core/docs/assessment/10-layer-cards/MV-verification-evidence.md

MV — Verification & Evidence (cross-cutting)

Kind: layer (cross-cut) · Parent: CORE · Assessor: Opus 5 (Phase 2) Re-verified: 39331dbc (2026-07-28) — see the arc note below. Verified at: 8927c563 (2026-07-27) Liveness: partial-wiring-debt · Fitness: strained · Topology role: benchmark/eval artifact + tooling surface

How CORE knows what it knows about itself. MV is cross-cutting because replay is a property of every layer and an apparatus of this one. Its philosophical charge follows directly from the thesis: a decoding system's failures are locatable, so the machinery that locates them is not overhead — it is the difference between a structured failure that can be fixed and a stochastic one that can only be regularized. MV is also the layer this assessment most depends on, and therefore the one whose gaps most threaten the assessment's own conclusions.

Telos stages: replay deterministically (primary); learn (evidence for calibration) Macro role: Produces the evidence that every other layer's claims are measured against, and fails loudly when a claim drifts.

Arc note — 2026-07-28 (39331dbc). PR-4 (curated-suite membership ratchet + the gate-guarding pin promoted onto the gate), PR-6 (three doctrine prohibitions pinned), G-7 and G-9 CLOSED. The layer's central finding is sharper than the card states: the recurring defect is not absent pins but pins that exist and do not run — four instances found this arc (the H-13 lifecycle pin, the smoke/CI parity pin, the safety-pack fail-closed contract, the ANN import ban), each indistinguishable from coverage in any document.

---

What it is / What it does

evals/ (375 modules) holds the lane runners and their contracts. tests/ (881 modules) holds the pins. core/cli_test.py defines 21 suites as hand-curated tuples — fast, smoke, runtime, cognition, teaching, packs, algebra, sensorium, pulse, formation, proof, refusal, margin, rotor, inner-loop, phase5, phase6, adr-0024, math, deductive, full. CLAIMS.md is machine-generated from in-tree state: Tier 1 = five ratified domains from core.capability.ledger_report; Tier 2 = eleven lanes pinned by report SHA-256, where mismatch is a CI failure. workbench/ (22 modules) is the read-only operator/auditor projection. scripts/verify_lane_shas.py is the Tier-2 verifier.

The validation doctrine is local-first: the in-worktree run is the merge bar (uv run core test --suite smoke -q pre-push, larger suites pre-merge, [Verification]: on the PR). A pre-push hook runs smoke plus the warmed_session consistency lane pin, deliberately targeted at a regression class smoke does not cover. scripts/ci/local-ci.sh reads suite membership from the CLI rather than restating it, so the runner cannot drift from the hook. The interpreter contract is fail-closed: requires-python == "3.12.13" exactly, and a degraded run stamps itself NON-CANONICAL on every line — a degraded run is legitimate, a degraded run reporting itself as the real thing is not.

---

Contract

• Inputs: every other layer's runtime behavior; lane runners; pinned SHAs.
• Outputs: lane reports (JSON), CLAIMS.md rows, trace_hash, CognitivePipelineRecord, telemetry events, capability ledger, [Verification]: evidence.
• Invariants:
  • Tier-2 lane report SHA-256 must reproduce byte-for-byte; drift is a CI failure and demotes the claim.
  • Replay treats pipeline_record as critical evidence — divergence is evidence against equivalence, not wall-clock noise.
  • Pre-widening journal rows must show missing_evidence, never a synthesized green pipeline.
  • derive_evidence_digest is deterministic in field order; re-running against on-disk lane results reproduces claim_digest.
  • Unknown lane ids fail closed (lane <id> has no registered shape — introduce via ADR amendment); a broken reviewers.yaml yields an empty registry rather than silently granting audit_passed=true.

---

Design vs build

• Design: ADR-0092/0093/0096/0098/0099 (registry, contract validation, fabrication control, demo composition, public demo), ADR-0106/0109 (expert-demo promotion, lane-shape registry), ADR-0119.x (GSM8K lane + seal discipline), ADR-0167 (audit-as-evidence), docs/testing-lanes.md, docs/eval_methodology.md, AGENTS.md §Local-First CI Validation Protocol.
• Build: partial-wiring-debt.
• Evidence:
  • Eleven Tier-2 lanes carry pinned SHAs verified by CI — CLAIMS.md + .github/workflows/lane-shas.yml — would-fail-if-absent: yes.
  • Suite membership is read from the CLI by the local runner, so hook and runner cannot diverge — code-read — scripts/ci/local-ci.sh — would-fail-if-absent: yes.
  • Fabrication-control lane enforces per-class refused == n, fabricated == 0 — lane shape — core/capability/expert_demo.py — would-fail-if-absent: yes.
  • Sealed holdout discipline: 1,319 GSM8K cases age-encrypted, plaintext never on disk, no CI workflow sets CORE_HOLDOUT_KEY, team operates blind — code-read + contract.
  • Suite coverage is incomplete in a load-bearing way — measurement — no l10/always_on test appears in any of the 21 suite tuples (see M6).

