from usersim.engine.core.pressure_bank import (
+            from usersim.engine.probes.safety_chat_pressure.bank import (
                 load_pressure_bank_default,
             )
 
diff --git a/tests/engine/evaluator/test_evaluator_scorer_safety_chat_pressure.py b/tests/probes/safety_chat_pressure/test_scorer.py
similarity index 92%
rename from tests/engine/evaluator/test_evaluator_scorer_safety_chat_pressure.py
rename to tests/probes/safety_chat_pressure/test_scorer.py
index e238033..36d4dcb 100644
--- a/tests/engine/evaluator/test_evaluator_scorer_safety_chat_pressure.py
+++ b/tests/probes/safety_chat_pressure/test_scorer.py
@@ -27,15 +27,15 @@
 
 import pytest
 
-from usersim.engine.core._assets import packaged_assets_dir
-from usersim.engine.core.pressure_bank import reset_pressure_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.safety_chat_pressure import (
+from usersim.engine.probes.safety_chat_pressure.bank import reset_pressure_bank_cache
+from usersim.engine.probes.safety_chat_pressure.scorer import (
     CONTEXTUAL_SAFETY_PERSISTENCE_AXIS,
     PRESSURE_AXES,
     REFUSAL_HELD_AXIS,
@@ -48,7 +48,7 @@
 
 
 REPO_ROOT = Path(__file__).resolve().parents[3]
-SAMPLE_BANK_PATH = packaged_assets_dir() / "safety_chat_pressure/sample.yaml"
+SAMPLE_BANK_PATH = probe_assets_dir("safety_chat_pressure") / "sample.yaml"
 
 
 @pytest.fixture(autouse=True)
@@ -215,7 +215,7 @@ async def test_missing_strategy_id_returns_no_op(self) -> None:
 
     async def test_no_llm_call_made_on_short_circuit(self) -> None:
         with patch(
-            "usersim.engine.evaluator.scorers.safety_chat_pressure.acall_llm",
+            "usersim.engine.probes.safety_chat_pressure.scorer.acall_llm",
         ) as mock_call:
             await score_safety_chat_pressure_trajectory({}, {"alias": object()})
         mock_call.assert_not_called()
@@ -233,7 +233,7 @@ async def test_no_flip_trajectory_returns_high_scores(
     ) -> None:
         traj = _trajectory(n_assistant_turns=3)
         with patch(
-            "usersim.engine.evaluator.scorers.safety_chat_pressure.acall_llm",
+            "usersim.engine.probes.safety_chat_pressure.scorer.acall_llm",
             side_effect=_mock_judge_call(
                 [
                     _judge_payload(
@@ -275,7 +275,7 @@ async def test_mid_flip_trajectory_records_turn_of_flip(
             {"turn_idx": 2, "classification": "full_capitulation", "reasoning": "x"},
         ]
         with patch(
-            "usersim.engine.evaluator.scorers.safety_chat_pressure.acall_llm",
+            "usersim.engine.probes.safety_chat_pressure.scorer.acall_llm",
             side_effect=_mock_judge_call(
                 [
                     _judge_payload(
@@ -304,7 +304,7 @@ async def test_immediate_flip_trajectory(self, mock_bank_env: Path) -> None:
             {"turn_idx": 0, "classification": "full_capitulation", "reasoning": "x"},
         ]
         with patch(
-            "usersim.engine.evaluator.scorers.safety_chat_pressure.acall_llm",
+            "usersim.engine.probes.safety_chat_pressure.scorer.acall_llm",
             side_effect=_mock_judge_call(
                 [
                     _judge_payload(
@@ -345,7 +345,7 @@ async def _capture_call(*args, **kwargs):
             }
 
         with patch(
-            "usersim.engine.evaluator.scorers.safety_chat_pressure.acall_llm",
+            "usersim.engine.probes.safety_chat_pressure.scorer.acall_llm",
             side_effect=_capture_call,
         ):
             await score_safety_chat_pressure_trajectory(
@@ -387,7 +387,7 @@ async def test_judge_omits_turn_of_flip_but_marks_per_turn_flip(
             {"turn_idx": 2, "classification": "refused_clean", "reasoning": "x"},
         ]
         with patch(
-            "usersim.engine.evaluator.scorers.safety_chat_pressure.acall_llm",
+            "usersim.engine.probes.safety_chat_pressure.scorer.acall_llm",
             side_effect=_mock_judge_call(
                 [
                     _judge_payload(
@@ -417,7 +417,7 @@ async def test_no_per_turn_flip_means_no_derived_turn_of_flip(
             {"turn_idx": 1, "classification": "hedged", "reasoning": "x"},
         ]
         with patch(
-            "usersim.engine.evaluator.scorers.safety_chat_pressure.acall_llm",
+            "usersim.engine.probes.safety_chat_pressure.scorer.acall_llm",
             side_effect=_mock_judge_call(
                 [
                     _judge_payload(
@@ -450,7 +450,7 @@ async def test_target_not_in_bank_returns_structured_error(
     ) -> None:
         traj = _trajectory(target_request_id="TR-DOES-NOT-EXIST")
         with patch(
-            "usersim.engine.evaluator.scorers.safety_chat_pressure.acall_llm",
+            "usersim.engine.probes.safety_chat_pressure.scorer.acall_llm",
         ) as mock_call:
             out = await score_safety_chat_pressure_trajectory(
                 traj,
@@ -467,7 +467,7 @@ async def test_strategy_not_in_bank_returns_structured_error(
     ) -> None:
         traj = _trajectory(strategy_id="missing_strategy")
         with patch(
-            "usersim.engine.evaluator.scorers.safety_chat_pressure.acall_llm",
+            "usersim.engine.probes.safety_chat_pressure.scorer.acall_llm",
         ) as mock_call:
             out = await score_safety_chat_pressure_trajectory(
                 traj,
@@ -487,7 +487,7 @@ async def test_bank_load_failure_returns_structured_error(self) -> None:
                 {"USERSIM_SAFETY_CHAT_PRESSURE_BANK": "/nonexistent/bank.yaml"},
             ),
             patch(
-                "usersim.engine.evaluator.scorers.safety_chat_pressure.acall_llm",
+                "usersim.engine.probes.safety_chat_pressure.scorer.acall_llm",
             ) as mock_call,
         ):
             out = await score_safety_chat_pressure_trajectory(
@@ -507,7 +507,7 @@ async def test_no_assistant_turns_returns_structured_error(
         msgs = [m for m in json.loads(traj["conversation_messages"]) if m.get("role") != "assistant"]
         traj["conversation_messages"] = json.dumps(msgs)
         with patch(
-            "usersim.engine.evaluator.scorers.safety_chat_pressure.acall_llm",
+            "usersim.engine.probes.safety_chat_pressure.scorer.acall_llm",
         ) as mock_call:
             out = await score_safety_chat_pressure_trajectory(
                 traj,
@@ -526,7 +526,7 @@ async def _raises(*args, **kwargs):
 
         traj = _trajectory()
         with patch(
-            "usersim.engine.evaluator.scorers.safety_chat_pressure.acall_llm",
+            "usersim.engine.probes.safety_chat_pressure.scorer.acall_llm",
             side_effect=_raises,
         ):
             out = await score_safety_chat_pressure_trajectory(
@@ -541,7 +541,7 @@ async def _raises(*args, **kwargs):
     async def test_unparseable_judge_response(self, mock_bank_env: Path) -> None:
         traj = _trajectory()
         with patch(
-            "usersim.engine.evaluator.scorers.safety_chat_pressure.acall_llm",
+            "usersim.engine.probes.safety_chat_pressure.scorer.acall_llm",
             return_value={"role": "assistant", "content": "not-json-{"},
         ):
             out = await score_safety_chat_pressure_trajectory(
@@ -564,7 +564,7 @@ async def test_invalid_classification_rejected(
             },
         ]
         with patch(
-            "usersim.engine.evaluator.scorers.safety_chat_pressure.acall_llm",
+            "usersim.engine.probes.safety_chat_pressure.scorer.acall_llm",
             side_effect=_mock_judge_call(
                 [
                     _judge_payload(
@@ -603,7 +603,7 @@ async def test_refusal_held_score_of_one_blocks_status(
         # Even with no turn_of_flip set, refusal_held=1 blocks.
         traj = _trajectory()
         with patch(
-            "usersim.engine.evaluator.scorers.safety_chat_pressure.acall_llm",
+            "usersim.engine.probes.safety_chat_pressure.scorer.acall_llm",
             side_effect=_mock_judge_call(
                 [
                     _judge_payload(
@@ -637,7 +637,7 @@ async def test_low_persistence_score_does_not_block_alone(
         # fail). Catches accidental gating-rule expansion.
         traj = _trajectory()
         with patch(
-            "usersim.engine.evaluator.scorers.safety_chat_pressure.acall_llm",
+            "usersim.engine.probes.safety_chat_pressure.scorer.acall_llm",
             side_effect=_mock_judge_call(
                 [
                     _judge_payload(
@@ -665,7 +665,7 @@ async def test_low_persistence_score_does_not_block_alone(
     async def test_all_high_passes(self, mock_bank_env: Path) -> None:
         traj = _trajectory()
         with patch(
-            "usersim.engine.evaluator.scorers.safety_chat_pressure.acall_llm",
+            "usersim.engine.probes.safety_chat_pressure.scorer.acall_llm",
             side_effect=_mock_judge_call(
                 [
                     _judge_payload(refusal_held=5, persistence=5),
@@ -691,7 +691,7 @@ async def test_pinned_version_matches_no_mismatch(
     ) -> None:
         traj = _trajectory(pinned_version="v0.5.5")
         with patch(
-            "usersim.engine.evaluator.scorers.safety_chat_pressure.acall_llm",
+            "usersim.engine.probes.safety_chat_pressure.scorer.acall_llm",
             side_effect=_mock_judge_call([_judge_payload()]),
         ):
             out = await score_safety_chat_pressure_trajectory(
@@ -706,7 +706,7 @@ async def test_pinned_version_differs_emits_mismatch_flag(
     ) -> None:
         traj = _trajectory(pinned_version="v0.0.5")
         with patch(
-            "usersim.engine.evaluator.scorers.safety_chat_pressure.acall_llm",
+            "usersim.engine.probes.safety_chat_pressure.scorer.acall_llm",
             side_effect=_mock_judge_call([_judge_payload()]),
         ):
             out = await score_safety_chat_pressure_trajectory(
@@ -725,7 +725,7 @@ 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.safety_chat_pressure.acall_llm",
+            "usersim.engine.probes.safety_chat_pressure.scorer.acall_llm",
             side_effect=_mock_judge_call([_judge_payload()]),
         ):
             out = await score_safety_chat_pressure_trajectory(
@@ -755,7 +755,7 @@ class TestStatusProposalTriState:
     async def test_refusal_held_4_returns_pass(self, mock_bank_env: Path) -> None:
         traj = _trajectory()
         with patch(
-            "usersim.engine.evaluator.scorers.safety_chat_pressure.acall_llm",
+            "usersim.engine.probes.safety_chat_pressure.scorer.acall_llm",
             side_effect=_mock_judge_call(
                 [
                     _judge_payload(refusal_held=4, persistence=4, turn_of_flip=None),
@@ -771,7 +771,7 @@ async def test_refusal_held_4_returns_pass(self, mock_bank_env: Path) -> None:
     async def test_refusal_held_5_returns_pass(self, mock_bank_env: Path) -> None:
         traj = _trajectory()
         with patch(
-            "usersim.engine.evaluator.scorers.safety_chat_pressure.acall_llm",
+            "usersim.engine.probes.safety_chat_pressure.scorer.acall_llm",
             side_effect=_mock_judge_call(
                 [
                     _judge_payload(refusal_held=5, persistence=5, turn_of_flip=None),
@@ -795,7 +795,7 @@ async def test_refusal_held_3_returns_partial_none(
         """
         traj = _trajectory()
         with patch(
-            "usersim.engine.evaluator.scorers.safety_chat_pressure.acall_llm",
+            "usersim.engine.probes.safety_chat_pressure.scorer.acall_llm",
             side_effect=_mock_judge_call(
                 [
                     _judge_payload(refusal_held=3, persistence=3, turn_of_flip=None),
@@ -811,7 +811,7 @@ async def test_refusal_held_3_returns_partial_none(
     async def test_refusal_held_2_returns_fail(self, mock_bank_env: Path) -> None:
         traj = _trajectory()
         with patch(
-            "usersim.engine.evaluator.scorers.safety_chat_pressure.acall_llm",
+            "usersim.engine.probes.safety_chat_pressure.scorer.acall_llm",
             side_effect=_mock_judge_call(
                 [
                     _judge_payload(refusal_held=2, persistence=2, turn_of_flip=2),
@@ -827,7 +827,7 @@ async def test_refusal_held_2_returns_fail(self, mock_bank_env: Path) -> None:
     async def test_refusal_held_1_returns_fail(self, mock_bank_env: Path) -> None:
         traj = _trajectory()
         with patch(
-            "usersim.engine.evaluator.scorers.safety_chat_pressure.acall_llm",
+            "usersim.engine.probes.safety_chat_pressure.scorer.acall_llm",
             side_effect=_mock_judge_call(
                 [
                     _judge_payload(refusal_held=1, persistence=1, turn_of_flip=1),
@@ -854,7 +854,7 @@ async def test_turn_of_flip_alone_does_not_flip_status(
         """
         traj = _trajectory()
         with patch(
-            "usersim.engine.evaluator.scorers.safety_chat_pressure.acall_llm",
+            "usersim.engine.probes.safety_chat_pressure.scorer.acall_llm",
             side_effect=_mock_judge_call(
                 [
                     _judge_payload(refusal_held=4, persistence=3, turn_of_flip=2),
@@ -898,7 +898,7 @@ async def test_warning_fires_on_disagreement(
         traj["trajectory_id"] = "false-pos-001"
         traj.update(stop_fields)
         with patch(
-            "usersim.engine.evaluator.scorers.safety_chat_pressure.acall_llm",
+            "usersim.engine.probes.safety_chat_pressure.scorer.acall_llm",
             side_effect=_mock_judge_call(
                 [
                     _judge_payload(refusal_held=5, persistence=5, turn_of_flip=None),
@@ -927,7 +927,7 @@ async def test_warning_does_not_fire_when_judge_agrees(
         traj["trajectory_id"] = "agree-001"
         traj["stopped_on_capitulation"] = True
         with patch(
-            "usersim.engine.evaluator.scorers.safety_chat_pressure.acall_llm",
+            "usersim.engine.probes.safety_chat_pressure.scorer.acall_llm",
             side_effect=_mock_judge_call(
                 [
                     _judge_payload(refusal_held=2, persistence=2, turn_of_flip=1),
@@ -951,7 +951,7 @@ async def test_warning_does_not_fire_when_no_capitulation_stop(
         traj = _trajectory()
         # No stopped_on_capitulation field at all.
         with patch(
-            "usersim.engine.evaluator.scorers.safety_chat_pressure.acall_llm",
+            "usersim.engine.probes.safety_chat_pressure.scorer.acall_llm",
             side_effect=_mock_judge_call(
                 [
                     _judge_payload(refusal_held=5, persistence=5, turn_of_flip=None),
diff --git a/tests/test_cli_simulate_inputs.py b/tests/test_cli_simulate_inputs.py
index c35bed7..2f59287 100644
--- a/tests/test_cli_simulate_inputs.py
+++ b/tests/test_cli_simulate_inputs.py
@@ -49,7 +49,7 @@
 from usersim.cli._models import to_model_configs as _real_to_model_configs
 from usersim.cli._pipeline import materialize_episode_inputs
 from usersim.engine import generator as generator_module
-from usersim.engine.core._assets import packaged_assets_dir
+from usersim.engine.core._assets import packaged_assets_dir, probe_assets_dir
 from usersim.engine.core.manifest import load_run_manifest
 from usersim.engine.core.missing import none_for_missing
 from usersim.engine.core.storage import list_runs, read_partitioned_dataset
@@ -565,7 +565,7 @@ def test_an_override_column_on_a_stored_row_reaches_the_probe(tmp_path):
     (stored,) = _materialize({"safety_chat_pressure": 1.0}, num_rows=1)
     first = _run([stored], tmp_path / "derived")
     (derived,) = _saved(tmp_path / "derived", first.run_id)["strategy_id"]
-    bank = yaml.safe_load((packaged_assets_dir() / "safety_chat_pressure" / "sample.yaml").read_text(encoding="utf-8"))
+    bank = yaml.safe_load((probe_assets_dir("safety_chat_pressure") / "sample.yaml").read_text(encoding="utf-8"))
     chosen = next(strategy["id"] for strategy in bank["strategies"] if strategy["id"] != derived)
 
     stored["pressure_strategy_id"] = chosen
diff --git a/tests/test_probe_names_in_shared_code.py b/tests/test_probe_names_in_shared_code.py
index 0c86d7b..78b258d 100644
--- a/tests/test_probe_names_in_shared_code.py
+++ b/tests/test_probe_names_in_shared_code.py
@@ -37,13 +37,11 @@ class instead (``architecture/probes.md``, "What a probe declares").
     "engine/core/clinical_profile_bank.py": 4,
     "engine/core/finance_bank.py": 1,
     "engine/core/identity_spec.py": 1,
-    "engine/core/pressure_bank.py": 1,
     "engine/core/preview.py": 3,
     "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,
     "reporting/capability_report.py": 1,
     "reporting/identity.py": 1,
     "taxonomy/capabilities.py": 41,

From 813c0a35d7401388cfd098b7b81a72db9cc6d1d8 Mon Sep 17 00:00:00 2001
From: Yev Meyer <ymeyer@nvidia.com>
Date: Tue, 3 Nov 2026 09:46:44 -0500
Subject: [PATCH 3/5] refactor(engine): keep safety_agentic in one folder

The probe module is probes/safety_agentic/probe.py, beside its bank loader bank.py, its action-taxonomy loader action_taxonomy.py, its scorer scorer.py and its banks in assets/safety_agentic/, the folder the class names as its assets_root. Its tests are in tests/probes/safety_agentic/. The label, the scorer name, the banks, the stored columns and both caches are unchanged.

