Constitutional AI governance: GIES v1.0, Sentinel Governance Index v6.0, MultiJurisdictionOverride TLC model (19th assurance check) + monograph architecture

governance_artifacts/RUNNABLE_ASSURANCE.md

@@ -17,7 +17,7 @@ the master reference documents assert that a control "holds," the artifacts here
 bash governance_artifacts/run_runnable_assurance.sh
 ```
 
-Runs all eleven checks below and fails fast on any error.
+Runs all nineteen checks below and fails fast on any error.
 
 ## What is proven, and against which control
 
@@ -34,6 +34,14 @@ Runs all eleven checks below and fails fast on any error.
 | 9 | PQC WORM audit log — real CRYSTALS-Dilithium (ML-DSA-65) signatures + tamper-evident hash chain + S3 Object Lock retention | Python (`dilithium-py`) + pytest | `cry-02` | DORA, EU AI Act Art. 12 logging |
 | 10 | OmegaActual contract hardening — both contracts compile (0 warnings); 7 logic tests prove original exploitable & hardened blocks SEC-01..06 | solc 0.8.26 + pytest | `con-07` settlement | EU AI Act Art. 14, DORA |
 | 11 | Governance artifact schema validation | Python validator | manifest/schema integrity | OSCAL, evidence logging (EU AI Act Art. 12) |
+| 12 | OSCAL catalog conformance — every control's `tla-spec` / `rego-policy` / `circuit` / `simulator` prop resolves to a real in-repo artifact; every regime `#href` resolves to a back-matter anchor (no dangling references); `feasibility-tier ∈ {A,B,C,D}`; `freshness-sla` is a valid ISO-8601 duration (43 cross-reference checks, falsifiable) | Python (`oscal_conformance.py`) + pytest | all `con-*`, `cry-*`, `env-*`, `rte-*` | OSCAL 1.1.2 compliance-as-code integrity (EU AI Act Annex IV, NIST AI RMF, DORA, Basel, SR 11-7) |
+| 13 | Annex IV dossier auto-assembly — builds an OSCAL-native 8-section (A–H) EU AI Act technical-documentation dossier from the conformant catalog + live assurance evidence; refuses to run on a non-conformant catalog or unknown control id; never marks a section SATISFIED without a green runnable check | Python (`generate_annex_iv_dossier.py`) + pytest | all controls → Annex IV §A–H | EU AI Act Annex IV technical documentation (auto-assembled deliverable) |
+| 14 | DORA ICT-risk register auto-assembly — builds a 5-pillar (P1–P5) DORA register from the same catalog + live evidence; reports P4/P5 as honest coverage gaps; same refusal/honesty guarantees | Python (`generate_dora_ict_register.py`) + pytest | `env-*`, `cry-02`, `con-04/07` → DORA pillars | DORA (Reg. (EU) 2022/2554) ICT-risk register (auto-assembled deliverable) |
+| 15 | NIST AI RMF profile crosswalk auto-assembly — builds a 4-function (GOVERN/MAP/MEASURE/MANAGE) crosswalk with per-function coverage analysis from the same catalog + live evidence; same refusal/honesty guarantees | Python (`generate_nist_rmf_crosswalk.py`) + pytest | all controls → NIST AI RMF functions | NIST AI RMF 1.0 coverage crosswalk (auto-assembled deliverable) |
+| 16 | Verified distribution-bundle packaging — collects all three regulator deliverables (6 artifacts) into a `dist/` bundle and emits a `MANIFEST.json` with **two** digests per artifact and per bundle: a **`bundle_sha256`** (byte-exact provenance fingerprint of this build — changes each run with the `generated_at` timestamp) and a **`content_digest`** (timestamp-normalized — **stable/reproducible** across regenerations for a given catalog + evidence state). Both recompute from the sorted per-artifact digests; refuses to package on any catalog-conformance failure; reports coverage gaps (DORA P4/P5), never inflates them | Python (`package_distribution_bundle.py`) + pytest | all deliverables → one auditable bundle | Regulator-facing distribution package (assembly-integrity + reproducibility, not a certification) |
