neural/trident/audit/full-release-preparation.md

π 0.0% 4351 words · 22 min

Full release preparation — 2026-09-16

Status: published — Trident 0.3.0, Trisha 0.3.0 and Joy 0.5.0.

All scoped release gates are closed. See the published release report for exact source identities, six native targets, the 198-proof baseline gate, the 36-pair proof exchange, Neptune admission and artifact links. Broader nox and live database features remain roadmap work.

Historical pre-release checkpoints

The following entries describe earlier candidates and do not represent current release blockers.

2026-09-16 continuation: FINAL6 remains blocked by the confirmed source-RAM clobbers in the scratch review. Its proof/node/packaging gates were stopped before execution. Live-source return cleanup, deep-stack access and inline assembly now preserve RAM in the regression suite. All133 baseline execution fixtures pass; across99 positive cases51 use fewer cycles and48 are unchanged. Three self-hash-bound inputs were independently regenerated; the other96 inputs and all hand programs remain unchanged. Six fresh CLI counterexample STARKs verify and six changed public outputs reject. The exact new RAM smoke section adds two fresh proofs through Trident delegation and two altered-output rejections. Exact evidence: RAM repair. Full workspace reruns pass: Trident896, Trisha460 (six explicit ignored proof gates), Joy71; all24 release-tooling tests pass. A Keccak struct performance regression found during the first full rerun was fixed without raising its 10-million-cycle limit; the complete rerun passes with both50-word vectors. Fresh coordinated platform/packaging proof gates remain open. Earlier receipts do not certify these edits. The prior FINAL5 full-baseline receipt and its /tmp work directory are absent on this continuation; that run's completion is unconfirmed.

Current authoritative checkpoint: frozen FINAL5 has native macOS/Linux builds, complete installed proof smoke, byte-identical binary repacks, fresh-unpack verification in both directions and actual isolated Neptune node admission. The exact frozen macOS candidate had started the complete198-proof baseline gate; its final receipt has not been recovered. Later live tooling/architecture work is absent from FINAL5 and requires explicit source/artifact reconciliation. See FINAL5 evidence. Historical 198-proof results below retain their documented proof-time provenance limits.

The next exact source snapshot is FINAL6, SHA-256 0219287c79902bad8b31433a065310766d5f8f3c9f0701f11422998e786ec0c3. Its11 repositories/2534 source entries and557 vendor entries verify after extraction and reproduce byte-for-byte. Native rebuilds and installed checks are tracked separately from FINAL5. The preserved Neptune intent has now aged past its10-hour acceptance window; FINAL6 node admission requires a fresh genuine SingleProof. No timestamp/expiry rule is relaxed to reuse it.

Post-FINAL5 live changes: the shared TIR now uses the selected XField width; Trisha exclusively legalizes machine stack depths. All133 baseline executions pass,48 positive fixtures use fewer cycles and none regresses, but40/84 native program identities change. Three custom-token fixture statements were rebound to the new actual self-hash through the independent codec generator. Exact comparison: owner legalization. Changed native programs need fresh proofs; frozen FINAL5 evidence cannot be relabeled. Ordinary Python baseline startup now preserves source inventory. Joy/Trisha artifact readers reject blocking/special inputs and enforce bounded reads; actual CLI regressions and old-proof compatibility pass. See file admission review.

The experimental tagged noun kernel has a separate draft contract and21 native/ constraint tests covering shape, hash, equality, word operations and quoted composition. It is not selected by any production proof format. A separate Trisha checker integration executes both branch shapes under the same program; its genuine proof test now passes with two fresh Triton7 STARKs and complete opposite-branch claim-transplant rejection. This remains experimental, with no production protocol dispatch. Actual proof evidence. Pure scalar helper formal analysis passes17 actual CLI/Z3 tests. The repeated48-module inventory still has1,320 UNKNOWN and four UNSAFE guard verdicts out of1,324 functions; no module is certified. See formal evidence. These are bounded steps toward the open general nox and whole-library formal objectives, not completion claims.

