3mei · Nov 3, 2026 · Nov 3, 2026 · Nov 3, 2026 · Nov 3, 2026 · Nov 3, 2026
@@ -257,9 +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
+`general_educational`, `general_open_ended`, the three `sov_ai_*` probes,
+`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>/`.
 
@@ -1223,8 +1224,8 @@ 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
-abort policy), `tests/engine/probes/test_safety_agentic.py` (custom
+(verbatim), `tests/probes/safety_chat_pressure/test_probe.py` (reframings and
+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

@@ -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).
@@ -1547,9 +1547,9 @@ 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.safety_agentic.generator",
-    "usersim.engine.probes.safety_chat_pressure.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",
     "usersim.engine.probes.sov_ai_facts.probe",
     "usersim.engine.probes.sov_ai_multilingual_parity.probe",

@@ -317,14 +317,14 @@ 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.evaluator.scorers.safety_agentic",
+    "usersim.engine.probes.safety_chat_pressure.scorer",
+    "usersim.engine.probes.safety_agentic.scorer",
     "usersim.engine.evaluator.scorers.financial_services",
     "usersim.engine.evaluator.scorers.language_compliance",
     "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",
 )
 
 

@@ -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",))

@@ -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

@@ -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"

@@ -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 "

@@ -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),

@@ -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",

@@ -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.
 """
 

@@ -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,

@@ -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`.
 """
@@ -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"

@@ -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),

@@ -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
@@ -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
 

@@ -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 / "