STP Traceability Matrix¶

Every requirement and named test script from the original System Test Plan, mapped to the code and tests that satisfy it. Status legend: ✅ done · 🟡 in progress · ⬜ planned.

Functional Requirements¶

Req	Description	Implemented in	Test (STP script)	Status
FR-001	Upload models (.h5/.pt/.pkl)	`ethiclens_api/ingestion.py` + `POST /api/models/upload` (sandboxed)	`TS-FUNC-001` → `test_models_and_governance.py`	✅
FR-002	Synthetic personas, configurable	`fairness_core.profiles.generate_profiles`	`TS-UNIT-002` → `test_generator.py`	✅
FR-003	Compute DI / SPD / EO + composite	`fairness_core.metrics`, `audit.run_audit`	`TS-UNIT-001/003/004` → `test_metrics_known_values.py`	✅
FR-004	Create audit session	`audit_session` table + `POST /api/sessions/create`	`TS-FUNC-002` → `test_sessions_e2e.py`	✅
FR-005	Ranked mitigations where DI<0.8	`fairness_core.mitigation.recommender` + `GET /api/sessions/{id}/recommendations`	`TS-INT-002`, `TS-FUNC-004` → `test_mitigation.py`, `test_sessions_e2e.py`	✅
FR-006	Explanation + estimated impact	`mitigation.recommender`	`TS-FUNC-004` → `test_mitigation.py`	✅
FR-007	Fairness Scorecard PDF	`fairness_core.reporting.scorecard` + `GET /api/sessions/{id}/report` (ReportLab)	`TS-FUNC-003` → `test_reporting.py`, `test_sessions_e2e.py`	✅
FR-008	Escalation workflow	descoped — sign-off (FR-010) is the governance gate	`TS-UAT-001`	⬜ descoped
FR-009	Re-audit after mitigation	`parent_session_id` lineage + `POST /api/sessions/{id}/mitigate`	`TS-UAT-002` → `test_sessions_e2e.py`	✅
FR-010	Governance sign-off portal	`POST /api/sessions/{id}/sign-off` (RBAC)	`TS-UAT-003` → `test_models_and_governance.py`	✅
FR-011	Lock + clearance notification	server-enforced lock ✅; notification is a log line only (`routers/governance.py`)	`TS-UAT-003` → `test_models_and_governance.py`	🟡 lock done, notifier roadmap

Non-Functional Requirements¶

Req	Description	Approach	Test	Status
NFR-PERF-001	≤90s audit for ≤500MB model	arq async + batched ONNX inference	`TS-PERF-001`	⬜ not yet benchmarked
NFR-PERF-002	≥50 concurrent, <20% degradation	arq queue + DB pool	`TS-PERF-002`	⬜ not yet benchmarked
NFR-PERF-003	PDF ≤15s for 10 attributes	async report job, precomputed metrics	`TS-PERF-003`	⬜ not yet benchmarked
NFR-DB-001	Persist results ≤5s	worker writes in completion handler	`TS-INT-003` step 5	🟡 implemented, not timed

Golden reference & superseded tests¶

Item	Where	Status
Golden `calibrated_bias_model.pkl` (DI ≈ 0.55, pinned)	`models/golden/` + `.github/workflows/golden-audit.yml`	✅
Fairlearn 1e-9 parity proof	`test_vs_fairlearn.py`	✅
Hypothesis property invariants	`test_properties.py`	✅

Consciously superseded: any STP assertion that treats the synthetic-persona audit as ground truth is deliberately replaced — see methodology.md. The pure-math unit scripts (TS-UNIT-001/003/004) pass verbatim; we do not claim 100% pass-through of the synthetic-audit path, because that path was methodologically unsound.