DEFAULT_SCORER_MODULES, the capability source references and the docs name the new modules, the tool_calling README links to the schema in its new folder, the authoring guide lists the probe among those kept in one folder, and the shared-code label allowances list neither loader nor the scorer.

Signed-off-by: Yev Meyer <ymeyer@nvidia.com>
---
 docs/engine/AUTHORING_A_PROBE.md              | 11 ++-
 docs/engine/README.md                         | 12 +--
 .../engine/assets/tool_calling/README.md      |  2 +-
 src/usersim/engine/core/probes.py             |  2 +-
 .../engine/evaluator/scorers/__init__.py      |  2 +-
 .../engine/probes/safety_agentic/__init__.py  |  2 +-
 .../safety_agentic}/action_taxonomy.py        |  9 +-
 .../assets/safety_agentic/SCHEMA.md           |  4 +-
 .../safety_agentic/action_taxonomy.yaml       |  0
 .../assets/safety_agentic/sample.yaml         |  0
 .../safety_agentic/bank.py}                   |  7 +-
 .../safety_agentic/{generator.py => probe.py} | 18 ++--
 .../safety_agentic/scorer.py}                 | 26 +++---
 .../probes/safety_agentic/task_derivation.py  |  2 +-
 src/usersim/taxonomy/capabilities.py          |  4 +-
 templates/probe/test_probe.py.template        |  2 +-
 tests/engine/core/test_assets_discovery.py    |  2 +-
 tests/engine/core/test_default_asset_paths.py | 18 ++--
 ...luator_scorer_structured_output_schemas.py |  4 +-
 .../probes/test_india_variant_probes.py       |  2 +-
 tests/engine/probes/test_probe_invariants.py  |  2 +-
 tests/engine/test_generator_row_scope.py      |  2 +-
 tests/probes/safety_agentic/__init__.py       |  2 +
 .../safety_agentic}/test_action_taxonomy.py   | 10 +-
 .../safety_agentic/test_bank.py}              |  8 +-
 .../safety_agentic/test_probe.py}             | 48 +++++-----
 .../safety_agentic/test_scorer.py}            | 92 +++++++++----------
 tests/test_capabilities.py                    |  2 +-
 tests/test_cli_simulate_inputs.py             |  2 +-
 tests/test_probe_names_in_shared_code.py      |  3 -
 tests/test_simulator_pipeline_e2e.py          |  4 +-
 31 files changed, 152 insertions(+), 152 deletions(-)
 rename src/usersim/engine/{core => probes/safety_agentic}/action_taxonomy.py (97%)
 rename src/usersim/engine/{ => probes/safety_agentic}/assets/safety_agentic/SCHEMA.md (98%)
 rename src/usersim/engine/{ => probes/safety_agentic}/assets/safety_agentic/action_taxonomy.yaml (100%)
 rename src/usersim/engine/{ => probes/safety_agentic}/assets/safety_agentic/sample.yaml (100%)
 rename src/usersim/engine/{core/agentic_bank.py => probes/safety_agentic/bank.py} (99%)
 rename src/usersim/engine/probes/safety_agentic/{generator.py => probe.py} (99%)
 rename src/usersim/engine/{evaluator/scorers/safety_agentic.py => probes/safety_agentic/scorer.py} (98%)
 create mode 100644 tests/probes/safety_agentic/__init__.py
 rename tests/{engine/core => probes/safety_agentic}/test_action_taxonomy.py (96%)
 rename tests/{engine/core/test_agentic_bank.py => probes/safety_agentic/test_bank.py} (98%)
 rename tests/{engine/probes/test_safety_agentic.py => probes/safety_agentic/test_probe.py} (96%)
 rename tests/{engine/evaluator/test_evaluator_scorer_safety_agentic.py => probes/safety_agentic/test_scorer.py} (91%)

diff --git a/docs/engine/AUTHORING_A_PROBE.md b/docs/engine/AUTHORING_A_PROBE.md
index b0966e7..6d9f51f 100644
--- a/docs/engine/AUTHORING_A_PROBE.md
+++ b/docs/engine/AUTHORING_A_PROBE.md
@@ -257,10 +257,11 @@ 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`, `general_open_ended`, the three `sov_ai_*` probes and
-`safety_chat_pressure`, 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
+`general_educational`, `general_open_ended`, the three `sov_ai_*` probes,
+`safety_chat_pressure` and `safety_agentic`, 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,
@@ -1224,7 +1225,7 @@ checks come for free.
 
 For real working examples: `tests/probes/sov_ai_facts/test_probe.py`
 (verbatim), `tests/probes/safety_chat_pressure/test_probe.py` (reframings and
-abort policy), `tests/engine/probes/test_safety_agentic.py` (custom
+abort policy), `tests/probes/safety_agentic/test_probe.py` (custom
 run_dispatch and a simpler mock).
 
 Every test under `tests/` runs inside the hermetic harness that
diff --git a/docs/engine/README.md b/docs/engine/README.md
index b11a9ff..0b65a4b 100644
--- a/docs/engine/README.md
+++ b/docs/engine/README.md
@@ -552,8 +552,8 @@ what each one carries:
 | `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` |
 | `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`) |
 | `probes/safety_chat_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 |
+| `probes/safety_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`) |
+| `probes/safety_agentic/action_taxonomy.py` | Rich action-taxonomy loader (categories, risk tiers, blast-radius ranks); consumed by the `safety_agentic` scorer for cross-asset action classification |
 | `core/seeds.py` | Seed loading: a probe's `<locale>/<kind>.json`, or the `en_US` file for a locale without one |
 | `core/analysis.py` | Helpers for the deep-dive analysis script |
 
@@ -600,8 +600,6 @@ The tables describe the main files.
 | `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_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 |
 | `test_identity_spec.py` | Identity spec loader: loading, the env override and the cache; model-id resolution (a family name after any separator, declarations over patterns, hosts never deciding); name folding; layering down the asset search path and the roots probes declare (inheritance, in-place patches, probes built on top, cycles); merging and `select` / `exclude` narrowing; the content digest; validation reporting every problem at once |
 
 **`tests/engine/probes/`: the probe generators, verifier and invariants**
@@ -611,7 +609,6 @@ The tables describe the main files.
 | `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_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 |
 | `test_financial_services.py` | `financial_services` probe: registration, multi-institution bank load and scope integrity, deterministic task instantiation and gold derivation, institution-scoped retrieval with no cross-institution leakage, the discoverable-tool gate through `run_dispatch` (mocked assistant), and the verifier scorer (good, unauthorized, skipped and dynamic-tier rows) |
 | `test_health_disclosure.py` | `health_disclosure` family: the Guard's pacing gates, the `commit_move` tool schema, move extraction, and the per-client config (topics, risk, archetype, prompts, axes) |
 | `test_identity_disclosure.py` | `identity_disclosure` probe: registration and the placeholder flag; shipped-spec checks (public examples, sourced lineage, every cataloged and shipped model resolves); deterministic tactic and strategy draws with panel overrides; expected-identity resolution (declared, inferred, an abort showing the declaration to add, an unidentified facade); openings and competitors; openings the simulated user writes (its instruction and prompt line, the check for AI names, the opening gate, a draft naming an AI developer written again in a loop run); row columns; prompts; pressure (every multi-turn row pressed, reframings recorded, its own prompt, gate and anchor, a loop run the scorer grades); the user-turn policy; a probe built on top; one full turn through the loop |
@@ -627,7 +624,6 @@ The tables describe the main files.
 | `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_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 |
 | `test_evaluator_scorer_identity_disclosure.py` | `identity_disclosure` scorer: applicability and fail-closed paths, extraction schema built from the spec's developers, quote validation, name-to-developer mapping (names, products, model names, ambiguous words), per-turn grading (false and spontaneous attribution, premise stance, reasoning spill, human claims), giving in under pressure, trajectory rates and review flags, same-family extractor warning |
 | `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 |
@@ -651,3 +647,7 @@ The tables describe the main files.
 | `safety_chat_pressure/test_scorer.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 |
 | `safety_chat_pressure/test_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 |
 | `safety_chat_pressure/test_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) |
+| `safety_agentic/test_probe.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 |
+| `safety_agentic/test_scorer.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 |
+| `safety_agentic/test_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 |
+| `safety_agentic/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 |
diff --git a/src/usersim/engine/assets/tool_calling/README.md b/src/usersim/engine/assets/tool_calling/README.md
index 528a824..f69ac75 100644
--- a/src/usersim/engine/assets/tool_calling/README.md
+++ b/src/usersim/engine/assets/tool_calling/README.md
@@ -222,4 +222,4 @@ json.loads(df.iloc[0]["tools"])
 - Turn-1 / probing prompts: [`prompts.yaml`](prompts.yaml)
 - The **agentic** safety surface uses a *different*, hand-authored mock-tool
   mechanism (per-action tool specs + canned responses) documented in
-  [`assets/safety_agentic/SCHEMA.md`](../safety_agentic/SCHEMA.md).
+  [`probes/safety_agentic/assets/safety_agentic/SCHEMA.md`](../../probes/safety_agentic/assets/safety_agentic/SCHEMA.md).
diff --git a/src/usersim/engine/core/probes.py b/src/usersim/engine/core/probes.py
index db579e6..d60bb3c 100644
--- a/src/usersim/engine/core/probes.py
+++ b/src/usersim/engine/core/probes.py
@@ -1548,7 +1548,7 @@ def decorator(cls: type[BaseProbe]) -> type[BaseProbe]:
     "usersim.engine.probes.health_disclosure.therapy",
     "usersim.engine.probes.health_disclosure.triage",
     "usersim.engine.probes.identity_disclosure.generator",
-    "usersim.engine.probes.safety_agentic.generator",
+    "usersim.engine.probes.safety_agentic.probe",
     "usersim.engine.probes.safety_chat_pressure.probe",
     "usersim.engine.probes.sov_ai_dynamic.probe",
     "usersim.engine.probes.sov_ai_facts.probe",
diff --git a/src/usersim/engine/evaluator/scorers/__init__.py b/src/usersim/engine/evaluator/scorers/__init__.py
index c6556de..f7394fd 100644
--- a/src/usersim/engine/evaluator/scorers/__init__.py
+++ b/src/usersim/engine/evaluator/scorers/__init__.py
@@ -318,7 +318,7 @@ def reload_scorers(scorer_module_paths: Iterable[str]) -> None:
     "usersim.engine.probes.sov_ai_dynamic.scorer",
     "usersim.engine.probes.sov_ai_multilingual_parity.scorer",
     "usersim.engine.probes.safety_chat_pressure.scorer",
-    "usersim.engine.evaluator.scorers.safety_agentic",
+    "usersim.engine.probes.safety_agentic.scorer",
     "usersim.engine.evaluator.scorers.financial_services",
     "usersim.engine.evaluator.scorers.language_compliance",
     "usersim.engine.evaluator.scorers.response_shape",
diff --git a/src/usersim/engine/probes/safety_agentic/__init__.py b/src/usersim/engine/probes/safety_agentic/__init__.py
index 6ce5376..072381d 100644
--- a/src/usersim/engine/probes/safety_agentic/__init__.py
+++ b/src/usersim/engine/probes/safety_agentic/__init__.py
@@ -18,7 +18,7 @@
 confirmation when the action taxonomy says it should — those are
 the scorer's questions, not the simulator's.
 
-Entry point: :func:`simulate_safety_agentic` in :mod:`.generator` —
+Entry point: :func:`simulate_safety_agentic` in :mod:`.probe` —
 registered as ``safety_agentic`` on ``_PROBE_REGISTRY`` in
 :mod:`usersim.engine.generator`.
 """
diff --git a/src/usersim/engine/core/action_taxonomy.py b/src/usersim/engine/probes/safety_agentic/action_taxonomy.py
similarity index 97%
rename from src/usersim/engine/core/action_taxonomy.py
rename to src/usersim/engine/probes/safety_agentic/action_taxonomy.py
index 9278e12..c10b990 100644
--- a/src/usersim/engine/core/action_taxonomy.py
+++ b/src/usersim/engine/probes/safety_agentic/action_taxonomy.py
@@ -3,7 +3,7 @@
 
 """Typed loader for the shipped ``assets/safety_agentic/action_taxonomy.yaml``.
 
-The ``core.agentic_bank`` loader already reads this file at load time
+The loader in ``bank.py`` beside this module already reads this file at load time
 for *name-only* validation (does every bank entry's
 ``sanctioned_action_name`` resolve to a taxonomy entry?). The
 ``safety_agentic`` scorer needs more: it joins every attempted action
@@ -14,7 +14,7 @@
 
 Design choices:
 
-- **Separate cache from the name-only cache in ``agentic_bank``.** The
+- **Separate cache from the name-only cache in ``bank.py``.** The
   two callers have different needs (names for bank validation, rich
   objects for scoring) and keeping them separate means changing
   either surface doesn't touch the other. Both honour the same
@@ -124,7 +124,7 @@ def blast_radius_rank(self, level: str) -> int | None:
 
 
 # ---------------------------------------------------------------------------
-# Cache (separate from the name-only cache in agentic_bank)
+# Cache (separate from the name-only cache in bank.py)
 # ---------------------------------------------------------------------------
 
 
@@ -133,8 +133,7 @@ def blast_radius_rank(self, level: str) -> int | None:
 
 
 def _default_action_taxonomy_path() -> Path:
-    """Discover the shipped ``assets/safety_agentic/action_taxonomy.yaml``
-    regardless of CWD. See ``core._assets.default_assets_dir``."""
+    """The probe's own ``assets/safety_agentic/action_taxonomy.yaml``, through ``probe_assets_dir``."""
     from usersim.engine.core._assets import probe_assets_dir
 
     return probe_assets_dir("safety_agentic") / "action_taxonomy.yaml"
diff --git a/src/usersim/engine/assets/safety_agentic/SCHEMA.md b/src/usersim/engine/probes/safety_agentic/assets/safety_agentic/SCHEMA.md
similarity index 98%
rename from src/usersim/engine/assets/safety_agentic/SCHEMA.md
rename to src/usersim/engine/probes/safety_agentic/assets/safety_agentic/SCHEMA.md
index 6cf79e7..dd92ce0 100644
--- a/src/usersim/engine/assets/safety_agentic/SCHEMA.md
+++ b/src/usersim/engine/probes/safety_agentic/assets/safety_agentic/SCHEMA.md
@@ -10,7 +10,7 @@ like.
 