The complete current Zheng workspace passes211 tests, including both exhaustive wire-corruption tests, with zero failures/ignored/Rust warnings. Its release workspace serde check also passes. Hemera's reviewed external changes retain the existing constants/profile/framing and fixed vectors; eight actual vector/ profile tests pass. BBG and other sibling changes preserve the reviewed product dependency identity set and public certificate encoding. Their exact compatibility reviews remain distinct from source stability and cryptographic assurance.

Current complete Trident neural workspace:894 tests pass. Current Joy workspace: 71 tests pass, including real private proofs. All24 release Python tests pass, covering exact source verification, snapshot fidelity, reproducible archives, binary receipt binding and complete proof-event accounting. Exact downstream workspace/check results are retained in their continuation receipts; the Trisha ordinary suite passed450 tests before a missing-artifact doctest failure, whose separate workspace documentation rerun passes after dependency rebuilding. The original failure and actual rerun are both retained in workspace evidence. Joy's all-target/ all-feature check also passes. Exact24-test tooling receipt: release tooling. These live-source checks do not replace the frozen installed-platform gates.

Dependency migration is applied: the default backend and recursive verifier now use pinned Triton VM/AIR/ISA/tasm-lib7.0.0, with native claim version5. The prior Triton2 backend lacked the AIR soundness fixes shipped in Triton4–7, including the Program Table fix in7. Historical Triton2 roundtrips below are regression evidence only; their release proof gates must be rerun on7. Upstream changelog.

Trisha owns the pinned dependencies, GPU overlays and native/CCS formats. Joy's JOYZK003 envelope requires zheng-ccs-triton7-zk-v2 and rejects prior private formats. Joy's current 62-test release suite passes on7, including genuine private execution and state proofs. The canonical Neptune SingleProof gate passes, including owner CLI submission and exact mempool admission on an isolated pinned Neptune 0.15.1 Testnet(1) node. All three archive-built macOS binaries pass complete installed smoke with recursive/public/private/state proofs. All 43 baselines now pass 198 fresh source/hand proofs for their 99 positive fixtures; all 34 rejection vectors pass execution. Later checkpoints supersede historical tests.

Current implementation

  • Real Triton #[test] runner: preserves actual main/helpers/imports/cfg, compiles selected tests into typed TIR, executes bounded VM tests independently.
  • Independent baseline vectors uncover and repair actual pow/hash/sponge/Merkle, RAM block, XField assignment and nested aggregate-field ABI defects. The suite keeps all 43 baseline files in the denominator and proves both candidates.
  • Zheng derives a bounded global execution CCS; atom calls validate private witnesses, inactive arms do not require successful inverse/word/call operations, pair equality binds all structural digest limbs. Public format v2 is distinct from the earlier development certificate.
  • Private execution: verifier-derived Zheng CCS checked by a deterministic Trisha-owned Triton7 program, native randomized STARK, Joy JOYZK003 artifact. Proof metadata cannot select another checker/program or replace public values.
  • Private queries over public state now select namespace/index/value inside the proved CCS from authenticated complete tables. All10 public namespaces required; 2048 total fields maximum. Actual CLI proof/fresh verification passed; private key absent from public statement, forged root/table/output rejected.
  • Public state: injective versioned BBG dimension serialization, Lens systematic commitment v2, complete authenticated dimension certificates. Zheng pins actual namespace/key/value/activity and four root limbs to the same execution witness. Joy JOYST001 carries the certificate and independently checks native agreement.
  • Nox compiler propagates roots through active local/imported state helpers and preserves stateless function ABI. Real public/private-state CLI proofs pass.
  • Portable Poseidon2-HL matches pinned upstream Plonky3 on Triton and nox. Full-width bigint modular arithmetic and low-product Python integer vectors pass. Compiler fixes cover conditional aggregate return frames, mixed-width tuple wildcard cleanup, branch-local result widths and simultaneous tuple assignment. Early returns preserve lexical scope and stop loops. Runtime array indexing checks bounds before RHS effects and preserves aggregate layouts and neighbors.
  • BBG public query verification authenticates root, exact entity key and cell context; private contextless payloads disclose no additional tables.
  • Joy build uses the shared compile path with project/target/profile identity, atomic no-overwrite output and the specified JSON-v1 envelope.
  • Generic owner-declared intrinsic signatures bind into target compilation identity and lower through typed TargetCall. Trisha's production recursive verifier binds the complete native claim and verifies the official STARK.
  • Lens opening v3 authenticates each selected column once, while consuming every original transcript draw and enforcing the exact sorted unique set. Commitment roots remain v2. Old repeated-column openings require regeneration.
  • Neptune offline deployment inspection returns the native program hash and verifies any attached execution proof. Node status rejects neptune-cli's exit-zero connection error and malformed height/mempool responses.
  • Trisha's proof codec preserves canonical existing bytes while rejecting modular field aliases, inconsistent lengths and trailing bytes before proving APIs. Core run/prove forward warrior-owned file witnesses and digest queues.

