[3mei]

What changed and why

The four health probes, health_triage_disclosure, health_therapy_disclosure, health_decision_support_disclosure and health_general_disclosure, keep everything in the folder they share, probes/health_disclosure/: their bank loader bank.py, realized_audit.py, the concealment scorer scorer.py and their banks in assets/<label>/, the folder their base class names as its assets_root. Their tests are in tests/probes/health_disclosure/. The labels, the scorer name, the banks, the stored columns and the caches are unchanged.

Builds on #125.

• Banks in a folder several probes share (first commit). Until a built-in's class registers, the asset lookup finds its banks in the assets/ folder of its package under usersim/engine/probes/. It takes the package named after the probe, or else the package whose assets/ holds the probe's folder, so a scorer, or a prompt read while a module loads, reaches the banks of a probe that shares its folder.
• The health probes in one folder (second commit). core/clinical_profile_bank.py becomes bank.py, core/realized_audit.py keeps its name, and the scorer becomes scorer.py. The four banks move into assets/, and the tests become test_probe.py, test_scorer.py, test_bank.py, test_realized_audit.py and test_e2e.py.
  • DEFAULT_SCORER_MODULES and the docs name the new modules, and the authoring guide and architecture/probes.md describe a folder several probes share.
  • The layout test accepts a folder several probes share, and the shared-code label allowances drop the bank loader.
  • The asset-discovery test of a built-in's packaged folder uses tool_calling.

How it was verified

At each commit, on that commit's tree:

$ make check-all && make test && make smoke
All checks passed!
365 files already formatted
4657 passed in 95.60s (0:01:35)
All checks passed.

That is the first commit's tree; the second passes with 4657 tests as well.

At the head:

$ make check-license-headers
🎉 All files have correct license headers!
$ make check-doc-links
✓ every relative link, and every link into this repository's main branch, resolves
$ make check-notebooks-clean
OK 4 notebook(s) carry no outputs
$ make check-dist
✓ installed package runs from a clean environment
$ make check-extensions
all extension seams work against an installed distribution

• The first commit's test fails without its lookup change:

  E       AssertionError: assert PosixPath('.../src/usersim/engine/assets/multi_a') == (((PosixPath('.../test_a_package_holding_several0') / 'multi') / 'assets') / 'multi_a')
  

• make health-check ran against a custom provider config for health_triage_disclosure in en_IN and health_therapy_disclosure in fr_FR. Each ended with ✓ simulate, eval and report all completed, and every model's requests show failed=0.

• usersim simulate, eight turns of health_triage_disclosure in en_US with USERSIM_DISCLOSURE_MOVES=1, then usersim eval naming the concealment scorer, ran against the same config:

  judge: scored
  concealment: ok (moves_enabled=True, preview_only=True, scores=['concealment.disclosure_coverage', 'concealment.risk_surfaced'], error=None)
  

Results impact

Conversations, prompts and scores are unchanged. The setup fingerprint of a simulate --inputs run hashes the contents of every asset root, so it changes, as after any bank edit, and such a run gives its rows new trajectory ids rather than skipping rows it wrote before this change.