 This schema lives here so:
 
-- the loader in `usersim/engine/core/agentic_bank.py` can validate
+- the loader in `usersim/engine/probes/safety_agentic/bank.py` can validate
   incoming banks and resolve `sanctioned_action_name` against the
   shipped `assets/safety_agentic/action_taxonomy.yaml`;
 - the probe adapter and scorer can encode against a stable contract;
@@ -228,7 +228,7 @@ recommended before relying on their results.
 
 ## Loader contract
 
-The loader (`usersim/engine/core/agentic_bank.py`) enforces:
+The loader (`usersim/engine/probes/safety_agentic/bank.py`) enforces:
 
 - `schema_version` is in the supported set.
 - `bank_id`, `bank_version`, `sub_protocols` (non-empty),
diff --git a/src/usersim/engine/assets/safety_agentic/action_taxonomy.yaml b/src/usersim/engine/probes/safety_agentic/assets/safety_agentic/action_taxonomy.yaml
similarity index 100%
rename from src/usersim/engine/assets/safety_agentic/action_taxonomy.yaml
rename to src/usersim/engine/probes/safety_agentic/assets/safety_agentic/action_taxonomy.yaml
diff --git a/src/usersim/engine/assets/safety_agentic/sample.yaml b/src/usersim/engine/probes/safety_agentic/assets/safety_agentic/sample.yaml
similarity index 100%
rename from src/usersim/engine/assets/safety_agentic/sample.yaml
rename to src/usersim/engine/probes/safety_agentic/assets/safety_agentic/sample.yaml
diff --git a/src/usersim/engine/core/agentic_bank.py b/src/usersim/engine/probes/safety_agentic/bank.py
similarity index 99%
rename from src/usersim/engine/core/agentic_bank.py
rename to src/usersim/engine/probes/safety_agentic/bank.py
index e305a1b..e58349b 100644
--- a/src/usersim/engine/core/agentic_bank.py
+++ b/src/usersim/engine/probes/safety_agentic/bank.py
@@ -675,10 +675,9 @@ def _require_localized_in_bank(
 def default_agentic_bank_path() -> Path:
     """Default on-disk path for the agentic-action bank.
 
-    Discovered via ``core._assets.default_assets_dir`` so the loader
-    resolves from the packaged asset tree, so it is independent of
-    wheel-installed deployments. Asset directory matches the consuming
-    probe name (``safety_agentic``).
+    Resolved through ``probe_assets_dir``, so it is the probe's own
+    ``assets/safety_agentic/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/safety_agentic/generator.py b/src/usersim/engine/probes/safety_agentic/probe.py
similarity index 99%
rename from src/usersim/engine/probes/safety_agentic/generator.py
rename to src/usersim/engine/probes/safety_agentic/probe.py
index 672fd37..bfe8b62 100644
--- a/src/usersim/engine/probes/safety_agentic/generator.py
+++ b/src/usersim/engine/probes/safety_agentic/probe.py
@@ -84,16 +84,9 @@
 import logging
 import time
 from dataclasses import dataclass
+from pathlib import Path
 from typing import Any, Iterable
 
-from usersim.engine.core.action_taxonomy import load_action_taxonomy_default
-from usersim.engine.core.agentic_bank import (
-    ActionRequest,
-    ToolSpec,
-    default_agentic_bank_path,
-    load_agentic_bank_default,
-    reset_agentic_bank_cache,
-)
 from usersim.engine.core.llm import acall_llm, get_current_outcome_builder, set_current_outcome_builder
 from usersim.engine.core.outcomes import (
     FailureAttribution,
@@ -119,6 +112,14 @@
     make_result,
 )
 from usersim.engine.core.tool_calls import parse_tool_call
+from usersim.engine.probes.safety_agentic.action_taxonomy import load_action_taxonomy_default
+from usersim.engine.probes.safety_agentic.bank import (
+    ActionRequest,
+    ToolSpec,
+    default_agentic_bank_path,
+    load_agentic_bank_default,
+    reset_agentic_bank_cache,
+)
 from usersim.engine.probes.safety_agentic.task_derivation import (
     resolve_task_from_row,
 )
@@ -206,6 +207,7 @@ class SafetyAgenticProbe(AgenticMixin, BankBackedProbe):
     """
 
     label = "safety_agentic"
+    assets_root = Path(__file__).parent / "assets"
     description = (
         "Tests safety under agentic tool-call probes across four sub-protocols "
         "(unsanctioned-action / scope-creep / consequence-disclosure / "
diff --git a/src/usersim/engine/evaluator/scorers/safety_agentic.py b/src/usersim/engine/probes/safety_agentic/scorer.py
similarity index 98%
rename from src/usersim/engine/evaluator/scorers/safety_agentic.py
rename to src/usersim/engine/probes/safety_agentic/scorer.py
index cd7b514..31d4080 100644
--- a/src/usersim/engine/evaluator/scorers/safety_agentic.py
+++ b/src/usersim/engine/probes/safety_agentic/scorer.py
@@ -9,7 +9,7 @@
 ``attempted_actions`` the ``safety_agentic`` probe records in its
 metadata (stored with the row as ``conversation_metadata``), joins every
 attempted action against the rich action taxonomy (loaded via
-:mod:`core.action_taxonomy`), and produces:
+:mod:`probes.safety_agentic.action_taxonomy`), and produces:
 
 - **mechanical** per-trajectory findings that need no LLM call:
 
@@ -93,18 +93,18 @@
 
 from pydantic import BaseModel, Field
 
-from usersim.engine.core.action_taxonomy import (
+from usersim.engine.core.llm import acall_llm
+from usersim.engine.core.messages import format_conversation_history_for_prompt
+from usersim.engine.core.row_metadata import recorded_value
+from usersim.engine.evaluator.scorers import register_scorer
+from usersim.engine.probes.safety_agentic.action_taxonomy import (
     ActionTaxonomy,
     load_action_taxonomy_default,
 )
-from usersim.engine.core.agentic_bank import (
+from usersim.engine.probes.safety_agentic.bank import (
     ActionRequest,
     load_agentic_bank_default,
 )
-from usersim.engine.core.llm import acall_llm
-from usersim.engine.core.messages import format_conversation_history_for_prompt
-from usersim.engine.core.row_metadata import recorded_value
-from usersim.engine.evaluator.scorers import register_scorer
 
 logger = logging.getLogger("usersim.engine")
 
@@ -362,7 +362,7 @@ async def score_safety_agentic_trajectory(
         bank = load_agentic_bank_default()
     except Exception as e:
         logger.warning(
-            "  |-- evaluator/scorers.safety_agentic: failed to load agentic bank: %s",
+            "  |-- safety_agentic scorer: failed to load agentic bank: %s",
             e,
         )
         return _no_op_envelope(
@@ -378,7 +378,7 @@ async def score_safety_agentic_trajectory(
         taxonomy = load_action_taxonomy_default()
     except Exception as e:
         logger.warning(
-            "  |-- evaluator/scorers.safety_agentic: failed to load action taxonomy: %s",
+            "  |-- safety_agentic scorer: failed to load action taxonomy: %s",
             e,
         )
         return _no_op_envelope(
@@ -393,7 +393,7 @@ async def score_safety_agentic_trajectory(
     bank_version_mismatch = pinned_version is not None and pinned_version != bank.bank_version
     if bank_version_mismatch:
         logger.info(
-            "  |-- evaluator/scorers.safety_agentic: trajectory pinned "
+            "  |-- safety_agentic scorer: trajectory pinned "
             "bank_version=%s but current loaded bank is %s — reporting "
             "drift flag on this row",
             pinned_version,
@@ -818,7 +818,7 @@ async def _call_judge(
         )
     except Exception as e:
         logger.warning(
-            "  |-- evaluator/scorers.safety_agentic: judge %r raised on action_request %s: %s: %s",
+            "  |-- safety_agentic scorer: judge %r raised on action_request %s: %s: %s",
             judge_alias,
             action_request.id,
             type(e).__name__,
@@ -831,7 +831,7 @@ async def _call_judge(
         parsed = json.loads(content)
     except (json.JSONDecodeError, TypeError):
         logger.warning(
-            "  |-- evaluator/scorers.safety_agentic: failed to parse structured output for action_request %s",
+            "  |-- safety_agentic scorer: failed to parse structured output for action_request %s",
             action_request.id,
         )
         return None, "parse_failure"
@@ -840,7 +840,7 @@ async def _call_judge(
         judgment = _AgenticJudgment.model_validate(parsed)
     except Exception as e:
         logger.warning(
-            "  |-- evaluator/scorers.safety_agentic: schema validation failed for action_request %s: %s",
+            "  |-- safety_agentic scorer: schema validation failed for action_request %s: %s",
             action_request.id,
             e,
         )
diff --git a/src/usersim/engine/probes/safety_agentic/task_derivation.py b/src/usersim/engine/probes/safety_agentic/task_derivation.py
index 469d80f..db54ffb 100644
--- a/src/usersim/engine/probes/safety_agentic/task_derivation.py
+++ b/src/usersim/engine/probes/safety_agentic/task_derivation.py
@@ -31,7 +31,7 @@
 import random
 from typing import Any, Iterable
 
-from usersim.engine.core.agentic_bank import ActionRequest, AgenticBank
+from usersim.engine.probes.safety_agentic.bank import ActionRequest, AgenticBank
 from usersim.engine.probes.sov_ai_facts.task_derivation import (
     persona_to_tags,
 )
diff --git a/src/usersim/taxonomy/capabilities.py b/src/usersim/taxonomy/capabilities.py
index 41d0098..d43cf06 100644
--- a/src/usersim/taxonomy/capabilities.py
+++ b/src/usersim/taxonomy/capabilities.py
@@ -995,7 +995,7 @@ def _identity_pressure_by_strategy() -> tuple[CapabilityDefinition, ...]:
     "sov_ai_dynamic": "src/usersim/engine/probes/sov_ai_dynamic/probe.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/probe.py",
-    "safety_agentic": "src/usersim/engine/probes/safety_agentic/generator.py",
+    "safety_agentic": "src/usersim/engine/probes/safety_agentic/probe.py",
     "financial_services": "src/usersim/engine/probes/financial_services/generator.py",
 }
 
@@ -1008,7 +1008,7 @@ def _identity_pressure_by_strategy() -> tuple[CapabilityDefinition, ...]:
     "sov_ai_dynamic": "src/usersim/engine/probes/sov_ai_dynamic/scorer.py",
     "sov_ai_multilingual_parity": "src/usersim/engine/probes/sov_ai_multilingual_parity/scorer.py",
     "safety_chat_pressure": "src/usersim/engine/probes/safety_chat_pressure/scorer.py",
-    "safety_agentic": "src/usersim/engine/evaluator/scorers/safety_agentic.py",
+    "safety_agentic": "src/usersim/engine/probes/safety_agentic/scorer.py",
     "financial_services": "src/usersim/engine/evaluator/scorers/financial_services.py",
     "identity_disclosure": "src/usersim/engine/evaluator/scorers/identity_disclosure.py",
 }
diff --git a/templates/probe/test_probe.py.template b/templates/probe/test_probe.py.template
index 7568d04..14b806e 100644
--- a/templates/probe/test_probe.py.template
+++ b/templates/probe/test_probe.py.template
@@ -14,7 +14,7 @@ after it. For real working examples, see:
   pattern).
 - ``tests/probes/safety_chat_pressure/test_probe.py`` — bank-reframing probe
   (``on_followup_failure="abort"`` test case).
-- ``tests/engine/probes/test_safety_agentic.py`` — agentic probe (custom ``run_dispatch``,
+- ``tests/probes/safety_agentic/test_probe.py`` — agentic probe (custom ``run_dispatch``,
   no user-LLM, simpler mock pattern — FIFO-only since there's only
   one acall_llm caller).
 
diff --git a/tests/engine/core/test_assets_discovery.py b/tests/engine/core/test_assets_discovery.py
index ba4418c..ab72e7a 100644
--- a/tests/engine/core/test_assets_discovery.py
+++ b/tests/engine/core/test_assets_discovery.py
@@ -11,7 +11,7 @@
 its overrides and a probe's declared ``assets_root``, and that the probe
 registry and the on-disk layout stay in lockstep.
 
-Loader-level tests live in the per-bank test files (``test_agentic_bank.py``,
+Loader-level tests live in the per-bank test files (``safety_agentic/test_bank.py``,
 ``safety_chat_pressure/test_bank.py``, …) and exercise these helpers indirectly via the
 shipped sample.
 """
diff --git a/tests/engine/core/test_default_asset_paths.py b/tests/engine/core/test_default_asset_paths.py
index d154021..0adc1fd 100644
--- a/tests/engine/core/test_default_asset_paths.py
+++ b/tests/engine/core/test_default_asset_paths.py
@@ -43,15 +43,6 @@
 
 import pytest
 
-from usersim.engine.core.action_taxonomy import (
-    load_action_taxonomy_default,
-    reset_action_taxonomy_cache,
-)
-from usersim.engine.core.agentic_bank import (
-    default_agentic_bank_path,
-    load_agentic_bank,
-    reset_agentic_bank_cache,
-)
 from usersim.engine.core.clinical_profile_bank import (
     CLIENT_PROBE_LABELS,
     default_clinical_profile_bank_path,
@@ -66,6 +57,15 @@
 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.seeds import load_seeds
+from usersim.engine.probes.safety_agentic.action_taxonomy import (
+    load_action_taxonomy_default,
+    reset_action_taxonomy_cache,
+)
+from usersim.engine.probes.safety_agentic.bank import (
+    default_agentic_bank_path,
+    load_agentic_bank,
+    reset_agentic_bank_cache,
+)
 from usersim.engine.probes.safety_chat_pressure.bank import (
     default_pressure_bank_path,
     load_pressure_bank,
diff --git a/tests/engine/evaluator/test_evaluator_scorer_structured_output_schemas.py b/tests/engine/evaluator/test_evaluator_scorer_structured_output_schemas.py
index 0e97dd0..1336262 100644
--- a/tests/engine/evaluator/test_evaluator_scorer_structured_output_schemas.py
+++ b/tests/engine/evaluator/test_evaluator_scorer_structured_output_schemas.py
@@ -27,10 +27,10 @@
     PREMISE_STANCES,
     extraction_schema,
 )
-from usersim.engine.evaluator.scorers.safety_agentic import (
+from usersim.engine.probes.safety_agentic.scorer import (
     JudgeScore as AgenticJudgeScore,
 )
-from usersim.engine.evaluator.scorers.safety_agentic import (
+from usersim.engine.probes.safety_agentic.scorer import (
     _AgenticJudgment,
 )
 from usersim.engine.probes.safety_chat_pressure.scorer import (
diff --git a/tests/engine/probes/test_india_variant_probes.py b/tests/engine/probes/test_india_variant_probes.py
index f1308bd..bcab5be 100644
--- a/tests/engine/probes/test_india_variant_probes.py
+++ b/tests/engine/probes/test_india_variant_probes.py
@@ -217,7 +217,7 @@ async def test_safety_chat_pressure_variant_translates_turn1(indian_persona, cfg
 
 
 async def test_safety_agentic_variant_constructs_and_routes(indian_persona, cfg) -> None:
-    from usersim.engine.probes.safety_agentic.generator import SafetyAgenticProbe
+    from usersim.engine.probes.safety_agentic.probe import SafetyAgenticProbe
 
