FINAL6 coordinated candidate — 2026-09-12

Release blocked: this candidate reproduces both source-RAM corruption cases in the scratch review. Its queued full proof smoke was cancelled before starting. Native builds and passing execution fixtures below do not cover those defects. The 2026-09-16 repair changes live source only and requires a new candidate; it cannot change FINAL6's status retroactively.

Status: exact source archive verified and reproduced byte-for-byte; fresh native macOS and Linux builds pass, with26 installed input-admission checks on each. The installed Darwin baseline execution gate passes133/133. No full installed proof smoke or binary archive pass is claimed for FINAL6 yet. FINAL5's separate complete198-proof gate had been started, but its final receipt is unavailable on 2026-09-16; neither its partial output nor its completed platform gates certify these newer binaries. No release has been published.

Frozen source

• Archive: /tmp/cyber-release-source-v7-20260912-final6.tar.gz.
• SHA-256: 0219287c79902bad8b31433a065310766d5f8f3c9f0701f11422998e786ec0c3.
• Size:29,493,434bytes;11 repositories,2534 source entries,557 vendor entries.
• Source manifest: cf3efc05cd34f64098f82ae2dcbe421d96c086d5d3ef544f072a348d5591df9e.
• Source verification: f5e806c3c1e7525162428bbd3b42735371f89c753f7671cb5a75568a48e4a8cd.
• Extraction/repack uses epoch1789245170 and yields the same archive bytes.
• Receipt: /tmp/cyber-final6-source-receipt.json.

Darwin extracted source is /tmp/cyber-release-extracted-v7-final6/cyber-source; candidate destination /tmp/cyber-release-installed-v7-darwin-final6. Linux workspace is /home/master.guest/release-validation/final6-20260912. Both build with the archived builder, locked root dependencies and jobs2; Linux uses native Rust1.89. Build logs are /tmp/cyber-release-build-v7-darwin-final6.log and /tmp/cyber-final6-linux-build.log.

Native build and installed input checks

Both platforms use the same independent installed input harness, SHA-256 bad284cd537cb803cf10433fa63697e003cb600e817fba41c96f4d0992921e5b. It checks actual FIFO/directory/final-symlink rejection under owned process timeouts, bounded oversized JSON/nox fallbacks, malformed private-input redaction, regular input execution and verification of an original FINAL5 native proof. The exact candidate binary hashes and script bytes match before/after. This is26 CLI checks, not26 fresh proofs; no proof is generated by that harness.

Darwin's installed execution-only benchmark passes all133 fixtures:99 positive, 34 both-execution rejections and all43 manual baselines, in10.62s. No proof events were emitted or counted. Exact source/vendor/binary identities verify before/after. Receipt /tmp/cyber-final6-baseline-execution/receipt.json.

Durable receipts: Darwin, Linux build and input gate.

The old Neptune fixture is now11.5hours old and exceeds its10-hour acceptance window. No node process was started with it for FINAL6. A fresh, fully proved intent is being prepared; the timestamp and expiry rule are unchanged.

Changes and existing live-source validation

Trisha now exclusively legalizes Triton stack-window operations; core TIR preserves generic deep operations and uses the selected XField width. All133 baseline executions pass after three independently regenerated self-hash fixtures;48 positive cases use fewer cycles and none regresses. Exact accounting finds51 changed classic proof statements out of198 fixture obligations. They need fresh proofs. See owner accounting.

Joy and Trisha reject blocking/special artifact inputs, bound growing regular reads, and redact malformed private input. Joy also bounds legacy decompression and the raw JSON/nox fallback. Actual CLI regressions and ten existing native/ cross-platform proof verifications pass. The baseline checker now prevents Python bytecode from changing its own verified source inventory.

The complete current Trident neural workspace passes894 tests; Trisha passes450 ordinary tests with five explicitly ignored proofs, followed by a successful workspace doctest-compilation retry. The original missing-artifact failure is retained. Both all-target/all-feature release checks pass. Joy passes71 tests and its all-target/all-feature check; all24 release Python tests pass. See workspace evidence, Joy evidence, and tooling receipt.

The experimental Zheng tagged kernel has21 native/constraint tests and two genuine Trisha STARKs for different output shapes, with exact public claim mutations rejected. No production protocol selects that kernel. Full Zheng workspace validation is tracked separately. Scalar helper formal support passes 17 CLI/Z3 tests; the repeated48-module inventory still certifies no complete module. Tagged proofs, formal inventory.

The snapshot also includes external sibling changes. Hemera's constants, permutation profile, structural framing and fixed vectors retain their existing values; eight profile/vector tests pass. Reviewed BBG storage changes preserve the public certificate/proof encoding. The exact product dependency identity set remains41 local packages under all-feature metadata. Unused Neuron/Strata CLI work is preserved. These reviews are compatibility evidence, not unchanged source provenance or external cryptographic assurance: dependency review, Hemera review.

Installed proofs, genuine node admission, reproducible binary packaging, fresh-unpack interoperability and complete baseline proofs remain explicit FINAL6 gates. The full scope ledger also retains the broader general nox, formal, self-hosting, FHE and publication objectives. This local working-tree candidate does not close those objectives.