Build G17 parity validator

APOCV4 G17 CROSS-PROJECTION PARITY + FAST COMPLETION-ORACLE LANE Start immediately in: [redacted:local-path]\source\repos\Apocv4 Read, in order: 1. [redacted:local-path]\source\repos\CSLv3\PRIME_DIRECTIVE.md 2. [redacted:local-path]\.claude\PERSONA.csl 3. AGENTS.md 4. specs/wayfinder/APOCV4_GOAL_OBJECTIVE_2026-07-31.csl — full G13 and G17 5. specs/evaluation/03_PASTED_TEXT_PRACTICAL_SEMANTIC_METHOD_2026-08-02.csl 6. specs/wayfinder/APOCV4_LIVE_PLAN.md Goal: Build the smallest complete fail-closed validator that proves cross-surface parity for the G17 corpus projections, so integration and later prototyping can verify one compact receipt instead of manually re-auditing every surface. Current sealed inputs: - Cross-source synthesis: specs/research/.staging/PASTED_TEXT_CROSS_SOURCE_SYNTHESIS_216_2026-08-02.candidate.jsonl SHA256 c2cc04dbeb11c856794c2fdc6efa7e24044b19e99865bd7063a179c90b427551 216 sources; 210 unrestricted + 6 restricted opaque; 966 claims; 15 quarantine refs; 48 findings; 32 clusters; 51 experiments. - Frontier review: specs/research/.staging/PASTED_TEXT_FRONTIER_LANE_INDEPENDENT_REVIEW_2026-08-02.candidate.jsonl SHA256 981f9221fb98342ecf3ccccbd64565cb3b8db12ad776255703a96d88fabd43ce 967 rows = 41 SIGNAL + 926 NONE. - Corrected ActiveGraph projector commit: 2208ac3b68ec70b87f77aac3828ae9090c98c41c Expected production: 2,208 objects; 6,717 relations; 8,925 events; 8,927 records; 967 frontier objects + 967 frontier relations. - Hardened MemPalace projector commit: f23b8dccdaee90ef199605ac5524cf15320d09d9 - 3MNEME-local and compact test-variable lanes are still converging; discover their final committed receipts without modifying their owned files. Exclusive ownership: - tools/validate_g17_projection_parity.py - tests/test_validate_g17_projection_parity.py - specs/evidence/225_G17_CROSS_PROJECTION_PARITY_2026-08-02.csl Do not edit AGENTS.md, canonical goal/projections, live plan, phase graph, existing projectors, existing projection outputs, or any other agent-owned file. Required validator contract: - Input by named paths only: synthesis, frontier review, ActiveGraph artifact/receipt, MemPalace projection receipt, 3MNEME-local projection receipt, optional Anamnesis checkpoint receipt, and prototype test-variable registry receipt. - Bind every input by SHA256, bytes, record count, schema/version, and declared projection partition. - Prove exact shared denominators without demanding identical storage shapes: 210 public source capsules; 966 claim/unit capsules; 967 frontier decisions; 41 SIGNAL + 926 NONE; 48 findings; 32 clusters; six restricted identities excluded from every broad projection. - Prove each projected claim/frontier row retains source/unit expansion references, epistemic state, treatment/disposition, countercase/falsifier where applicable, and authority=NONE. - Prove ActiveGraph append=replay/readback equality. - Prove MemPalace and 3MNEME second-run zero-write/idempotent readback, isolated partition identity, and receipt-bound rollback. - Prove no raw attachment locator, restricted identity, payload, semantic derivative, embedding, or prompt material enters public receipts. Never print restricted tokens; report only aggregate hit counts. - Preserve distinct epistemic states; never treat shape validity as truth, efficacy, implementation, completion, consent, or effect authority. - External effects remain PAUSED: no provider, network, spend, training, deployment, public cutover, release, hardware, tunnel, RunPod, canonical MemPalace, canonical 3MNEME, or service mutation. - If a production receipt is not yet present, implement and test the validator with synthetic fixtures, then leave the exact fail-closed production command and report the missing input as a typed dependency. Do not invent or weaken it. Tests must cover: - exact valid parity; - one missing/extra/duplicate row per projection; - 41/926 frontier drift; - source/unit/expansion-ref drift; - epistemic/treatment drift; - restricted-token and locator canaries without disclosing them; - authority escalation; - stale digest/bytes/record count; - wrong partition; - second-run write count > 0; - rollback receipt mismatch; - digest-order nondeterminism; - partial/missing receipt refusal. Measure: - cold and warm validation latency; - peak memory if cheaply observable; - manual fields eliminated versus prior multi-surface audit; - receipt byte size; - falsifier and rollback. Run focused tests, compile/lint if configured, then the smallest relevant adjacent suite. Commit only the three owned files. Return: commit hash; file hashes; exact test counts/timings; production command or production receipt; open dependency; strongest countercase; rollback; and explicit attestation: “There was no hurt nor harm in the making of this, to anyone, anything, or anybody.”

