WarrantOS layer status¶
Auto-generated by warrantos status --markdown or python -c "from provenance.status import render_markdown; print(render_markdown())".
Status values: BUILT (full SHALL coverage + tests + runtime enforcement); PARTIAL (code present, some SHALLs discipline-only); STARTER (built but documented as needing extension before production use); NOT_BUILT (documented v1.0+ deferral).
| Layer | Status | Module | Notes |
|---|---|---|---|
| L1 Context Classification (11 classes) | BUILT |
provenance.context_admissibility |
11 canonical SPEC §2.2 classes via regex; SPEC-L1-S005 review-role gating via source_agent kwarg. |
| L2 Provenance Ledger (5 sub-ledgers + override) | BUILT |
provenance.ledger_write, provenance.overrides |
INV-004 storage-level enforcement added in v0.9 via SQLite BEFORE UPDATE triggers. |
| L3 Applied Insight Compiler | BUILT |
provenance.context_admissibility, provenance.ledger_write |
SPEC-L3-N001 closure: persist_context_transform() writes the ledger row per derived requirement (v0.9). |
| L4 Context Admissibility Engine (6 actor roles) | BUILT |
provenance.context_admissibility, provenance.cbom |
Per-item flags + 6-role actor identity in CBOM v0.2. |
| L5 Clean-Room Writer Pack (5 sections) | BUILT |
provenance.writer_pack |
Clean Brief / Approved Sources / Style Rules / Acceptance Tests / Banned Residue. |
| L6 Clean-Room Generation (discipline + subprocess isolation) | BUILT |
provenance.clean_room |
WarrantOS does not call any LLM itself. The caller invokes their writer model through the InvocationPlan. |
| L7-G1 G1 Prose Boundary | BUILT |
provenance.context_admissibility |
v0.9 added prompt-template profile after empirical calibration on real briefs. |
| L7-G2 G2 Source & Warrant Check | BUILT |
provenance.verify, provenance.grade, provenance.extract |
Three graders: heuristic (default, free), Anthropic LLM (paid), Local LLM (free, OpenAI-compatible). Phase 1 (2026-06-11) added the decision trigger (closing the salience/detection misalignment) plus five categories (superlative, causal, numeric-approximation, named-body attribution, empirical comparison) with +0.30 salience weights for causal/comparison/body-attribution, and per-profile unsupported-fraction HOLD thresholds so an audit run cannot return a bare PASS with all claims unsupported. |
| L7-G3 G3 Non-Self-Grounding | BUILT |
provenance.gates |
Wired into warrantos CLI via --writer-model / --verifier-model. |
| L7-G4 G4 Safety & Contamination | BUILT |
provenance.gates, eval/corpus/contamination.jsonl |
v0.9 Phase 3 extended the generic starter set with policy-domain patterns and a 24-item labelled corpus (eval/corpus/contamination.jsonl), flipping corpus_completeness from 'starter' to 'domain-extended'. The list is still not exhaustive; production deployments SHOULD continue to extend it against their own documented threat model. |
| L7-G5 G5 Evaluation & Calibration | BUILT |
provenance.gates, eval/run_eval.py, cli.warrantos_cli |
warrantos calibrate runs the grader against the labelled eval corpus and writes .warrant/calibration.json (grader, corpus size, per-class recall, coverage estimate). check_calibration() accepts either live verdict rows (Brier-with-explicit-coverage) or the stored calibration.json. Confidence coverage is typically 0 with the offline HeuristicGrader (no confidence emitted); per-class recall is the meaningful measure until an LLM grader supplies numeric confidence. |
| L8 Human Review & Decision Authority | BUILT |
provenance.overrides, provenance.footer |
Escalation routing is a documented taxonomy not an automated workflow. |
| F-policy Foundation: Policy & Role Definitions | BUILT |
provenance.roles; docs/SPEC.md; docs/STACK.md |
v0.9: machine-readable role registry (provenance.roles) enumerates the six SPEC-F-S002 actor roles and validates a CBOM actor_identity map at runtime; the normative SPEC document is committed at docs/SPEC.md. Both gaps that held this row at PARTIAL are closed. |
| F-classification Foundation: Data Classification (sensitivity tiers) | BUILT |
provenance.classification |
4-tier default registry mirrors the reference adopter's data gate. Keyword heuristics (Cabinet, ministerial, legal advice, Crown Solicitor, HR/PIP/termination, $NNNM/B budget markers, credential patterns) are a documented STARTER set; production deployments SHALL extend them with a domain taxonomy. Unmatched text defaults to Official, never silently Public. |
| F-audit Foundation: Audit Logging | BUILT |
SQLite ledger + per-run JSON artefacts |
Audit lives in SQLite (cross-run) + per-run JSON (single-run snapshot) + shadow log (observation history). |
| F-integrity Foundation: Ledger Integrity & Attestation | BUILT |
provenance.merkle, provenance.attestation, provenance.warrant_bundle |
Stdlib Merkle integrity + fail-closed .warrant verification need no key; Ed25519 signing needs the optional [attestation] extra. Envelope is project-defined (DSSE/COSE migration under consideration). |
| F-retention Foundation: Retention & Deletion (tombstones) | BUILT |
provenance.retention, schema/provenance.sql |
INV-011 implemented as APPEND-ONLY tombstones in v0.9 Phase 3: no hard delete. Expiry of a retention window appends a tombstone marking the run logically retired while preserving every ledger row (INV-004 append-only is never violated). Per-run windows are set at run creation (column) or appended later via set_window() (latest override wins). Adopters still specify the window. |
| F-compliance Foundation: Compliance & Standards (ISO / NIST / Gov) | BUILT |
docs/SPEC.md, docs/COMPLIANCE.md; SPEC-IDs grep-traceable in code/tests |
BUILT = documented-mapping ceiling: the normative SPEC (docs/SPEC.md) and a self-assessment control mapping (docs/COMPLIANCE.md) are both committed, mapping the enforced controls (gates G1-G5, override ledger, classification, retention/tombstones, Merkle/attestation) to ISO/IEC 42001 and NIST AI RMF clauses. This is NOT certified conformance: no accredited body has assessed the project, and certification is explicitly out of scope (docs/COMPLIANCE.md disclaimer + §5). There is still no automated SPEC-ID conformance check; the grep-traceable SPEC IDs in code and tests remain the technical compliance surface. |
| F-override Foundation: Human Override | BUILT |
provenance.overrides |
Most complete of the foundation items. |
| F-metrics Foundation: Metrics & Monitoring | BUILT |
provenance.metrics, tools/warrantos-shadow-observe.py, cli.warrantos_cli |
v0.9.1: the shadow log is now aggregated, not just appended. provenance.metrics.aggregate_shadow_log() reads the JSON-lines log and computes the verdict distribution, the unsupported-claim rate over the observation window, and a simple improving/worsening/stable trend (earlier vs later half of the observed rows); a missing or empty log yields an honest empty aggregate (observed=0, trend=insufficient_data) rather than an error. warrantos metrics writes .warrant/metrics.json. The G5 link is gates.calibration_with_monitoring(), which keeps the corpus calibration and the operational shadow-rate as SEPARATE surfaces (the shadow rate is a monitoring signal, not a Brier-style confidence calibration). Tested in tests/test_metrics.py. |
Summary¶
- BUILT: 20
Generated against the running install of provenance.*. Re-run after each version bump to refresh.