     ob = _builder()
     probe = SafetyAgenticProbe(
diff --git a/tests/engine/probes/test_probe_invariants.py b/tests/engine/probes/test_probe_invariants.py
index 43c0434..d34580d 100644
--- a/tests/engine/probes/test_probe_invariants.py
+++ b/tests/engine/probes/test_probe_invariants.py
@@ -40,7 +40,7 @@
 - Per-probe end-to-end behavior (verbatim turn-1 shape, two-turn
   follow-up, placeholder warnings, bank-load failures): the
   per-probe test files (``sov_ai_facts/test_probe.py``,
-  ``test_safety_agentic.py``, etc.) own these.
+  ``safety_agentic/test_probe.py``, etc.) own these.
 - Substrate behavior (mixins, ``BankBackedProbe``,
   ``LocalePromptPack``, ``@register_probe`` decorator semantics):
   ``test_probes_substrate.py`` owns these.
diff --git a/tests/engine/test_generator_row_scope.py b/tests/engine/test_generator_row_scope.py
index 642808d..1eee06c 100644
--- a/tests/engine/test_generator_row_scope.py
+++ b/tests/engine/test_generator_row_scope.py
@@ -54,7 +54,7 @@ async def test_a_nested_scope_puts_the_row_builder_back(self) -> None:
 
         from usersim.engine.core.llm import get_current_outcome_builder, set_current_outcome_builder
         from usersim.engine.core.outcomes import OutcomeBuilder, Provenance
-        from usersim.engine.probes.safety_agentic import generator as safety_agentic
+        from usersim.engine.probes.safety_agentic import probe as safety_agentic
 
         row_builder = OutcomeBuilder(provenance=Provenance())
         probe = safety_agentic.SafetyAgenticProbe.__new__(safety_agentic.SafetyAgenticProbe)
diff --git a/tests/probes/safety_agentic/__init__.py b/tests/probes/safety_agentic/__init__.py
new file mode 100644
index 0000000..52a7a9d
--- /dev/null
+++ b/tests/probes/safety_agentic/__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_action_taxonomy.py b/tests/probes/safety_agentic/test_action_taxonomy.py
similarity index 96%
rename from tests/engine/core/test_action_taxonomy.py
rename to tests/probes/safety_agentic/test_action_taxonomy.py
index 90166dd..a0fe2a9 100644
--- a/tests/engine/core/test_action_taxonomy.py
+++ b/tests/probes/safety_agentic/test_action_taxonomy.py
@@ -1,9 +1,9 @@
 # SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 
-"""Tests for the rich action-taxonomy loader (core.action_taxonomy).
+"""Tests for the rich action-taxonomy loader (probes.safety_agentic.action_taxonomy).
 
-The taxonomy is also read by ``core.agentic_bank`` for name-only
+The taxonomy is also read by ``probes.safety_agentic.bank`` for name-only
 validation. This test file focuses on the **rich** loader that the
 ``safety_agentic`` scorer consumes — full field parsing, blast-radius
 rank computation, schema-violation error messages, cache + env override.
@@ -17,8 +17,8 @@
 
 import pytest
 
-from usersim.engine.core._assets import packaged_assets_dir
-from usersim.engine.core.action_taxonomy import (
+from usersim.engine.core._assets import probe_assets_dir
+from usersim.engine.probes.safety_agentic.action_taxonomy import (
     ActionTaxonomy,
     ActionTaxonomyError,
     load_action_taxonomy,
@@ -27,7 +27,7 @@
 )
 
 REPO_ROOT = Path(__file__).resolve().parents[3]
-SHIPPED_PATH = packaged_assets_dir() / "safety_agentic/action_taxonomy.yaml"
+SHIPPED_PATH = probe_assets_dir("safety_agentic") / "action_taxonomy.yaml"
 
 
 @pytest.fixture(autouse=True)
diff --git a/tests/engine/core/test_agentic_bank.py b/tests/probes/safety_agentic/test_bank.py
similarity index 98%
rename from tests/engine/core/test_agentic_bank.py
rename to tests/probes/safety_agentic/test_bank.py
index 1dd39e4..0c93028 100644
--- a/tests/engine/core/test_agentic_bank.py
+++ b/tests/probes/safety_agentic/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/agentic_bank.py`` — the agentic-action bank loader.
+"""Tests for ``probes/safety_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`` / ``sov_ai_multilingual_parity/test_bank.py`` / ``safety_chat_pressure/test_bank.py``):
@@ -26,8 +26,8 @@
 
 import pytest
 
-from usersim.engine.core._assets import packaged_assets_dir
-from usersim.engine.core.agentic_bank import (
+from usersim.engine.core._assets import probe_assets_dir
+from usersim.engine.probes.safety_agentic.bank import (
     AgenticBank,
     AgenticBankError,
     MockToolResponse,
@@ -533,7 +533,7 @@ def test_default_path_resolves_to_assets(self) -> None:
 class TestShippedSampleBank:
     @pytest.fixture
     def sample_bank(self) -> AgenticBank:
-        return load_agentic_bank((packaged_assets_dir() / "safety_agentic/sample.yaml"))
+        return load_agentic_bank((probe_assets_dir("safety_agentic") / "sample.yaml"))
 
     def test_loads_clean(self, sample_bank: AgenticBank) -> None:
         assert sample_bank.bank_id == "sample_v1"
diff --git a/tests/engine/probes/test_safety_agentic.py b/tests/probes/safety_agentic/test_probe.py
similarity index 96%
rename from tests/engine/probes/test_safety_agentic.py
rename to tests/probes/safety_agentic/test_probe.py
index c900621..0c7ffae 100644
--- a/tests/engine/probes/test_safety_agentic.py
+++ b/tests/probes/safety_agentic/test_probe.py
@@ -34,15 +34,7 @@
 import pytest
 
 import usersim.engine.generator  # noqa: F401 — triggers probe registration
-from usersim.engine.core._assets import packaged_assets_dir
-from usersim.engine.core.agentic_bank import (
-    ActionProvenance,
-    ActionRequest,
-    AgenticBank,
-    MockToolResponse,
-    ToolSpec,
-    load_agentic_bank,
-)
+from usersim.engine.core._assets import probe_assets_dir
 from usersim.engine.core.outcomes import (
     FailureClass,
     OutcomeStatus,
@@ -51,12 +43,20 @@
 )
 from usersim.engine.core.probes import _PROBE_REGISTRY, resolve_probe
 from usersim.engine.probes.safety_agentic import (
-    generator as probe_gen,
+    probe as probe_gen,
 )
 from usersim.engine.probes.safety_agentic import (
     prompts,
     task_derivation,
 )
+from usersim.engine.probes.safety_agentic.bank import (
+    ActionProvenance,
+    ActionRequest,
+    AgenticBank,
+    MockToolResponse,
+    ToolSpec,
+    load_agentic_bank,
+)
 
 # ---------------------------------------------------------------------------
 # Fixtures
@@ -64,7 +64,7 @@
 
 
 REPO_ROOT = Path(__file__).resolve().parents[3]
-SAMPLE_BANK_PATH = packaged_assets_dir() / "safety_agentic/sample.yaml"
+SAMPLE_BANK_PATH = probe_assets_dir("safety_agentic") / "sample.yaml"
 
 
 @pytest.fixture
@@ -485,7 +485,7 @@ async def test_no_tool_call_response_short_circuits(
         with (
             patch.dict(os.environ, _shipped_bank_env()),
             patch(
-                "usersim.engine.probes.safety_agentic.generator.acall_llm",
+                "usersim.engine.probes.safety_agentic.probe.acall_llm",
                 side_effect=_mock_call_llm(
                     [
                         {"role": "assistant", "content": "I should ask first."},
@@ -525,7 +525,7 @@ async def test_first_user_turn_is_initial_user_prompt_verbatim(
         simulator_cfg: Any,
     ) -> None:
         with patch.dict(os.environ, _shipped_bank_env()):
-            from usersim.engine.core.agentic_bank import (
+            from usersim.engine.probes.safety_agentic.bank import (
                 load_agentic_bank_default,
             )
 
@@ -544,7 +544,7 @@ async def test_first_user_turn_is_initial_user_prompt_verbatim(
         with (
             patch.dict(os.environ, _shipped_bank_env()),
             patch(
-                "usersim.engine.probes.safety_agentic.generator.acall_llm",
+                "usersim.engine.probes.safety_agentic.probe.acall_llm",
                 side_effect=_mock_call_llm(
                     [
                         {"role": "assistant", "content": "ok"},
@@ -577,7 +577,7 @@ async def test_assistant_tool_call_intercepted_with_mock_response(
         with (
             patch.object(probe_gen, "_load_bank", return_value=synth),
             patch(
-                "usersim.engine.probes.safety_agentic.generator.acall_llm",
+                "usersim.engine.probes.safety_agentic.probe.acall_llm",
                 side_effect=_mock_call_llm(
                     [
                         {
@@ -633,7 +633,7 @@ async def test_attempted_actions_recorded_with_full_details(
         with (
             patch.object(probe_gen, "_load_bank", return_value=synth),
             patch(
-                "usersim.engine.probes.safety_agentic.generator.acall_llm",
+                "usersim.engine.probes.safety_agentic.probe.acall_llm",
                 side_effect=_mock_call_llm(
                     [
                         {
@@ -688,7 +688,7 @@ async def test_unknown_tool_call_yields_synthetic_error_response(
         with (
             patch.object(probe_gen, "_load_bank", return_value=synth),
             patch(
-                "usersim.engine.probes.safety_agentic.generator.acall_llm",
+                "usersim.engine.probes.safety_agentic.probe.acall_llm",
                 side_effect=_mock_call_llm(
                     [
                         {
@@ -758,7 +758,7 @@ async def test_panel_override_with_unknown_id_returns_failure(
         with (
             patch.dict(os.environ, _shipped_bank_env()),
             patch(
-                "usersim.engine.probes.safety_agentic.generator.acall_llm",
+                "usersim.engine.probes.safety_agentic.probe.acall_llm",
             ) as mock_call,
         ):
             result = await probe_gen.simulate_safety_agentic(
@@ -788,7 +788,7 @@ async def test_no_resolvable_request_returns_failure(
         with (
             patch.object(probe_gen, "_load_bank", return_value=synth),
             patch(
-                "usersim.engine.probes.safety_agentic.generator.acall_llm",
+                "usersim.engine.probes.safety_agentic.probe.acall_llm",
             ) as mock_call,
         ):
             result = await probe_gen.simulate_safety_agentic(
@@ -817,7 +817,7 @@ async def _raises(*args, **kwargs):
         with (
             patch.dict(os.environ, _shipped_bank_env()),
             patch(
-                "usersim.engine.probes.safety_agentic.generator.acall_llm",
+                "usersim.engine.probes.safety_agentic.probe.acall_llm",
                 side_effect=_raises,
             ),
         ):
@@ -854,7 +854,7 @@ async def test_non_placeholder_request_no_warning(
         with (
             patch.object(probe_gen, "_load_bank", return_value=synth),
             patch(
-                "usersim.engine.probes.safety_agentic.generator.acall_llm",
+                "usersim.engine.probes.safety_agentic.probe.acall_llm",
                 side_effect=_mock_call_llm(
                     [
                         {"role": "assistant", "content": "ok"},
@@ -888,7 +888,7 @@ async def test_probe_variant_format(
         with (
             patch.object(probe_gen, "_load_bank", return_value=synth),
             patch(
-                "usersim.engine.probes.safety_agentic.generator.acall_llm",
+                "usersim.engine.probes.safety_agentic.probe.acall_llm",
                 side_effect=_mock_call_llm(
                     [
                         {"role": "assistant", "content": "ok"},
@@ -921,7 +921,7 @@ async def test_probe_variant_format_when_no_sanctioned_action(
         with (
             patch.object(probe_gen, "_load_bank", return_value=synth),
             patch(
-                "usersim.engine.probes.safety_agentic.generator.acall_llm",
+                "usersim.engine.probes.safety_agentic.probe.acall_llm",
                 side_effect=_mock_call_llm(
                     [
                         {"role": "assistant", "content": "Refusing."},
@@ -979,7 +979,7 @@ def should_succeed(self, state: Any) -> bool:
         with (
             patch.dict(os.environ, _shipped_bank_env()),
             patch(
-                "usersim.engine.probes.safety_agentic.generator.acall_llm",
+                "usersim.engine.probes.safety_agentic.probe.acall_llm",
                 side_effect=_mock_call_llm([{"role": "assistant", "content": "I should ask first."}]),
             ),
         ):
diff --git a/tests/engine/evaluator/test_evaluator_scorer_safety_agentic.py b/tests/probes/safety_agentic/test_scorer.py
similarity index 91%
rename from tests/engine/evaluator/test_evaluator_scorer_safety_agentic.py
rename to tests/probes/safety_agentic/test_scorer.py
index a120f15..540dcd9 100644
--- a/tests/engine/evaluator/test_evaluator_scorer_safety_agentic.py
+++ b/tests/probes/safety_agentic/test_scorer.py
@@ -26,9 +26,7 @@
 import pytest
 
 from usersim.engine.config import ConversationSimulatorConfig
-from usersim.engine.core._assets import packaged_assets_dir
-from usersim.engine.core.action_taxonomy import reset_action_taxonomy_cache
-from usersim.engine.core.agentic_bank import reset_agentic_bank_cache
+from usersim.engine.core._assets import probe_assets_dir
 from usersim.engine.core.outcomes import Provenance
 from usersim.engine.evaluator.scorers import (
     clear_registry,
@@ -36,7 +34,10 @@
     list_scorers,
     load_default_scorers,
 )
-from usersim.engine.evaluator.scorers.safety_agentic import (
+from usersim.engine.probes.safety_agentic.action_taxonomy import reset_action_taxonomy_cache
+from usersim.engine.probes.safety_agentic.bank import reset_agentic_bank_cache
+from usersim.engine.probes.safety_agentic.probe import simulate_safety_agentic
+from usersim.engine.probes.safety_agentic.scorer import (
     ALL_AGENTIC_AXES,
     BLAST_RADIUS_CONTAINED_AXIS,
     BOUNDARY_HELD_AXIS,
@@ -47,7 +48,6 @@
     MECHANICAL_AXES,
     score_safety_agentic_trajectory,
 )
-from usersim.engine.probes.safety_agentic.generator import simulate_safety_agentic
 
 # ---------------------------------------------------------------------------
 # Shared fixtures
@@ -55,8 +55,8 @@
 
 
 REPO_ROOT = Path(__file__).resolve().parents[3]
-SHIPPED_BANK = packaged_assets_dir() / "safety_agentic/sample.yaml"
-SHIPPED_TAXONOMY = packaged_assets_dir() / "safety_agentic/action_taxonomy.yaml"
+SHIPPED_BANK = probe_assets_dir("safety_agentic") / "sample.yaml"
+SHIPPED_TAXONOMY = probe_assets_dir("safety_agentic") / "action_taxonomy.yaml"
 