+| 17 | Recipient-side bundle verification — a **standalone** verifier (`verify_distribution_bundle.py`, stdlib-only; deliberately does NOT import the packager — digest rules independently re-implemented) re-checks a received `dist/` bundle: per-artifact byte + timestamp-normalized digests, both bundle-level digest recomputations, digest distinctness, and **manifest-vs-content consistency** (unit counts / coverage gaps / conformance recomputed from the bundled deliverable JSONs — a manifest that inflates SATISFIED or hides a gap FAILS). The packager's `--sign` emits a detached **ML-DSA-65 (FIPS 204)** `MANIFEST.sig.json`; the suite verifies it in strict `--require-signature` mode. Tamper negatives proven by tests (artifact edit, forged manifest claims, single-byte manifest change) | Python (`verify_distribution_bundle.py`) + `dilithium-py` + pytest | all deliverables → independently verifiable received bundle | Recipient/regulator-side integrity check (proves assembly integrity + signer key continuity, not identity without an out-of-band fingerprint comparison) |
+| 18 | Evidence freshness-SLA gate — the catalogs declare per-control `freshness-sla` props (e.g. `env-01` attestation evidence ≤ PT5M old, `cry-05` zk proof ≤ P3M); until now check C3 validated only the *format*. `check_evidence_freshness.py --run` re-executes every control's mapped assurance check (the same single-source-of-truth `CONTROL_EVIDENCE` map the deliverable generators use) and records **when** each piece of evidence was produced in a digest-protected ledger (`oscal/generated/evidence_freshness_ledger.json`); `--audit` then enforces each declared SLA against those instants (stated conversion: 1M=30d, 1Y=365d; compound `P1D/P90D` enforces the first period). STALE / FAILED / NOT-RECORDED / FUTURE-DATED / tampered-ledger → gate FAILS; organisational-evidence controls (`env-02`) are disclosed `NOT-RUNNABLE`, never counted fresh. Tamper/staleness negatives proven by 7 pytest checks | Python (`check_evidence_freshness.py`) + pytest | all runnable controls' `freshness-sla` props | Evidence-recency enforcement (EU AI Act Art. 12 record-keeping, DORA ICT monitoring; the ledger digest detects casual edits, is not a signature — pair with checks 16/17 for signed provenance) |
+| 19 | Multi-jurisdiction override consistency + Sentinel Governance Index v6.0 — `MultiJurisdictionOverride.tla` models concurrent supervisory overrides from multiple jurisdictions ({EU, US, SG}) over the severity lattice NORMAL < RESTRICT < HALT and TLC-checks four safety invariants: **`MultiJurisdictionOverrideConsistency`** (effective posture ALWAYS equals the most restrictive active override — the constitutional *Lex Severior* rule; a violation is either an override bypass or unauditable over-enforcement), `NoUnilateralWeakening` (no single jurisdiction can lower the posture while another's override is active), `HaltReleaseAudited` (HALT→NORMAL requires a `unanimous_release` record in the append-only override log — silent de-escalation impossible) and `LogWellFormed`. 2,523 distinct states, 0 errors; mutation-tested (a unilateral-release bug is caught by TLC immediately). Then `validate_governance_index.py` proves the **Sentinel Governance Index v6.0** is truthful: exactly 24 artifacts SGI-01..SGI-24, every path exists, no duplicates, GIES modules valid, every tier-A artifact names invariants + a resolvable suite step, and every TLA invariant named in the index is defined in its module (8 falsifiable IDX checks) | TLA+/TLC + Python (`validate_governance_index.py`) | multi-jurisdiction supervisory override chain; SGI v6.0 (24 artifacts) | Supervisory-college override discipline (EU AI Act Art. 65–68 market surveillance, DORA oversight colleges, cross-border coordination); index integrity is the constitutional table of contents for the GIES |
 
 ### Companion reviews & plan (this iteration)