Remaining gates (keep open until evidence is recorded)

  • Full baseline execution/proof coverage is closed for the declared 43-program inventory. The compiler pipeline remains a documented prototype and Trinity a finite-field arithmetic demonstration; those fixtures do not establish a complete self-hosted compiler or secure FHE. Broader promised implementations remain subject to their own acceptance requirements.
  • Dynamic continuations/variable noun shapes and scalability beyond bounded public table selection. Recursive Triton SDK proofs and pinned Neptune transaction validation have separate completed receipts below.
  • Hemera is explicitly novel and unaudited upstream. Correct execution of its constraints is not external cryptographic assurance for all roots/transcripts.
  • Funded Neptune wallet construction, public-network operation and block confirmation. Canonical output/intent/SingleProof validation, authenticated adapter submission and actual isolated Testnet(1) mempool admission pass. The host's older node installation was not used for that gate; the pinned Neptune 0.15.1 node was built and run inside an isolated Linux namespace.
  • Final supported-platform installed proof matrix and artifact checks after remaining source changes. Complete downstream suites and security/format regressions have current receipts below. Legacy contextless PCS APIs retain explicit limits.
  • Coordinated breaking versions, migration notes, fresh reproducible source and binary archives, isolated installed CLI smoke with new private/state protocols, checksums and release notes tied to final commits. Fresh working-tree source and binary rehearsals exist; they are not final committed release artifacts.

Detailed test logs and exact results are recorded when each gate completes; this file is a work ledger, not a release readiness assertion.