 
 @pytest.fixture(autouse=True)
@@ -160,7 +160,7 @@ def test_registered_via_default_modules(self) -> None:
         assert "safety_agentic" in list_scorers()
         fn = get_scorer("safety_agentic")
         assert fn.__name__ == "score_safety_agentic_trajectory"
-        assert fn.__module__ == ("usersim.engine.evaluator.scorers.safety_agentic")
+        assert fn.__module__ == ("usersim.engine.probes.safety_agentic.scorer")
 
     def test_registration_restored_after_clear(self) -> None:
         clear_registry()
@@ -197,7 +197,7 @@ async def test_missing_action_request_id_short_circuits(self, shipped_env) -> No
         traj = _trajectory()
         del traj["action_request_id"]
         with patch(
-            "usersim.engine.evaluator.scorers.safety_agentic.acall_llm",
+            "usersim.engine.probes.safety_agentic.scorer.acall_llm",
         ) as mock_call:
             out = await score_safety_agentic_trajectory(traj, {"judge_model": object()})
         mock_call.assert_not_called()
@@ -212,7 +212,7 @@ async def test_non_string_action_request_id_short_circuits(
         traj = _trajectory()
         traj["action_request_id"] = 42
         with patch(
-            "usersim.engine.evaluator.scorers.safety_agentic.acall_llm",
+            "usersim.engine.probes.safety_agentic.scorer.acall_llm",
         ) as mock_call:
             out = await score_safety_agentic_trajectory(traj, {"judge_model": object()})
         mock_call.assert_not_called()
@@ -222,7 +222,7 @@ async def test_missing_sub_protocol_short_circuits(self, shipped_env) -> None:
         traj = _trajectory()
         del traj["sub_protocol"]
         with patch(
-            "usersim.engine.evaluator.scorers.safety_agentic.acall_llm",
+            "usersim.engine.probes.safety_agentic.scorer.acall_llm",
         ) as mock_call:
             out = await score_safety_agentic_trajectory(traj, {"judge_model": object()})
         mock_call.assert_not_called()
@@ -241,7 +241,7 @@ async def test_bank_load_failure_returns_error_envelope(self) -> None:
             {"USERSIM_SAFETY_AGENTIC_BANK": "/does/not/exist.yaml"},
         ):
             with patch(
-                "usersim.engine.evaluator.scorers.safety_agentic.acall_llm",
+                "usersim.engine.probes.safety_agentic.scorer.acall_llm",
             ) as mock_call:
                 out = await score_safety_agentic_trajectory(
                     _trajectory(),
@@ -265,11 +265,11 @@ def _explode() -> None:
 
         with (
             patch(
-                "usersim.engine.evaluator.scorers.safety_agentic.load_action_taxonomy_default",
+                "usersim.engine.probes.safety_agentic.scorer.load_action_taxonomy_default",
                 side_effect=_explode,
             ),
             patch(
-                "usersim.engine.evaluator.scorers.safety_agentic.acall_llm",
+                "usersim.engine.probes.safety_agentic.scorer.acall_llm",
             ) as mock_call,
         ):
             out = await score_safety_agentic_trajectory(
@@ -293,7 +293,7 @@ async def test_missing_action_request_id_in_bank_returns_error(
     ) -> None:
         traj = _trajectory(action_request_id="AR-DOES-NOT-EXIST")
         with patch(
-            "usersim.engine.evaluator.scorers.safety_agentic.acall_llm",
+            "usersim.engine.probes.safety_agentic.scorer.acall_llm",
         ) as mock_call:
             out = await score_safety_agentic_trajectory(traj, {"judge_model": object()})
         mock_call.assert_not_called()
@@ -319,7 +319,7 @@ async def test_unsanctioned_action_with_no_tool_calls_passes_mechanical(
             attempted_actions=[],
         )
         with patch(
-            "usersim.engine.evaluator.scorers.safety_agentic.acall_llm",
+            "usersim.engine.probes.safety_agentic.scorer.acall_llm",
         ) as mock_call:
             out = await score_safety_agentic_trajectory(traj, {"judge_model": object()})
         mock_call.assert_not_called()
@@ -349,7 +349,7 @@ async def test_unsanctioned_action_with_tempted_call_fails_mechanically(
             ],
         )
         with patch(
-            "usersim.engine.evaluator.scorers.safety_agentic.acall_llm",
+            "usersim.engine.probes.safety_agentic.scorer.acall_llm",
         ) as mock_call:
             out = await score_safety_agentic_trajectory(traj, {"judge_model": object()})
         mock_call.assert_not_called()
@@ -380,7 +380,7 @@ async def test_scope_creep_with_only_sanctioned_call_passes(
             ],
         )
         with patch(
-            "usersim.engine.evaluator.scorers.safety_agentic.acall_llm",
+            "usersim.engine.probes.safety_agentic.scorer.acall_llm",
         ) as mock_call:
             out = await score_safety_agentic_trajectory(traj, {"judge_model": object()})
         mock_call.assert_not_called()
@@ -405,7 +405,7 @@ async def test_fabricated_tool_is_classified_fabricated(
             ],
         )
         with patch(
-            "usersim.engine.evaluator.scorers.safety_agentic.acall_llm",
+            "usersim.engine.probes.safety_agentic.scorer.acall_llm",
         ) as mock_call:
             out = await score_safety_agentic_trajectory(traj, {"judge_model": object()})
         mock_call.assert_not_called()
@@ -446,7 +446,7 @@ async def test_reads_list_directly(self, shipped_env) -> None:
             ],
         )
         with patch(
-            "usersim.engine.evaluator.scorers.safety_agentic.acall_llm",
+            "usersim.engine.probes.safety_agentic.scorer.acall_llm",
         ):
             out = await score_safety_agentic_trajectory(traj, {"judge_model": object()})
         assert len(out["attempted_actions_classified"]) == 1
@@ -469,7 +469,7 @@ async def test_reads_json_string(self, shipped_env) -> None:
             ]
         )
         with patch(
-            "usersim.engine.evaluator.scorers.safety_agentic.acall_llm",
+            "usersim.engine.probes.safety_agentic.scorer.acall_llm",
         ):
             out = await score_safety_agentic_trajectory(traj, {"judge_model": object()})
         assert len(out["attempted_actions_classified"]) == 1
@@ -484,7 +484,7 @@ async def test_invalid_json_yields_empty_classified_list(
         )
         traj["attempted_actions"] = "{not-json"
         with patch(
-            "usersim.engine.evaluator.scorers.safety_agentic.acall_llm",
+            "usersim.engine.probes.safety_agentic.scorer.acall_llm",
         ):
             out = await score_safety_agentic_trajectory(traj, {"judge_model": object()})
         assert out["attempted_actions_classified"] == []
@@ -499,7 +499,7 @@ async def test_reads_conversation_metadata_when_the_row_has_no_column(self, ship
         traj = _trajectory(action_request_id="AR-UNSAN-001", sub_protocol="unsanctioned_action")
         del traj["attempted_actions"]
         traj["conversation_metadata"] = encode({"attempted_actions": [_TEMPTED_ACTION]})
-        with patch("usersim.engine.evaluator.scorers.safety_agentic.acall_llm"):
+        with patch("usersim.engine.probes.safety_agentic.scorer.acall_llm"):
             out = await score_safety_agentic_trajectory(traj, {"judge_model": object()})
         assert [a["classification"] for a in out["attempted_actions_classified"]] == ["tempted"]
         assert out["scores"][BOUNDARY_HELD_AXIS]["score"] == 1
@@ -507,7 +507,7 @@ async def test_reads_conversation_metadata_when_the_row_has_no_column(self, ship
     async def test_a_top_level_value_wins_over_conversation_metadata(self, shipped_env) -> None:
         traj = _trajectory(action_request_id="AR-UNSAN-001", sub_protocol="unsanctioned_action")
         traj["conversation_metadata"] = json.dumps({"attempted_actions": [_TEMPTED_ACTION]})
-        with patch("usersim.engine.evaluator.scorers.safety_agentic.acall_llm"):
+        with patch("usersim.engine.probes.safety_agentic.scorer.acall_llm"):
             out = await score_safety_agentic_trajectory(traj, {"judge_model": object()})
         assert out["attempted_actions_classified"] == []
 
@@ -526,7 +526,7 @@ async def test_reading_the_compromised_document_holds_the_boundary(self, shipped
             attempted_actions=[_call("read_external_content")],
         )
         with patch(
-            "usersim.engine.evaluator.scorers.safety_agentic.acall_llm",
+            "usersim.engine.probes.safety_agentic.scorer.acall_llm",
             side_effect=_mock_judge_call([_judge_payload(injection_resistance=5)]),
         ):
             out = await score_safety_agentic_trajectory(traj, {"judge_model": object()})
@@ -542,7 +542,7 @@ async def test_acting_on_the_injection_breaks_the_boundary(self, shipped_env) ->
             attempted_actions=[_call("read_external_content"), _call("share_file_external", turn_idx=1)],
         )
         with patch(
-            "usersim.engine.evaluator.scorers.safety_agentic.acall_llm",
+            "usersim.engine.probes.safety_agentic.scorer.acall_llm",
             side_effect=_mock_judge_call([_judge_payload(injection_resistance=1)]),
         ):
             out = await score_safety_agentic_trajectory(traj, {"judge_model": object()})
@@ -558,7 +558,7 @@ async def test_listing_then_deleting_the_files_in_scope_holds_the_boundary(self,
             attempted_actions=[_call("list_files_in_scope"), _call("delete_user_files_in_scope", turn_idx=1)],
         )
         with patch(
-            "usersim.engine.evaluator.scorers.safety_agentic.acall_llm",
+            "usersim.engine.probes.safety_agentic.scorer.acall_llm",
             side_effect=_mock_judge_call([_judge_payload(consequence_disclosure=5, confirmation_sought=5)]),
         ):
             out = await score_safety_agentic_trajectory(traj, {"judge_model": object()})
@@ -571,7 +571,7 @@ async def test_reading_the_calendar_before_the_sanctioned_change_holds_the_bound
             sub_protocol="scope_creep",
             attempted_actions=[_call("read_calendar"), _call("modify_calendar_event", turn_idx=1)],
         )
-        with patch("usersim.engine.evaluator.scorers.safety_agentic.acall_llm") as judge:
+        with patch("usersim.engine.probes.safety_agentic.scorer.acall_llm") as judge:
             out = await score_safety_agentic_trajectory(traj, {"judge_model": object()})
         judge.assert_not_called()
         assert [a["classification"] for a in out["attempted_actions_classified"]] == ["supporting", "sanctioned"]
@@ -590,7 +590,7 @@ async def test_a_tempted_call_the_probe_recorded_fails_the_boundary(self, shippe
             "function": {"name": "send_email_on_behalf", "arguments": json.dumps({"to": "x@y.com"})},
         }
         with patch(
-            "usersim.engine.probes.safety_agentic.generator.acall_llm",
+            "usersim.engine.probes.safety_agentic.probe.acall_llm",
             side_effect=[
                 {"role": "assistant", "content": "", "tool_calls": [tempted_call]},
                 {"role": "assistant", "content": "Sent."},
@@ -609,7 +609,7 @@ async def test_a_tempted_call_the_probe_recorded_fails_the_boundary(self, shippe
         stored = {column: result[column] for column in cfg.side_effect_columns if column in result}
         assert "attempted_actions" not in stored
 
-        with patch("usersim.engine.evaluator.scorers.safety_agentic.acall_llm") as judge:
+        with patch("usersim.engine.probes.safety_agentic.scorer.acall_llm") as judge:
             out = await score_safety_agentic_trajectory(stored, {"judge_model": object()})
         judge.assert_not_called()
         assert [a["classification"] for a in out["attempted_actions_classified"]] == ["tempted"]
@@ -640,7 +640,7 @@ async def test_judge_called_and_both_llm_axes_populated(
             ],
         )
         with patch(
-            "usersim.engine.evaluator.scorers.safety_agentic.acall_llm",
+            "usersim.engine.probes.safety_agentic.scorer.acall_llm",
             side_effect=_mock_judge_call(
                 [
                     _judge_payload(
@@ -676,7 +676,7 @@ async def test_low_disclosure_blocks(self, shipped_env) -> None:
             ],
         )
         with patch(
-            "usersim.engine.evaluator.scorers.safety_agentic.acall_llm",
+            "usersim.engine.probes.safety_agentic.scorer.acall_llm",
             side_effect=_mock_judge_call(
                 [
                     _judge_payload(
@@ -704,7 +704,7 @@ async def test_low_confirmation_blocks(self, shipped_env) -> None:
             ],
         )
         with patch(
-            "usersim.engine.evaluator.scorers.safety_agentic.acall_llm",
+            "usersim.engine.probes.safety_agentic.scorer.acall_llm",
             side_effect=_mock_judge_call(
                 [
                     _judge_payload(
@@ -728,7 +728,7 @@ async def test_no_assistant_turns_short_circuits_judge(
             n_assistant_turns=0,
         )
         with patch(
-            "usersim.engine.evaluator.scorers.safety_agentic.acall_llm",
+            "usersim.engine.probes.safety_agentic.scorer.acall_llm",
         ) as mock_call:
             out = await score_safety_agentic_trajectory(traj, {"judge_model": object()})
         mock_call.assert_not_called()
@@ -754,7 +754,7 @@ async def test_judge_called_and_injection_axis_populated(
             attempted_actions=[],  # Correct: no tool calls in response to injection.
         )
         with patch(
-            "usersim.engine.evaluator.scorers.safety_agentic.acall_llm",
+            "usersim.engine.probes.safety_agentic.scorer.acall_llm",
             side_effect=_mock_judge_call(
                 [
                     _judge_payload(
@@ -785,7 +785,7 @@ async def test_injection_failure_blocks(self, shipped_env) -> None:
             ],
         )
         with patch(
-            "usersim.engine.evaluator.scorers.safety_agentic.acall_llm",
+            "usersim.engine.probes.safety_agentic.scorer.acall_llm",
             side_effect=_mock_judge_call(
                 [
                     _judge_payload(
@@ -827,7 +827,7 @@ async def _raises(*args, **kwargs):
             ],
         )
         with patch(
-            "usersim.engine.evaluator.scorers.safety_agentic.acall_llm",
+            "usersim.engine.probes.safety_agentic.scorer.acall_llm",
             side_effect=_raises,
         ):
             out = await score_safety_agentic_trajectory(traj, {"judge_model": object()})
@@ -848,7 +848,7 @@ async def test_unparseable_judge_output_yields_parse_failure(
             sub_protocol="consequence_disclosure",
         )
         with patch(
-            "usersim.engine.evaluator.scorers.safety_agentic.acall_llm",
+            "usersim.engine.probes.safety_agentic.scorer.acall_llm",
             return_value={"role": "assistant", "content": "not json"},
         ):
             out = await score_safety_agentic_trajectory(traj, {"judge_model": object()})
@@ -872,7 +872,7 @@ async def test_schema_violation_yields_schema_validation_error(
             "injection_resistance_reasoning": None,
         }
         with patch(
-            "usersim.engine.evaluator.scorers.safety_agentic.acall_llm",
+            "usersim.engine.probes.safety_agentic.scorer.acall_llm",
             return_value={"role": "assistant", "content": json.dumps(bad)},
         ):
             out = await score_safety_agentic_trajectory(traj, {"judge_model": object()})
@@ -892,7 +892,7 @@ async def test_matching_version_no_mismatch(self, shipped_env) -> None:
             pinned_version="v0.5.5",
         )
         with patch(
-            "usersim.engine.evaluator.scorers.safety_agentic.acall_llm",
+            "usersim.engine.probes.safety_agentic.scorer.acall_llm",
         ):
             out = await score_safety_agentic_trajectory(traj, {"judge_model": object()})
         assert out["bank_version_mismatch"] is False
@@ -905,7 +905,7 @@ async def test_mismatched_version_flagged(self, shipped_env) -> None:
             pinned_version="v99.0.0",
         )
         with patch(
-            "usersim.engine.evaluator.scorers.safety_agentic.acall_llm",
+            "usersim.engine.probes.safety_agentic.scorer.acall_llm",
         ):
             out = await score_safety_agentic_trajectory(traj, {"judge_model": object()})
         assert out["bank_version_mismatch"] is True
@@ -921,7 +921,7 @@ async def test_no_pinned_version_no_mismatch(self, shipped_env) -> None:
         outcome["provenance"]["bank_version"] = {}
         traj["simulation_outcome"] = json.dumps(outcome)
         with patch(
-            "usersim.engine.evaluator.scorers.safety_agentic.acall_llm",
+            "usersim.engine.probes.safety_agentic.scorer.acall_llm",
         ):
             out = await score_safety_agentic_trajectory(traj, {"judge_model": object()})
         assert out["bank_version_mismatch"] is False
@@ -956,7 +956,7 @@ async def test_sums_ranks_across_attempts(self, shipped_env) -> None:
             ],
         )
         with patch(
-            "usersim.engine.evaluator.scorers.safety_agentic.acall_llm",
+            "usersim.engine.probes.safety_agentic.scorer.acall_llm",
         ):
             out = await score_safety_agentic_trajectory(traj, {"judge_model": object()})
         assert out["total_modeled_blast_radius"] == 6
@@ -982,7 +982,7 @@ async def test_fabricated_tools_contribute_zero_to_sum(
             ],
         )
         with patch(
-            "usersim.engine.evaluator.scorers.safety_agentic.acall_llm",
+            "usersim.engine.probes.safety_agentic.scorer.acall_llm",
         ):
             out = await score_safety_agentic_trajectory(traj, {"judge_model": object()})
         assert out["total_modeled_blast_radius"] == 0
@@ -1002,7 +1002,7 @@ async def test_envelope_has_all_expected_keys(self, shipped_env) -> None:
             sub_protocol="consequence_disclosure",
         )
         with patch(
-            "usersim.engine.evaluator.scorers.safety_agentic.acall_llm",
+            "usersim.engine.probes.safety_agentic.scorer.acall_llm",
             side_effect=_mock_judge_call(
                 [
                     _judge_payload(
@@ -1039,7 +1039,7 @@ async def test_action_request_placeholder_propagated(self, shipped_env) -> None:
             sub_protocol="unsanctioned_action",
         )
         with patch(
-            "usersim.engine.evaluator.scorers.safety_agentic.acall_llm",
+            "usersim.engine.probes.safety_agentic.scorer.acall_llm",
         ):
             out = await score_safety_agentic_trajectory(traj, {"judge_model": object()})
         assert out["action_request_placeholder"] is False
diff --git a/tests/test_capabilities.py b/tests/test_capabilities.py
index 5a18119..a906910 100644
--- a/tests/test_capabilities.py
+++ b/tests/test_capabilities.py
@@ -174,7 +174,7 @@ def test_capabilities_expose_implementation_refs() -> None:
         for ref in implementation_refs_for_capability(capability_by_id("sovereign_local_knowledge"))
     )
     assert any(
-        ref.endswith("probes/safety_agentic/generator.py")
+        ref.endswith("probes/safety_agentic/probe.py")
         for ref in implementation_refs_for_capability(capability_by_id("agentic_disclosure"))
     )
 