---

Capacity

• Designed: every claim mechanically derived from in-tree state and verified by CI.
• Measured: 21 suites; 881 test modules; 11 SHA-pinned lanes; 5 Tier-1 domains; deductive suite at 20 files, smoke at 23, sensorium at 21, algebra at 15. The full ~12k fast-lane runs async by design so the pre-push hook stays targeted rather than gridlocking the push cycle.
• Ceilings: the smoke gate is not the full suite — a PR's own tests must be run against the rebased base pre-merge, which is process discipline rather than mechanism. Remote Actions are secondary observability only; GitHub mirror Actions are billing-locked dead signals. Roughly 31 red tests were retired historically. The sealed GSM8K holdout has never been opened against a parser with sufficient coverage.

---

Dependencies & provenance

witnesses → every layer; reads → M5 (capability ledger, reviewer registry), M4 (telemetry, trace); constrains → M0 (bit-exact parity pins).

---

Stage coverage

Stage	Verdict	Evidence
replay deterministically	covered	Tier-2 SHA pinning is mechanical and fails CI on drift; pipeline record is critical replay evidence
(evidence for learning)	covered	Sealed practice ledgers, capability ledger, lane-shape registry all fail closed on unknown shapes
(coverage of the telos-critical layer)	uncovered	No suite runs any L10/always-on pin

Zone roster: evals-determinism, tooling-cli-workbench-rs; capability ✱ (owned by M5, cross-referenced here).

Rollup note: weakest-link rollup. MV's mechanisms for the things it covers are excellent; its coverage map has a hole at precisely the layer with the least other protection.

---

Judgment

Fitness: strained. MV's design instincts are consistently right — machine-generated claims, fail-closed registries, SHA pinning, a non-canonical stamp for degraded runs, suite membership read rather than restated. These are the habits of a system that expects to be wrong and wants to find out. The strain is structural rather than qualitative: suite membership is hand-curated, so coverage is a curation artifact, and nothing detects the absence of a pin from every suite. That is how the most telos-critical layer in CORE came to be enforced by nothing while its tests exist and pass on demand.

Honest wrinkles:

• The hole is where it hurts most. M6 has a complete falsifiable soak harness with holds/bites predicate pairs — genuinely excellent test design — that no suite runs. The tests are not weak; they are unscheduled. A hand-curated suite list has no mechanism to notice.
• There is no meta-pin for coverage. The project has learned "a pin in no suite never runs" as doctrine, but nothing mechanically enforces it. A test file that belongs to zero suites is currently indistinguishable from a test file that runs everywhere. This is, in this assessor's reading, MV's single highest-leverage improvement and a strong Phase 4 Third-Door candidate: rather than curating harder, make orphaned pins detectable.
• The gap registers this layer ought to feed are dead (Phase 0 Finding 0-D): docs/gaps.md is 26/26 closed, the liveness ratchet is L10-blocked and stale, and docs/analysis/ is a chronological archive of ~130 documents with no aggregator. MV produces excellent point-in-time evidence and has no standing instrument that accumulates it.
• Long-horizon evidence has no home. The soak harness's __main__ exists; no results artifact does. Lanes are pinned by SHA precisely because point-in-time reports drift — but a lane nobody runs produces no report to pin.
• MV's honesty machinery (missing_evidence, NON-CANONICAL stamping, fail-closed unknown lanes) is a genuine model for the rest of the system and should be cited as such in Phase 5 rather than only audited.

Open questions:

• Add a meta-pin asserting every tests/*.py belongs to ≥1 suite (or an explicit exclusion list)? (→ Phase 4; likely highest-leverage single change in MV)
• Which suite should own the L10/always-on pins, given they are soaks? (→ Phase 4 / ruling)
• Should the assessment's gap register become the standing instrument MV lacks? (→ Phase 4)