Verified current checkpoints

  • Joy complete workspace60tests pass on Triton7 with zero warnings, including build5tests and real JOYZK003 private execution/state proofs (/tmp/joy-triton7-full-release.log). Old envelope and backend formats reject.
  • Zheng execution32tests pass, including malicious read-provider witnesses that fail the CCS itself, all four root limbs, private calls and pair equality. /tmp/zheng-private-state-tests.log. Default library172tests pass after legacy migration. Fresh full release/serde workspace185tests pass, including178 library tests and exhaustive wire mutations (/tmp/zheng-opening-v3-full.log). A subsequent37-test execution run adds five exact-budget branch regressions: only the selected branch's bound cost consumes budget, with inactive-operation and forged-cost checks retained.
  • BBG71tests pass against opening v3 (/tmp/bbg-opening-v3-full.log); full Lens workspace129tests pass with all features in release after independent review (/tmp/lens-opening-v3-reviewed-full.log). This includes canonical column bounds before hashing, independent full transcript draws and legacy-wire rejection. No sampled PCS128bit claim is made.
  • Bigint full8 regression tests pass in release (7.52s), plus6 broader boundary tests under both debug/release source profiles. Logs: /tmp/trisha-bigint-release.log, /tmp/trisha-bigint-boundaries.log. Four frame and five control-flow real-VM regressions pass; early returns9 and runtime arrays9 pass in both source profiles (/tmp/trisha-early-return-final.log, /tmp/trisha-array-final.log). Trident763 default and807 with neural pass after the array edits (/tmp/trident-post-array-full.log, /tmp/trident-neural-post-array.log). The later CLI transport regression and five existing dispatch tests pass (/tmp/trident-input-transport.log).
  • Standard Poseidon2-HL matches pinned Plonky3 on Triton and nox. The independent hand/source fixtures produce equal digests: source7105cycles, hand7027cycles.
  • Historical Triton2 recursive SDK checkpoint (must regenerate on7): tests pass against actual official proofs, including changed claim program/version/input/output/length and malformed proof rejection (/tmp/trisha-recursive-production-tests.log). The explicit expensive outer proof test passed: 426962cycles, padded524288, 130402encoded proof fields, 88.656s generation, peak19.23GiB, no swaps (/tmp/trisha-recursive-outer-proof.log). The test is deliberately ignored in ordinary suites and must be invoked with --ignored for the release gate. All four recursive baselines pass full proofs for both source and hand programs: eight generated and freshly verified STARKs (/tmp/trisha-recursive-baselines-full.log). Aggregation852915/852860 cycles, padded1048576 each; two proofs598.64s, peak RSS34.00GiB, no swaps. OS peak memory footprint51.26GiB includes compressed memory; this differs from peak RSS and must remain visible in resource guidance.
  • Right-shift lowering uses the correct numerator/divisor order and keeps the quotient. Two word-operation tests cover512 real VM vectors; two canonical proof-wire regressions pass (/tmp/trisha-word-wire-final.log).
  • CLI file witnesses and explicit expected-claim preparation pass real process tests, including recursive execution, incorrect claims, private file mode and input bounds. Installed smoke now includes the full recursive outer proof; it has not yet been run against a fresh final archive.
  • Neptune inspection native-hash/real-proof test and two CLI tests pass (/tmp/trisha-deploy-rs-tests.log, /tmp/trisha-neptune-cli-tests.log). These inspect programs and mocked/read-only status; they do not validate or submit Neptune transactions.
  • Archive rehearsal built isolated Trident+Trisha+Joy without warnings, from a worktree snapshot. That rehearsal predates final private-stateformatv2 and current crypto edits; a fresh final archive/smoke is still required.

Latest verification after the backend migration

  • Trident release suite765tests passes without warnings after event and witness transport fixes (/tmp/trident-release-final.log); neural feature compilation passes (/tmp/trident-neural-final.log). Earlier807-test neural run predates the event changes and is not the final feature-suite receipt.
  • Five actual-VM event tests pass on Triton7 for debug/release source profiles: declaration-order Reveal/Seal payloads, aggregate fields, preserved locals, once-only field effects and independent Tip5 commitments. Huge aggregate and duplicate fields reject; nox explicitly rejects unsupported event operations. Logs: /tmp/events-triton7-final.log, /tmp/events-nox.log.
  • Four canonical native proof input tests pass on7, including exact proof-item consumption, hostile log-height values without panic, canonical field bytes, bounded lengths and trailing-byte rejection (/tmp/trisha7-proof-wire-guards.log).
  • Quantum hand assembly has been replaced from independent complex arithmetic and RAM algorithms. Every21pure gate is exercised on four independent vectors; four complete baseline fixtures agree (source27388–27390cycles, hand739). RAM gate fixtures cover all targets, complex coefficients, state neighbors, highest valid U32 boundary and invalid addresses/dimensions before mutation. This is finite-field simulation; legacy measurement is an exact predicate, not probabilistic quantum measurement. Logs: /tmp/quantum-pure-baselines.log, /tmp/quantum-all-gates-final.log; full Triton7 proofs remain pending.
  • Full SHA-256 compression and Keccak24rounds now have explicit bounded RAM interfaces. Independent hash/matrix vectors remain the oracle. RAM double-SHA uses188106cycles and Keccak826062cycles; these are execution observations, not proof receipts. All baseline43files stay in the denominator.

