Files
Jonathan Boniface 3b3ca64de0
Some checks failed
check agents / check (push) Failing after 14s
F02: K-02 lookup in code (ADR 0013), represents rule in the profiler; first passing run
- Critic 0.3.0: the hooks look up every disposition citation in the verified
  Discovery package (app/evidence.py). An unresolved citation fails K-02 and
  rejects the disposition; whether the records support the basis stays the
  model's judgement.
- Critic 0.3.1: K-02 judges a disposition's citations together (evidence_ref
  is the primary record), ADR 0013 addendum.
- Profiler 0.1.1: represents edges are not dependencies (ADR 0012, now covering
  ALZ-ASM-02).
- Scaffold: builds every prompt before deleting anything.
- Tests on Discovery run exec_58ba6a71d00b's package and runs 007 and 009 (15).
- Run exec_49e355141382 (RUN-ASM-AZURE-012) passed on crucible-kab-dev.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-06 15:59:01 +00:00

4.9 KiB

[0.3.1] - 2026-11-06

Changed

  • K-02 judges support on a disposition's citations taken together (ADR 0013, addendum of 2026-11-06): evidence_ref is the primary record and evidence_refs the rest. A disposition covers a whole application, so it is never failed because its primary record alone covers one asset. Run exec_659b13e5f947 failed K-02 on the one evidence_ref (ALZ-DSC-03/CI-001) although it found that the 19 citations together supported the basis.

[0.3.0] - 2026-11-06

Added

  • K-02's lookup in code (ADR 0013, app/evidence.py): every disposition's evidence_ref and evidence_refs must resolve to a record in the verified Discovery package (by agent and record key). A citation that doesn't resolve fails K-02 and rejects the disposition. Whether the cited records support the rule-trace basis stays the model's judgement.
  • The prompt says that a resolved citation is valid, and that the rate card is cited by rate_id in the TCO line items, never as Discovery evidence. Run exec_1f518b04d6e4 failed K-02 on both: it said ALZ-DSC-05/CI-002 and CI-010 don't resolve (they do), and that the rate card isn't a Discovery record.
  • Tests on Discovery run exec_58ba6a71d00b's real package and the records of runs exec_4d88027a4d26 and exec_1f518b04d6e4 (15 in all).

[0.2.1] - 2026-11-05

Fixed

  • A represents edge is not a dependency (ADR 0012): it joins a CMDB record to the inventory asset it stands for (F01 ADR 0019). Run exec_2cd7a953335b counted eight of them as dependents of APP-001: fan-in 1.0, a risk score of 1.00 and the high-risk tail, and retire ruled out for eight dependents.
  • K-07 checks the disposition agent's own signoff_queue against its dispositions below 0.80. In run exec_2cd7a953335b it compared the critic's queue after rejections, and failed when the model rejected APP-001 under K-02.

[0.2.0] - 2026-11-05

Added

  • Code checks (ADR 0011, app/arithmetic.py, app/business_logic.py): pre_process works out K-04 (risk), K-05 (confidence), K-06 (TCO), K-07 (sign-off routing) and K-11 (schema and 7R enum) from the producers' records and shows the result to the model as checked_in_code; post_process sets those checks, the sign-off queue, the rejected dispositions and the verdict. The model's checks (K-01 to K-03, K-08 to K-10, K-12) are kept as it wrote them; a different reading of a code check is kept in unverified_claims.
  • Tests on the records of run exec_4d88027a4d26, where the critic model expected a score of 0.70 for 1.00 and summed 0.90 - 0.10 - 0.20 - 0.10 to 0.60. With the code checks, that run's verdict is pass.

[0.1.6] - 2026-11-05

Fixed

  • The prompt carries the producers' rules it recalculates with (ADR 0001): ALZ-ASM-03's factor, dependent, scoring and tail rules, and ALZ-ASM-04's 7R rule set, confidence and TCO rules, copied from their prompts by the scaffold. Run exec_1779863642d6 failed K-04 because the critic had no risk rules; the K-04 and K-05 passes of earlier runs had no rules behind them either.

[0.1.5] - 2026-11-05

Fixed

  • K-08 says that listing security findings as advisory in the scoring note, as ALZ-ASM-04 must (Objective 5), passes; only a confidence adjustment or rule-trace basis citing one fails. Run exec_5fd4a145a5d9 rejected a disposition for its scoring note alone.

[0.1.4] - 2026-11-05

Changed

  • K-12 is settled when the workflow is built (ADR 0010); the critic reports it from critic_model and executor_models alone. Run exec_0b0afb70b8b9 failed K-12 again on a Discovery note about Discovery's own critic model, with every other check passing.

[0.1.3] - 2026-11-05

Fixed

  • The verdict rule is stated plainly: an unverified K-10 never fails the run on its own (ADR 0006). Run exec_effdf1042137 passed every other check and still returned fail on K-10 unverified under the earlier wording.
  • K-10 names where ALZ-ASM-05's output arrives (posture_status, security_findings). The same run reported that no ALZ-ASM-05 output was supplied.

[0.1.2] - 2026-11-05

Fixed

  • K-12 compares only critic_model with executor_models. Run exec_4be193061c41 failed K-12 on a note in the Discovery package that Discovery's network agent runs on the critic's model: a Discovery agent is not an executor of this run.
  • New input in_cloud_estate, the planner's flag, so K-10 can check that ALZ-ASM-05 reports not_applicable with no findings when the estate is not in the cloud. Run exec_4be193061c41 reported K-10 unverified partly because the flag never reached it.

[0.1.1] - 2026-11-05

Changed

  • rate_card_document comes from the run input of that name, the rate card's JSON as text, not from an uploaded file (ADR 0009). The input's description says so.

[0.1.0] - 2026-11-02

Added

  • First release: Independently validates the Assessment: coverage, evidence, dependents, recalculated scores, confidence and TCO, the interim and advisory rules, and routes confidence below 0.80 to disposition sign-off.