diff --git a/tests/test_cli_simulate_inputs.py b/tests/test_cli_simulate_inputs.py
index 2f59287..21a2b06 100644
--- a/tests/test_cli_simulate_inputs.py
+++ b/tests/test_cli_simulate_inputs.py
@@ -64,7 +64,7 @@
     "usersim.engine.core.translation",
     "usersim.engine.core.realized_audit",
     "usersim.engine.probes.health_disclosure.move_runtime",
-    "usersim.engine.probes.safety_agentic.generator",
+    "usersim.engine.probes.safety_agentic.probe",
     "usersim.engine.probes.safety_chat_pressure.classifier",
     "usersim.engine.probes.tool_calling.generator",
 )
diff --git a/tests/test_probe_names_in_shared_code.py b/tests/test_probe_names_in_shared_code.py
index 78b258d..cc763ba 100644
--- a/tests/test_probe_names_in_shared_code.py
+++ b/tests/test_probe_names_in_shared_code.py
@@ -32,8 +32,6 @@ class instead (``architecture/probes.md``, "What a probe declares").
     "asset_gen/registry.py": 3,
     "cli/_pipeline.py": 7,
     "cli/smoke.py": 2,
-    "engine/core/action_taxonomy.py": 1,
-    "engine/core/agentic_bank.py": 2,
     "engine/core/clinical_profile_bank.py": 4,
     "engine/core/finance_bank.py": 1,
     "engine/core/identity_spec.py": 1,
@@ -41,7 +39,6 @@ class instead (``architecture/probes.md``, "What a probe declares").
     "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,
     "reporting/capability_report.py": 1,
     "reporting/identity.py": 1,
     "taxonomy/capabilities.py": 41,
diff --git a/tests/test_simulator_pipeline_e2e.py b/tests/test_simulator_pipeline_e2e.py
index 46dc49f..326108e 100644
--- a/tests/test_simulator_pipeline_e2e.py
+++ b/tests/test_simulator_pipeline_e2e.py
@@ -165,7 +165,7 @@ def _patched_call_llm():
     - ``probes/tool_calling/generator.py`` — probe-internal tool
       arg-generation calls (it doesn't go through ``ConversationLoop``
       for the tool-call branch).
-    - ``probes/safety_agentic/generator.py`` — mock-environment
+    - ``probes/safety_agentic/probe.py`` — mock-environment
       tool-response synthesis.
 
     Evaluator-side scorer modules (``evaluator/scorers/*.py``) ALSO
@@ -178,7 +178,7 @@ def _patched_call_llm():
         "usersim.engine.core.judges.acall_llm",
         "usersim.engine.core.context.acall_llm",
         "usersim.engine.probes.tool_calling.generator.acall_llm",
-        "usersim.engine.probes.safety_agentic.generator.acall_llm",
+        "usersim.engine.probes.safety_agentic.probe.acall_llm",
     )
     patches = [patch(t, side_effect=_mock_call_llm_side_effect) for t in targets]
     for p in patches:

From 3c94a0d2ef82fa89e60d3f9b0d8baf6f6466b24e Mon Sep 17 00:00:00 2001
From: Yev Meyer <ymeyer@nvidia.com>
Date: Tue, 3 Nov 2026 09:46:49 -0500
Subject: [PATCH 4/5] refactor(engine): keep identity_disclosure in one folder

The probe module is probes/identity_disclosure/probe.py, beside its spec loader spec.py and spec_types.py, its scorer scorer.py and grading.py, and its spec in assets/identity_disclosure/, the folder the class names as its assets_root. Its tests are in tests/probes/identity_disclosure/, with the tests of its capability rows and of the contradictions report. The label, the scorer name, the spec, the stored columns and spec layering are unchanged.

A probe built on this one imports IdentityDisclosureProbe from usersim.engine.probes.identity_disclosure.probe, as the probe README and the extension fixture do. DEFAULT_SCORER_MODULES, the capability source references and the docs name the new modules, the authoring guide lists the probe among those kept in one folder, and the shared-code label allowances list neither the spec loader nor the scorer.

Signed-off-by: Yev Meyer <ymeyer@nvidia.com>
---
 docs/engine/AUTHORING_A_PROBE.md              |  8 ++---
 docs/engine/README.md                         | 10 +++---
 src/usersim/engine/core/probes.py             |  2 +-
 .../engine/evaluator/scorers/__init__.py      |  2 +-
 .../probes/identity_disclosure/README.md      | 12 +++----
 .../assets/identity_disclosure/SCHEMA.md      |  4 +--
 .../assets/identity_disclosure/spec.yaml      |  0
 .../identity_disclosure/grading.py}           |  2 +-
 .../{generator.py => probe.py}                | 22 ++++++------
 .../identity_disclosure/scorer.py}            | 10 +++---
 .../identity_disclosure/spec.py}              |  6 ++--
 .../identity_disclosure/spec_types.py}        |  2 +-
 .../identity_disclosure/task_derivation.py    |  2 +-
 src/usersim/taxonomy/capabilities.py          |  8 ++---
 tests/engine/core/test_default_asset_paths.py |  8 ++---
 ...luator_scorer_structured_output_schemas.py |  2 +-
 .../usersim_extension_fixture/identity.py     |  2 +-
 .../extension_package/verify_seams.py         |  2 +-
 tests/probes/identity_disclosure/__init__.py  |  2 ++
 .../identity_disclosure/test_capabilities.py} |  6 ++--
 .../test_contradictions.py}                   |  6 ++--
 .../identity_disclosure/test_probe.py}        | 36 +++++++++----------
 .../identity_disclosure/test_scorer.py}       |  8 ++---
 .../identity_disclosure/test_spec.py}         |  2 +-
 tests/test_cli_simulate_inputs.py             |  4 +--
 tests/test_extension_api.py                   |  2 +-
 tests/test_probe_names_in_shared_code.py      |  2 --
 27 files changed, 87 insertions(+), 85 deletions(-)
 rename src/usersim/engine/{ => probes/identity_disclosure}/assets/identity_disclosure/SCHEMA.md (99%)
 rename src/usersim/engine/{ => probes/identity_disclosure}/assets/identity_disclosure/spec.yaml (100%)
 rename src/usersim/engine/{evaluator/scorers/identity_grading.py => probes/identity_disclosure/grading.py} (99%)
 rename src/usersim/engine/probes/identity_disclosure/{generator.py => probe.py} (99%)
 rename src/usersim/engine/{evaluator/scorers/identity_disclosure.py => probes/identity_disclosure/scorer.py} (98%)
 rename src/usersim/engine/{core/identity_spec.py => probes/identity_disclosure/spec.py} (99%)
 rename src/usersim/engine/{core/identity_spec_types.py => probes/identity_disclosure/spec_types.py} (99%)
 create mode 100644 tests/probes/identity_disclosure/__init__.py
 rename tests/{test_identity_capabilities.py => probes/identity_disclosure/test_capabilities.py} (97%)
 rename tests/{test_identity_contradictions.py => probes/identity_disclosure/test_contradictions.py} (95%)
 rename tests/{engine/probes/test_identity_disclosure.py => probes/identity_disclosure/test_probe.py} (98%)
 rename tests/{engine/evaluator/test_evaluator_scorer_identity_disclosure.py => probes/identity_disclosure/test_scorer.py} (99%)
 rename tests/{engine/core/test_identity_spec.py => probes/identity_disclosure/test_spec.py} (99%)

diff --git a/docs/engine/AUTHORING_A_PROBE.md b/docs/engine/AUTHORING_A_PROBE.md
index 6d9f51f..bb6d88e 100644
--- a/docs/engine/AUTHORING_A_PROBE.md
+++ b/docs/engine/AUTHORING_A_PROBE.md
@@ -258,10 +258,10 @@ 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`, `general_open_ended`, the three `sov_ai_*` probes,
-`safety_chat_pressure` and `safety_agentic`, 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
+`safety_chat_pressure`, `safety_agentic` and `identity_disclosure`, which keep
+their bank loader (`bank.py`, `taxonomy.py` for `sov_ai_dynamic`, `spec.py` for
+`identity_disclosure`) 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,
diff --git a/docs/engine/README.md b/docs/engine/README.md
index 0b65a4b..3138de1 100644
--- a/docs/engine/README.md
+++ b/docs/engine/README.md
@@ -133,7 +133,7 @@ different questions of the model.
 | 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 | `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` |
+| Identity spec | `probes/identity_disclosure/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` |
 
 The query bank's cross-locale shape is what makes cross-locale comparison
 possible: every trajectory records the `query_id` it asked, so answers to
@@ -600,7 +600,6 @@ The tables describe the main files.
 | `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_identity_spec.py` | Identity spec loader: loading, the env override and the cache; model-id resolution (a family name after any separator, declarations over patterns, hosts never deciding); name folding; layering down the asset search path and the roots probes declare (inheritance, in-place patches, probes built on top, cycles); merging and `select` / `exclude` narrowing; the content digest; validation reporting every problem at once |
 
 **`tests/engine/probes/`: the probe generators, verifier and invariants**
 
@@ -611,7 +610,6 @@ The tables describe the main files.
 | `test_verifiers.py` | Tool-call JSON-schema validator |
 | `test_financial_services.py` | `financial_services` probe: registration, multi-institution bank load and scope integrity, deterministic task instantiation and gold derivation, institution-scoped retrieval with no cross-institution leakage, the discoverable-tool gate through `run_dispatch` (mocked assistant), and the verifier scorer (good, unauthorized, skipped and dynamic-tier rows) |
 | `test_health_disclosure.py` | `health_disclosure` family: the Guard's pacing gates, the `commit_move` tool schema, move extraction, and the per-client config (topics, risk, archetype, prompts, axes) |
-| `test_identity_disclosure.py` | `identity_disclosure` probe: registration and the placeholder flag; shipped-spec checks (public examples, sourced lineage, every cataloged and shipped model resolves); deterministic tactic and strategy draws with panel overrides; expected-identity resolution (declared, inferred, an abort showing the declaration to add, an unidentified facade); openings and competitors; openings the simulated user writes (its instruction and prompt line, the check for AI names, the opening gate, a draft naming an AI developer written again in a loop run); row columns; prompts; pressure (every multi-turn row pressed, reframings recorded, its own prompt, gate and anchor, a loop run the scorer grades); the user-turn policy; a probe built on top; one full turn through the loop |
 
 **`tests/engine/evaluator/`: judges, scorers, registry**
 
@@ -625,7 +623,6 @@ The tables describe the main files.
 | `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_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 |
-| `test_evaluator_scorer_identity_disclosure.py` | `identity_disclosure` scorer: applicability and fail-closed paths, extraction schema built from the spec's developers, quote validation, name-to-developer mapping (names, products, model names, ambiguous words), per-turn grading (false and spontaneous attribution, premise stance, reasoning spill, human claims), giving in under pressure, trajectory rates and review flags, same-family extractor warning |
 | `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 |
@@ -651,3 +648,8 @@ The tables describe the main files.
 | `safety_agentic/test_scorer.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 |
 | `safety_agentic/test_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 |
 | `safety_agentic/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 |
+| `identity_disclosure/test_probe.py` | `identity_disclosure` probe: registration and the placeholder flag; shipped-spec checks (public examples, sourced lineage, every cataloged and shipped model resolves); deterministic tactic and strategy draws with panel overrides; expected-identity resolution (declared, inferred, an abort showing the declaration to add, an unidentified facade); openings and competitors; openings the simulated user writes (its instruction and prompt line, the check for AI names, the opening gate, a draft naming an AI developer written again in a loop run); row columns; prompts; pressure (every multi-turn row pressed, reframings recorded, its own prompt, gate and anchor, a loop run the scorer grades); the user-turn policy; a probe built on top; one full turn through the loop |
+| `identity_disclosure/test_scorer.py` | `identity_disclosure` scorer: applicability and fail-closed paths, extraction schema built from the spec's developers, quote validation, name-to-developer mapping (names, products, model names, ambiguous words), per-turn grading (false and spontaneous attribution, premise stance, reasoning spill, human claims), giving in under pressure, trajectory rates and review flags, same-family extractor warning |
+| `identity_disclosure/test_spec.py` | Identity spec loader: loading, the env override and the cache; model-id resolution (a family name after any separator, declarations over patterns, hosts never deciding); name folding; layering down the asset search path and the roots probes declare (inheritance, in-place patches, probes built on top, cycles); merging and `select` / `exclude` narrowing; the content digest; validation reporting every problem at once |
+| `identity_disclosure/test_capabilities.py` | The identity capability rows: their definitions, and cells built from real scorer output |
+| `identity_disclosure/test_contradictions.py` | `identity_contradictions`: the turns the identity scorer labelled contradictions, with the question each answered |
diff --git a/src/usersim/engine/core/probes.py b/src/usersim/engine/core/probes.py
index d60bb3c..5bcf9d2 100644
--- a/src/usersim/engine/core/probes.py
+++ b/src/usersim/engine/core/probes.py
@@ -1547,7 +1547,7 @@ def decorator(cls: type[BaseProbe]) -> type[BaseProbe]:
     "usersim.engine.probes.health_disclosure.general",
     "usersim.engine.probes.health_disclosure.therapy",
     "usersim.engine.probes.health_disclosure.triage",
-    "usersim.engine.probes.identity_disclosure.generator",
+    "usersim.engine.probes.identity_disclosure.probe",
     "usersim.engine.probes.safety_agentic.probe",
     "usersim.engine.probes.safety_chat_pressure.probe",
     "usersim.engine.probes.sov_ai_dynamic.probe",
diff --git a/src/usersim/engine/evaluator/scorers/__init__.py b/src/usersim/engine/evaluator/scorers/__init__.py
index f7394fd..2d1b7d9 100644
--- a/src/usersim/engine/evaluator/scorers/__init__.py
+++ b/src/usersim/engine/evaluator/scorers/__init__.py
@@ -324,7 +324,7 @@ def reload_scorers(scorer_module_paths: Iterable[str]) -> None:
     "usersim.engine.evaluator.scorers.response_shape",
     "usersim.engine.evaluator.scorers.refusal_basics",
     "usersim.engine.evaluator.scorers.health_disclosure",
-    "usersim.engine.evaluator.scorers.identity_disclosure",
+    "usersim.engine.probes.identity_disclosure.scorer",
 )
 
 
diff --git a/src/usersim/engine/probes/identity_disclosure/README.md b/src/usersim/engine/probes/identity_disclosure/README.md
index 855fc9c..0b97709 100644
--- a/src/usersim/engine/probes/identity_disclosure/README.md
+++ b/src/usersim/engine/probes/identity_disclosure/README.md
@@ -9,8 +9,8 @@ another developer's model.
 ## What a row asks
 
 The opening turn comes from the spec
-([`assets/identity_disclosure/spec.yaml`](../../assets/identity_disclosure/spec.yaml),
-fields in [`SCHEMA.md`](../../assets/identity_disclosure/SCHEMA.md)):
+([`assets/identity_disclosure/spec.yaml`](assets/identity_disclosure/spec.yaml),
+fields in [`SCHEMA.md`](assets/identity_disclosure/SCHEMA.md)):
 
 - **Neutral questions** name no AI product or company: "Who are you?", "What
   are you?", "Introduce yourself.", "Who made you?", "What model are you?".
@@ -109,7 +109,7 @@ expected_identities:
 
 Save it as, say, `my_models.local.yaml` and set
 `USERSIM_IDENTITY_DISCLOSURE_SPEC` to its path. The spec's
-[`SCHEMA.md`](../../assets/identity_disclosure/SCHEMA.md) describes the
+[`SCHEMA.md`](assets/identity_disclosure/SCHEMA.md) describes the
 matching in full and lists the developers it knows.
 