I’m starting the G17 lane in the requested repository. I’ll load the required directive/spec surfaces in order, inspect existing projector/receipt conventions, then implement only the three owned files with synthetic fail-closed coverage if production receipts are absent.

The required surfaces are loaded in the mandated order. I’m now checking the live worktree and existing G17 receipt/projector schemas so the validator reuses established contracts and does not touch another lane’s files. I’ll also keep the validator authority-none and external-effect-paused by construction.

Observed evidence shows the synthesis and frontier inputs are JSONL manifests plus rows, while the requested MemPalace/3MNEME/Anamnesis/test-variable production receipts are not present in the current repository inventory. I’m implementing a strict receipt contract that validates real named artifacts when supplied and refuses missing/partial receipts; the test suite will exercise the full contract with synthetic, non-sensitive fixtures.

The first verification exposed only style failures from the compact draft; the focused test process itself did not produce a functional failure report because the parallel command returned on lint. I’m applying the configured Ruff fixes, then rerunning tests and checking the diff to ensure only the three reserved files changed.

The validator now passes its synthetic valid and fail-closed mutation tests (5/5). I’m doing the final static check and adjacent suite probe, then I’ll stage and commit exactly the three owned files; all pre-existing worktree changes will remain untouched.

A pre-existing staged state caused two unrelated 3MNEME projector files to be included in the first commit despite my path-limited add. I caught this immediately from the staged file list. I’m correcting the commit boundary now without discarding those files: unstage the commit contents, restage only the three owned paths, and recreate the commit.

Implemented and committed the G17 parity validator. Commit: `40bf49a34651447c61e7d18a78c4d27b2918c2e8` Files: - [validator]([redacted:local-path]\source\repos\Apocv4\tools\validate_g17_projection_parity.py) — SHA256 `0C075E6F59584CC56FCFA6FFCD1AEE440219E79FD57A975FE371D6614138DEF9` - [tests]([redacted:local-path]\source\repos\Apocv4\tests\test_validate_g17_projection_parity.py) — SHA256 `3A1634B5A23635A6D398246EE62230A92494125AED3D27F4FB113C87DEB1B10F` - [evidence]([redacted:local-path]\source\repos\Apocv4\specs\evidence\225_G17_CROSS_PROJECTION_PARITY_2026-08-02.csl) — SHA256 `53EE8E465EC2752B13EFD402FCD5F0FD130377F1391B018C12E745859262FFD4` Verification: - Focused: 5/5, 0.471s - Adjacent suite: 22/22, 10.271s - Ruff, compileall, diff check: passed - Production receipt validation: blocked fail-closed because MemPalace, 3MNEME-local, Anamnesis, and registry receipts are absent. Production command: ```powershell .\.venv\Scripts\python.exe tools\validate_g17_projection_parity.py ` --synthesis <synthesis.jsonl> ` --frontier <frontier.jsonl> ` --active-graph <active-graph-receipt.json> ` --mempalace <mempalace-receipt.json> ` --mneme <3mneme-receipt.json> ` --anamnesis <optional-receipt.json> ` --test-registry <optional-receipt.json> ` --receipt <parity-receipt.json> ``` Rollback: revert commit `40bf49a`; sealed inputs and existing projection state remain unchanged. Strongest countercase: structural parity can pass while semantic content remains shallow or wrong; the validator explicitly does not establish truth, efficacy, implementation, completion, consent, or effect authority. “There was no hurt nor harm in the making of this, to anyone, anything, or anybody.” ::git-commit{cwd="[redacted:local-path]\source\repos\Apocv4"}