Files
CASAN/AINative_OKR_CASAN5/packages/casan-harness/agentops/hallucination-tracking.yaml
T
thanhnvandClaude Opus 4.8 7101af9fd4 feat(plan-01): Phase 4c — hard cutoff, remove .specify compat facade
Repoint every remaining literal `.specify/...` / `docs/input` reference to the real
package/domain location and delete all compat symlinks. The harness now runs purely via
packages/casan-harness/... with no .specify facade; .specify holds ONLY runtime state
(logs/, agentops/alerts.log, level5/central-governance).

Refs fixed (Phase 0.5 only caught `$VAR/.specify/` — these were bare/`__file__`/literal):
- secrets-scan.sh: scan-target excludes -> packages/casan-harness/... (+ apps/okr/domain/corpus)
- loop_common.py: loop-policy.yaml -> harness config (package-relative)
- evidence-pack-build.py: judge-gate test + traceability-matrix.py -> harness/sibling
- phase10-traceability: REQ -> $CASAN_DOMAIN_ROOT/input
- run-casan-pipeline.mjs: model-fallback/drift-detect/rollback-manager -> HARNESS_BASH,
  golden -> GOLDEN_PLAN (apps/okr/domain), with .specify/logs state kept
- casan-step.mjs: requirement fallback restored to docs/input for hermetic sandboxes
- descriptive config (tool-registry/harness-package/drift-policy/hallucination/risk-registry/
  loop-policy.schema + docstrings) repointed for accuracy
- policy-bundle.yaml: 8 policy paths -> packages/casan-harness/...; manifest regenerated +
  re-signed (POLICY_HASHES_VALID files=8, POLICY_SIGNATURE_VALID)

Removed 22 .specify code/config symlinks + docs/input symlink.
Full gate via packages path, NO facade: PASS=64 FAIL=0 SKIP=3.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2026-07-08 12:28:37 +09:00

77 lines
1.8 KiB
YAML

version: 1.0
description: CASAN H6 per-step hallucination signal tracking
tracking:
enabled: true
method: cross_reference_check
log_path: .specify/logs/cost/metrics.jsonl
field_name: hallucination_signals
steps_with_tracking:
- step: step-1-srs
reference_docs:
- apps/okr/domain/input/okr-requirement.md
signals:
- ungrounded_feature_claim
- invented_requirement
- undefined_acronym
keywords_indicating_hallucination:
- "I assume"
- "typically"
- "usually includes"
- "standard practice suggests"
- step: step-3-spec
reference_docs:
- docs/output/ipa-docs/srs/
- docs/output/ipa-docs/bd/
signals:
- invented_api_endpoint
- schema_mismatch
- undefined_term
- ungrounded_business_rule
keywords_indicating_hallucination:
- "assumed from context"
- "common pattern would be"
- "likely the intent is"
- step: step-6-plan
reference_docs:
- docs/output/specs/
signals:
- invented_dependency
- wrong_module_assignment
- unspecified_library
keywords_indicating_hallucination:
- "might need"
- "could potentially use"
- "I believe"
- step: step-10-implement
reference_docs:
- docs/output/specs/
signals:
- undocumented_endpoint
- wrong_field_type
- missing_required_field
- hardcoded_value_not_in_spec
keywords_indicating_hallucination:
- "// TODO: confirm"
- "// assumed"
- "// placeholder"
- step: step-11-review-code
reference_docs:
- docs/output/specs/
- backend/
- frontend/
signals:
- false_security_clearance
- invented_test_result
- unverified_compliance_claim
thresholds:
warn: 3
block: 7
note: "warn=log warning; block=force retry of step"