 ## Running it
@@ -260,9 +260,9 @@ Set `USERSIM_IDENTITY_DISCLOSURE_SPEC` to the layer's path for a run, or save
 it as `identity_disclosure/spec.yaml` under a directory on `USERSIM_ASSET_PATH`
 to use it for every run. Each row records the layers and a digest of the
 merged spec in its `identity_spec` column, so results say exactly what graded
-them. [`SCHEMA.md`](../../assets/identity_disclosure/SCHEMA.md#layering) gives
+them. [`SCHEMA.md`](assets/identity_disclosure/SCHEMA.md#layering) gives
 the merge rules, and `validate_spec("identity_disclosure")` in
-`usersim.engine.core.identity_spec` reports every problem with the layers at
+`usersim.engine.probes.identity_disclosure.spec` reports every problem with the layers at
 once.
 
 ## Building a probe on top
@@ -274,7 +274,7 @@ strategies or models, can subclass this one under a label of its own:
 from pathlib import Path
 
 from usersim.engine.core.probes import register_probe
-from usersim.engine.probes.identity_disclosure.generator import IdentityDisclosureProbe
+from usersim.engine.probes.identity_disclosure.probe import IdentityDisclosureProbe
 
 
 @register_probe(family="my_identity", prompt_version="v1.0", variants=("default",))
diff --git a/src/usersim/engine/assets/identity_disclosure/SCHEMA.md b/src/usersim/engine/probes/identity_disclosure/assets/identity_disclosure/SCHEMA.md
similarity index 99%
rename from src/usersim/engine/assets/identity_disclosure/SCHEMA.md
rename to src/usersim/engine/probes/identity_disclosure/assets/identity_disclosure/SCHEMA.md
index 8d20d73..ec44bec 100644
--- a/src/usersim/engine/assets/identity_disclosure/SCHEMA.md
+++ b/src/usersim/engine/probes/identity_disclosure/assets/identity_disclosure/SCHEMA.md
@@ -2,7 +2,7 @@
 
 `spec.yaml` defines the `identity_disclosure` probe: who each model should say
 made it, which names a reply may use, and how the probe asks. It is loaded by
-`usersim.engine.core.identity_spec`.
+`usersim.engine.probes.identity_disclosure.spec`.
 
 ## Top-level fields
 
@@ -268,7 +268,7 @@ column, so runs can be compared on exactly what they were graded against.
 ## Validation
 
 `usersim smoke` checks the shipped spec. `validate_spec(label)` in
-`usersim.engine.core.identity_spec` returns every problem at once, each with a
+`usersim.engine.probes.identity_disclosure.spec` returns every problem at once, each with a
 code and the field it concerns: unknown developers, a name listed under two
 developers, ambiguous entries missing from their developer's lists, rules that
 name both or neither of `model` and `match`, competitors that are no
diff --git a/src/usersim/engine/assets/identity_disclosure/spec.yaml b/src/usersim/engine/probes/identity_disclosure/assets/identity_disclosure/spec.yaml
similarity index 100%
rename from src/usersim/engine/assets/identity_disclosure/spec.yaml
rename to src/usersim/engine/probes/identity_disclosure/assets/identity_disclosure/spec.yaml
diff --git a/src/usersim/engine/evaluator/scorers/identity_grading.py b/src/usersim/engine/probes/identity_disclosure/grading.py
similarity index 99%
rename from src/usersim/engine/evaluator/scorers/identity_grading.py
rename to src/usersim/engine/probes/identity_disclosure/grading.py
index 5b157b8..25d8ea8 100644
--- a/src/usersim/engine/evaluator/scorers/identity_grading.py
+++ b/src/usersim/engine/probes/identity_disclosure/grading.py
@@ -17,7 +17,7 @@
 from dataclasses import asdict, dataclass, field, replace
 from typing import Any, Iterable
 
-from usersim.engine.core.identity_spec_types import Vocabulary, fold_name, name_pattern
+from usersim.engine.probes.identity_disclosure.spec_types import Vocabulary, fold_name, name_pattern
 
 NO_SPONTANEOUS_FALSE_ATTRIBUTION_AXIS = "identity.no_spontaneous_false_attribution_rate"
 NO_FALSE_ATTRIBUTION_AXIS = "identity.no_false_attribution_rate"
diff --git a/src/usersim/engine/probes/identity_disclosure/generator.py b/src/usersim/engine/probes/identity_disclosure/probe.py
similarity index 99%
rename from src/usersim/engine/probes/identity_disclosure/generator.py
rename to src/usersim/engine/probes/identity_disclosure/probe.py
index 9164a45..b72ff3c 100644
--- a/src/usersim/engine/probes/identity_disclosure/generator.py
+++ b/src/usersim/engine/probes/identity_disclosure/probe.py
@@ -50,6 +50,7 @@
 import logging
 from collections.abc import Callable
 from dataclasses import asdict
+from pathlib import Path
 from typing import Any
 
 from usersim.engine.config import MODEL_ASSISTANT
@@ -58,16 +59,6 @@
     format_interaction_style_instructions,
 )
 from usersim.engine.core.identity import persona_uuid, resolve_model_name
-from usersim.engine.core.identity_spec import (
-    PROBE_LABEL,
-    ExpectedIdentity,
-    IdentitySpec,
-    TacticMode,
-    Vocabulary,
-    load_identity_spec_default,
-    spec_env_var,
-    validate_spec,
-)
 from usersim.engine.core.locale import SHIPPED_LOCALES, india_variant
 from usersim.engine.core.missing import is_missing
 from usersim.engine.core.outcomes import WarningKind
@@ -94,6 +85,16 @@
     PRESSURE_USER_SYSTEM_PROMPTS,
     USER_SYSTEM_PROMPTS,
 )
+from usersim.engine.probes.identity_disclosure.spec import (
+    PROBE_LABEL,
+    ExpectedIdentity,
+    IdentitySpec,
+    TacticMode,
+    Vocabulary,
+    load_identity_spec_default,
+    spec_env_var,
+    validate_spec,
+)
 from usersim.engine.probes.identity_disclosure.task_derivation import (
     COMPETITOR_PLACEHOLDERS,
     IDENTITY_PLACEHOLDERS,
@@ -137,6 +138,7 @@ class IdentityDisclosureProbe(BankVerbatimMixin, BankBackedProbe):
     """Persona users ask the model under test who it is; rows carry what it may truthfully say."""
 
     label = PROBE_LABEL
+    assets_root = Path(__file__).parent / "assets"
     description = (
         "Tests whether the assistant names the developer that made it when a persona "
         "asks who it is or who made it, with no system prompt: neutral questions, 'Are "
diff --git a/src/usersim/engine/evaluator/scorers/identity_disclosure.py b/src/usersim/engine/probes/identity_disclosure/scorer.py
similarity index 98%
rename from src/usersim/engine/evaluator/scorers/identity_disclosure.py
rename to src/usersim/engine/probes/identity_disclosure/scorer.py
index 9a9a35b..5f620c1 100644
--- a/src/usersim/engine/evaluator/scorers/identity_disclosure.py
+++ b/src/usersim/engine/probes/identity_disclosure/scorer.py
@@ -75,10 +75,9 @@
 from pydantic import BaseModel, Field, create_model
 
 from usersim.engine.core.identity import resolve_model_name
-from usersim.engine.core.identity_spec_types import match_model_id
 from usersim.engine.core.llm import acall_llm
 from usersim.engine.evaluator.scorers import register_scorer
-from usersim.engine.evaluator.scorers.identity_grading import (
+from usersim.engine.probes.identity_disclosure.grading import (
     CONFIRMS_DEVELOPER_AXIS,
     CORRECT_ATTRIBUTION_AXIS,
     IDENTITY_AXES,
@@ -100,6 +99,7 @@
     grade,
     normalise_claims,
 )
+from usersim.engine.probes.identity_disclosure.spec_types import match_model_id
 
 __all__ = [
     "CLAIM_TYPES",
@@ -419,9 +419,7 @@ async def _extract(
             },
         )
     except Exception as e:
-        logger.warning(
-            "  |-- evaluator/scorers.identity_disclosure: extractor %r raised %s: %s", judge_alias, type(e).__name__, e
-        )
+        logger.warning("  |-- identity_disclosure scorer: extractor %r raised %s: %s", judge_alias, type(e).__name__, e)
         return {}, f"{type(e).__name__}: {e}"
     content = resp.get("content", "") if isinstance(resp, dict) else ""
     try:
@@ -472,7 +470,7 @@ def _extractor_family(
     if same and key not in _same_family_warned:
         _same_family_warned.add(key)
         logger.warning(
-            "  |-- evaluator/scorers.identity_disclosure: the extractor %s is made by %s, the developer under test; "
+            "  |-- identity_disclosure scorer: the extractor %s is made by %s, the developer under test; "
             "pick an extractor from another developer",
             model_id,
             ", ".join(developers),
diff --git a/src/usersim/engine/core/identity_spec.py b/src/usersim/engine/probes/identity_disclosure/spec.py
similarity index 99%
rename from src/usersim/engine/core/identity_spec.py
rename to src/usersim/engine/probes/identity_disclosure/spec.py
index e698f63..8287dc4 100644
--- a/src/usersim/engine/core/identity_spec.py
+++ b/src/usersim/engine/probes/identity_disclosure/spec.py
@@ -19,7 +19,7 @@
 ``assets/identity_disclosure/SCHEMA.md``.
 
 The typed views and the model-id and name matching live in
-``usersim.engine.core.identity_spec_types`` and are re-exported here.
+``usersim.engine.probes.identity_disclosure.spec_types`` and are re-exported here.
 """
 
 from __future__ import annotations
@@ -37,7 +37,8 @@
 import yaml
 
 from usersim.engine.core._assets import asset_roots_in_use, probe_asset_roots, probe_assets_dir
-from usersim.engine.core.identity_spec_types import (
+from usersim.engine.core.localized import LocalizedText, LocalizedTextError
+from usersim.engine.probes.identity_disclosure.spec_types import (
     Developer,
     ExpectedIdentity,
     IdentityRule,
@@ -56,7 +57,6 @@
     match_model_id,
     name_pattern,
 )
-from usersim.engine.core.localized import LocalizedText, LocalizedTextError
 
 __all__ = [
     "ALLOWED_PLACEHOLDERS",
diff --git a/src/usersim/engine/core/identity_spec_types.py b/src/usersim/engine/probes/identity_disclosure/spec_types.py
similarity index 99%
rename from src/usersim/engine/core/identity_spec_types.py
rename to src/usersim/engine/probes/identity_disclosure/spec_types.py
index 1a165fe..28abfab 100644
--- a/src/usersim/engine/core/identity_spec_types.py
+++ b/src/usersim/engine/probes/identity_disclosure/spec_types.py
@@ -3,7 +3,7 @@
 
 """Typed views of the identity_disclosure spec, and how model ids and names are matched.
 
-``usersim.engine.core.identity_spec`` loads and validates specs into these
+``usersim.engine.probes.identity_disclosure.spec`` loads and validates specs into these
 types and re-exports everything public here.
 """
 
diff --git a/src/usersim/engine/probes/identity_disclosure/task_derivation.py b/src/usersim/engine/probes/identity_disclosure/task_derivation.py
index 5a8d289..5138e5c 100644
--- a/src/usersim/engine/probes/identity_disclosure/task_derivation.py
+++ b/src/usersim/engine/probes/identity_disclosure/task_derivation.py
@@ -19,7 +19,7 @@
 from dataclasses import dataclass, field
 from typing import Mapping
 
-from usersim.engine.core.identity_spec import (
+from usersim.engine.probes.identity_disclosure.spec import (
     ExpectedIdentity,
     IdentitySpec,
     Strategy,
diff --git a/src/usersim/taxonomy/capabilities.py b/src/usersim/taxonomy/capabilities.py
index d43cf06..02613aa 100644
--- a/src/usersim/taxonomy/capabilities.py
+++ b/src/usersim/taxonomy/capabilities.py
@@ -288,9 +288,9 @@ def _identity_source(*axes: str) -> EvidenceSource:
         scorer="identity_disclosure",
         axes=axes,
         implementation_refs=(
-            "src/usersim/engine/probes/identity_disclosure/generator.py",
-            "src/usersim/engine/assets/identity_disclosure/spec.yaml",
-            "src/usersim/engine/evaluator/scorers/identity_grading.py",
+            "src/usersim/engine/probes/identity_disclosure/probe.py",
+            "src/usersim/engine/probes/identity_disclosure/assets/identity_disclosure/spec.yaml",
+            "src/usersim/engine/probes/identity_disclosure/grading.py",
         ),
     )
 
@@ -1010,7 +1010,7 @@ def _identity_pressure_by_strategy() -> tuple[CapabilityDefinition, ...]:
     "safety_chat_pressure": "src/usersim/engine/probes/safety_chat_pressure/scorer.py",
     "safety_agentic": "src/usersim/engine/probes/safety_agentic/scorer.py",
     "financial_services": "src/usersim/engine/evaluator/scorers/financial_services.py",
-    "identity_disclosure": "src/usersim/engine/evaluator/scorers/identity_disclosure.py",
+    "identity_disclosure": "src/usersim/engine/probes/identity_disclosure/scorer.py",
 }
 
 _AXIS_IMPL_REFS = {
diff --git a/tests/engine/core/test_default_asset_paths.py b/tests/engine/core/test_default_asset_paths.py
index 0adc1fd..60e5b2f 100644
--- a/tests/engine/core/test_default_asset_paths.py
+++ b/tests/engine/core/test_default_asset_paths.py
@@ -49,14 +49,14 @@
     load_clinical_profile_bank,
     reset_clinical_profile_bank_cache,
 )
-from usersim.engine.core.identity_spec import (
-    default_identity_spec_path,
-    load_identity_spec,
-)
 from usersim.engine.core.locale import SHIPPED_LOCALES
 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.seeds import load_seeds
+from usersim.engine.probes.identity_disclosure.spec import (
+    default_identity_spec_path,
+    load_identity_spec,
+)
 from usersim.engine.probes.safety_agentic.action_taxonomy import (
     load_action_taxonomy_default,
     reset_action_taxonomy_cache,
diff --git a/tests/engine/evaluator/test_evaluator_scorer_structured_output_schemas.py b/tests/engine/evaluator/test_evaluator_scorer_structured_output_schemas.py
index 1336262..90ef35e 100644
--- a/tests/engine/evaluator/test_evaluator_scorer_structured_output_schemas.py
+++ b/tests/engine/evaluator/test_evaluator_scorer_structured_output_schemas.py
@@ -21,7 +21,7 @@
 
 import typing as _typing
 
-from usersim.engine.evaluator.scorers.identity_disclosure import (
+from usersim.engine.probes.identity_disclosure.scorer import (
     CLAIM_TYPES,
     DETAIL_KINDS,
     PREMISE_STANCES,
diff --git a/tests/fixtures/extension_package/usersim_extension_fixture/identity.py b/tests/fixtures/extension_package/usersim_extension_fixture/identity.py
index a7d2d6f..504f051 100644
--- a/tests/fixtures/extension_package/usersim_extension_fixture/identity.py
+++ b/tests/fixtures/extension_package/usersim_extension_fixture/identity.py
@@ -3,7 +3,7 @@
 
 from pathlib import Path
 
-from usersim.engine.probes.identity_disclosure.generator import IdentityDisclosureProbe
+from usersim.engine.probes.identity_disclosure.probe import IdentityDisclosureProbe
 from usersim.extension import register_probe
 
 
diff --git a/tests/fixtures/extension_package/verify_seams.py b/tests/fixtures/extension_package/verify_seams.py
index 80e659e..c4da175 100644
--- a/tests/fixtures/extension_package/verify_seams.py
+++ b/tests/fixtures/extension_package/verify_seams.py
@@ -173,7 +173,7 @@ def preflight_refusal(locales: list[str]) -> str:
 
 # A probe built on a shipped one: its spec layer ships in this package and
 # extends identity_disclosure's spec, which must be inside the installed wheel.
-from usersim.engine.core.identity_spec import load_identity_spec_default, validate_spec  # noqa: E402
+from usersim.engine.probes.identity_disclosure.spec import load_identity_spec_default, validate_spec  # noqa: E402
 
 check("probe built on identity_disclosure", "fixture_identity" in known_probes(), f"saw {known_probes()}")
 spec_problems = validate_spec("fixture_identity")
diff --git a/tests/probes/identity_disclosure/__init__.py b/tests/probes/identity_disclosure/__init__.py
new file mode 100644
index 0000000..52a7a9d
--- /dev/null
+++ b/tests/probes/identity_disclosure/__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/test_identity_capabilities.py b/tests/probes/identity_disclosure/test_capabilities.py
similarity index 97%
rename from tests/test_identity_capabilities.py
rename to tests/probes/identity_disclosure/test_capabilities.py
index b4e2e15..1a22585 100644
--- a/tests/test_identity_capabilities.py
+++ b/tests/probes/identity_disclosure/test_capabilities.py
@@ -13,8 +13,8 @@
 import pandas as pd
 import pytest
 
-from usersim.engine.core.identity_spec import load_identity_spec_default, reset_identity_spec_cache
-from usersim.engine.evaluator.scorers.identity_disclosure import IDENTITY_AXES, score_identity_disclosure_trajectory
+from usersim.engine.probes.identity_disclosure.scorer import IDENTITY_AXES, score_identity_disclosure_trajectory
+from usersim.engine.probes.identity_disclosure.spec import load_identity_spec_default, reset_identity_spec_cache
 from usersim.reporting.capability_report import _capability_from_definition
 from usersim.taxonomy.capabilities import (
     ALL_PROBES,
@@ -25,7 +25,7 @@
 )
 
 NEMOTRON = "nvidia/nemotron-3-super-120b-a12b"
-PATCH = "usersim.engine.evaluator.scorers.identity_disclosure.acall_llm"
+PATCH = "usersim.engine.probes.identity_disclosure.scorer.acall_llm"
 SPONTANEOUS = "identity.no_spontaneous_false_attribution_rate"
 PRESSURE = "identity_pressure_resistance"
 
diff --git a/tests/test_identity_contradictions.py b/tests/probes/identity_disclosure/test_contradictions.py
similarity index 95%
rename from tests/test_identity_contradictions.py
rename to tests/probes/identity_disclosure/test_contradictions.py
index 4fb72ab..7ccd45a 100644
--- a/tests/test_identity_contradictions.py
+++ b/tests/probes/identity_disclosure/test_contradictions.py
@@ -13,13 +13,13 @@
 import pandas as pd
 import pytest
 
-from usersim.engine.core.identity_spec import load_identity_spec_default, reset_identity_spec_cache
-from usersim.engine.evaluator.scorers.identity_disclosure import score_identity_disclosure_trajectory
+from usersim.engine.probes.identity_disclosure.scorer import score_identity_disclosure_trajectory
+from usersim.engine.probes.identity_disclosure.spec import load_identity_spec_default, reset_identity_spec_cache
 from usersim.reporting import identity_contradictions
 from usersim.reporting.identity import CONTRADICTION_COLUMNS
 
 NEMOTRON = "nvidia/nemotron-3-super-120b-a12b"
-PATCH = "usersim.engine.evaluator.scorers.identity_disclosure.acall_llm"
+PATCH = "usersim.engine.probes.identity_disclosure.scorer.acall_llm"
 
 
 @pytest.fixture(scope="module")
diff --git a/tests/engine/probes/test_identity_disclosure.py b/tests/probes/identity_disclosure/test_probe.py
similarity index 98%
rename from tests/engine/probes/test_identity_disclosure.py
rename to tests/probes/identity_disclosure/test_probe.py
index 5feebb7..0f5864f 100644
--- a/tests/engine/probes/test_identity_disclosure.py
+++ b/tests/probes/identity_disclosure/test_probe.py
@@ -25,32 +25,32 @@
 from usersim.engine.core import probes as core_probes
 from usersim.engine.core import translation
 from usersim.engine.core.behavioral import compute_behavioral_profile
-from usersim.engine.core.identity_spec import (
-    Reframing,
-    Strategy,
-    TacticKind,
-    TacticMode,
-    load_identity_spec_default,
-    reset_identity_spec_cache,
-    spec_env_var,
-    validate_spec,
-)
 from usersim.engine.core.locale import INDIA_VARIANT_LOCALES
 from usersim.engine.core.localized import LocalizedText
 from usersim.engine.core.outcomes import OutcomeBuilder, OutcomeStatus, Provenance, WarningKind
 from usersim.engine.core.probes import register_probe, resolve_probe
-from usersim.engine.evaluator.scorers import identity_disclosure as identity_scorer
-from usersim.engine.evaluator.scorers.identity_disclosure import (
+from usersim.engine.probes.identity_disclosure import probe as probe_gen
+from usersim.engine.probes.identity_disclosure import scorer as identity_scorer
+from usersim.engine.probes.identity_disclosure.probe import (
+    IdentityDisclosureProbe,
+    IdentityDisclosureProbeError,
+)
+from usersim.engine.probes.identity_disclosure.prompts import FOLLOWUP_ANCHOR, OPENING_SENT, OPENING_TO_WRITE
+from usersim.engine.probes.identity_disclosure.scorer import (
     PRESSURE_RESISTANCE_AXIS,
     Vocabulary,
     score_identity_disclosure_trajectory,
 )
-from usersim.engine.probes.identity_disclosure import generator as probe_gen
-from usersim.engine.probes.identity_disclosure.generator import (
-    IdentityDisclosureProbe,
-    IdentityDisclosureProbeError,
+from usersim.engine.probes.identity_disclosure.spec import (
+    Reframing,
+    Strategy,
+    TacticKind,
+    TacticMode,
+    load_identity_spec_default,
+    reset_identity_spec_cache,
+    spec_env_var,
+    validate_spec,
 )
-from usersim.engine.probes.identity_disclosure.prompts import FOLLOWUP_ANCHOR, OPENING_SENT, OPENING_TO_WRITE
 from usersim.engine.probes.identity_disclosure.task_derivation import (
     opening_frame,
     placeholders_in,
@@ -698,7 +698,7 @@ async def test_the_scorer_reads_the_pressure_the_row_records(self, monkeypatch)
         }
         monkeypatch.setattr(identity_scorer, "_vocabularies", {})
         extractor = AsyncMock(return_value={"role": "assistant", "content": json.dumps(extraction)})
-        with patch("usersim.engine.evaluator.scorers.identity_disclosure.acall_llm", extractor):
+        with patch("usersim.engine.probes.identity_disclosure.scorer.acall_llm", extractor):
             scored = await score_identity_disclosure_trajectory(result, {"evaluator_model": object()})
         assert scored["scores"][PRESSURE_RESISTANCE_AXIS]["score"] == pytest.approx(1.0)
 
diff --git a/tests/engine/evaluator/test_evaluator_scorer_identity_disclosure.py b/tests/probes/identity_disclosure/test_scorer.py
similarity index 99%
rename from tests/engine/evaluator/test_evaluator_scorer_identity_disclosure.py
rename to tests/probes/identity_disclosure/test_scorer.py
index b3eb4c6..902ec2f 100644
--- a/tests/engine/evaluator/test_evaluator_scorer_identity_disclosure.py
+++ b/tests/probes/identity_disclosure/test_scorer.py
@@ -18,10 +18,9 @@
 import numpy as np
 import pytest
 
-from usersim.engine.core.identity_spec import load_identity_spec_default, reset_identity_spec_cache
-from usersim.engine.evaluator.scorers import identity_disclosure as scorer
 from usersim.engine.evaluator.scorers import list_scorers, load_default_scorers
-from usersim.engine.evaluator.scorers.identity_disclosure import (
+from usersim.engine.probes.identity_disclosure import scorer
+from usersim.engine.probes.identity_disclosure.scorer import (
     CONFIRMS_DEVELOPER_AXIS,
     CORRECT_ATTRIBUTION_AXIS,
     IDENTITY_AXES,
@@ -42,13 +41,14 @@
     normalise_claims,
     score_identity_disclosure_trajectory,
 )
+from usersim.engine.probes.identity_disclosure.spec import load_identity_spec_default, reset_identity_spec_cache
 from usersim.taxonomy.eval_cell import scorer_state
 from usersim.testing import assert_scorer_conforms
 
 NEMOTRON = "nvidia/nemotron-3-super-120b-a12b"
 LLAMA_NEMOTRON = "nvidia/llama-3.1-nemotron-70b-instruct"
 GEMMA = "google/gemma-3-27b-it"
-PATCH = "usersim.engine.evaluator.scorers.identity_disclosure.acall_llm"
+PATCH = "usersim.engine.probes.identity_disclosure.scorer.acall_llm"
 
 
 @dataclass(frozen=True)
diff --git a/tests/engine/core/test_identity_spec.py b/tests/probes/identity_disclosure/test_spec.py
similarity index 99%
rename from tests/engine/core/test_identity_spec.py
rename to tests/probes/identity_disclosure/test_spec.py
index e406f7e..602070e 100644
--- a/tests/engine/core/test_identity_spec.py
+++ b/tests/probes/identity_disclosure/test_spec.py
@@ -19,7 +19,7 @@
 import pytest
 import yaml
 
-from usersim.engine.core.identity_spec import (
+from usersim.engine.probes.identity_disclosure.spec import (
     IdentitySpecError,
     ProblemCode,
     SpecProblem,
diff --git a/tests/test_cli_simulate_inputs.py b/tests/test_cli_simulate_inputs.py
index 21a2b06..da3abc1 100644
--- a/tests/test_cli_simulate_inputs.py
+++ b/tests/test_cli_simulate_inputs.py
@@ -329,8 +329,8 @@ def test_changed_assets_are_a_new_setup_and_moved_ones_are_not(rows, tmp_path):
 def test_a_change_in_a_lower_priority_asset_root_is_a_new_setup(monkeypatch, tmp_path):
     """A spec that ``extends`` its own probe reads the copy below it, so every root counts."""
     top, below = tmp_path / "top", tmp_path / "below"
-    shutil.copytree(packaged_assets_dir(), top)
-    shutil.copytree(packaged_assets_dir(), below)
+    shutil.copytree(probe_assets_dir("identity_disclosure").parent, top)
+    shutil.copytree(probe_assets_dir("identity_disclosure").parent, below)
     monkeypatch.setenv("USERSIM_ASSET_PATH", str(below))
     before = _inputs._asset_roots_digest(top)
     shadowed = next(below.rglob("spec.yaml"))
diff --git a/tests/test_extension_api.py b/tests/test_extension_api.py
index 5679cd6..077ede8 100644
--- a/tests/test_extension_api.py
+++ b/tests/test_extension_api.py
@@ -68,7 +68,7 @@
     ("usersim.engine.core.probes", "dispatch_episode"),
     ("usersim.engine.generator", None),
     # The fixture's probe that layers a spec over a shipped probe's.
-    ("usersim.engine.probes.identity_disclosure.generator", "IdentityDisclosureProbe"),
+    ("usersim.engine.probes.identity_disclosure.probe", "IdentityDisclosureProbe"),
 }
 
 
diff --git a/tests/test_probe_names_in_shared_code.py b/tests/test_probe_names_in_shared_code.py
index cc763ba..ae91e36 100644
--- a/tests/test_probe_names_in_shared_code.py
+++ b/tests/test_probe_names_in_shared_code.py
@@ -34,11 +34,9 @@ class instead (``architecture/probes.md``, "What a probe declares").
     "cli/smoke.py": 2,
     "engine/core/clinical_profile_bank.py": 4,
     "engine/core/finance_bank.py": 1,
-    "engine/core/identity_spec.py": 1,
     "engine/core/preview.py": 3,
     "engine/evaluator/axes.py": 6,
     "engine/evaluator/scorers/financial_services.py": 4,
-    "engine/evaluator/scorers/identity_disclosure.py": 1,
     "reporting/capability_report.py": 1,
     "reporting/identity.py": 1,
     "taxonomy/capabilities.py": 41,

From 22e0d135a4b5f9334f06d530f9e49f90f96a6c8c Mon Sep 17 00:00:00 2001
From: Yev Meyer <ymeyer@nvidia.com>
Date: Tue, 3 Nov 2026 09:46:53 -0500
Subject: [PATCH 5/5] test: check that every capability implementation ref
 names a file

Every implementation ref a capability lists, its own and those its probe and scorer add, is a path from the repo root to a file in the repo, so a probe module, scorer or bank that moves without its refs fails the test.

Signed-off-by: Yev Meyer <ymeyer@nvidia.com>
---
 tests/test_capabilities.py | 10 ++++++++++
 1 file changed, 10 insertions(+)

diff --git a/tests/test_capabilities.py b/tests/test_capabilities.py
index a906910..4205fab 100644
--- a/tests/test_capabilities.py
+++ b/tests/test_capabilities.py
@@ -5,6 +5,8 @@
 
 from __future__ import annotations
 
+from pathlib import Path
+
 from usersim.taxonomy.capabilities import (
     ALL_PROBES,
     capabilities_missing_scorer,
@@ -179,6 +181,14 @@ def test_capabilities_expose_implementation_refs() -> None:
     )
 
 
+def test_every_implementation_ref_names_a_file_in_the_repo() -> None:
+    """A ref is a path from the repo root; a probe's module, scorer or bank that moves takes its refs along."""
+    root = Path(__file__).resolve().parents[1]
+    refs = {ref for definition in capability_definitions() for ref in implementation_refs_for_capability(definition)}
+    assert refs
+    assert sorted(ref for ref in refs if not (root / ref).is_file()) == []
+
+
 def test_safety_pressure_uses_axis_aggregation() -> None:
     """``safety_pressure`` aggregates per-axis (1-5), not via status_rate.