2026-09-12 continuation

  • Canonical Neptune0.15.1/HardforkGamma SingleProof generated and independently verified on Triton7 for a deterministic empty transaction, after generating and verifying all five actual subproofs. Native trace1155385cycles, padded2097152; final proof518.771seconds, pipeline531.64seconds. Peak RSS 25828950016bytes, OS peak footprint31741336592bytes, zero swaps. Four Rayon threads and TVM_LDE_TRACE=no_cache; no network submission or mock proofs. /tmp/neptune-single-proof-full.log; compiled policy wrapper/outer proof integration remains a separate gate owned by the Trisha fixture work.
  • CPU and wgpu now share the exact native input/claim/proof codec and verifier. Both reject noncanonical input, oversized FRI index domains, appended proof items and binary aliases. Private execution errors suppress VM witness dumps. ProgramInput is capped at8Mi field words, counting every digest coordinate. JSON uses an eight-byte canonical Field representation and bounded sequence visitors, with tighter expected-claim limits. Six SDK/wire tests, two JSON tests and one real-proof GPU acceptance test pass without warnings: /tmp/trisha7-input-guards.log, /tmp/trisha7-json-input-guards.log, /tmp/trisha7-gpu-proof-guards.log. All-feature/all-target workspace check passed before the coordinated product-version bump.
  • Full current Zheng release/serde workspace190tests passes, including183 library tests and both exhaustive wire mutation scans; library612.90seconds. /tmp/zheng-release-current-final.log. These include the five exact-budget regressions added after the earlier185-test receipt.
  • Coordinated versions are applied: Trident0.4, Trisha0.3, Joy0.5, Zheng0.4, BBG0.3, nox0.3 and affected Lens packages0.2. Compiler API2 rejects stale warriors; package JSON schema1 remains unchanged. All eight Cargo roots pass locked offline metadata checks, with no registry package changes. Complete map and32 focused test receipts: version closure. Individual registry crates still lack sibling resources/root patch propagation; the implemented distribution route is the full source closure plus binaries.
  • Snapshot fidelity test passes; candidate-builder Nushell syntax passes. Final archives and installed binary smoke remain pending final source changes.

Current coordinated suite and platform receipts

  • All133 reference fixtures pass:99 positive and34 rejection vectors, covering all43 manual baselines. No H0003 hints or unused-import diagnostics remain in the latest complete execution run (/tmp/h0003-bench.log). This is execution coverage. Full source/hand proofs are a separate required gate:6 heavy baselines have6 positive/12 negative fixtures, and the other37 have93 positive/22 negative fixtures. Their union is the exact original43/133 set.
  • Trisha release workspace with all features passes428 tests, zero failures and zero compiler warnings, across55 test/doc-test binaries. Three explicit heavy outer/deployment proof tests are excluded from this ordinary suite and have separate required gates. The default CLI package passes33 tests across7 binaries, with one intentional deployment-proof exclusion. The PLUMB compile regressions check actual old/new root and shared-path mutations. See trisha/audit/final-workspace-validation.md in the coordinated tree.
  • Coordinated release suites pass: Joy60, Zheng190, Lens129, BBG71, nox169 default/175 all-feature tests. Trident's full neural workspace now passes851 tests across18 test/doc-test binaries with zero failures, ignored tests or warnings after formal-audit and registry state-selection fixes (/tmp/formal-trident-workspace-neural-final.log).
  • Default CPU builds for Trident/Trisha/Joy pass Rust1.89 compilation. Linux x86_64 cross-checks pass for all three, with all features checked for Trisha and nox. Apple-only mining dependencies and dispatch are target-gated; the portable Tip5 path matches official hashes, including raw Montgomery words and padding. These are compile checks, not Linux binary execution receipts. Detailed commands: trisha/audit/platform-validation.md.
  • nox jet admission now checks exponents and exact bounded shapes before shifts, allocation or jet charging. Declined acceleration preserves normal reduction, result/error, trace and budget. The existing NTT pure anchor is only a single butterfly; the repair no longer substitutes a full transform for that different formula. A full recursive NTT anchor remains unimplemented.

Subsequent release preparation

  • All six heavy baseline programs pass their twelve fresh source/hand outer STARKs on Triton 7. The separate compiled SDK outer-proof gate also passes: 430989 cycles, padded 524288, 130.947 s proof generation. Exact paired cycles, wall times and distinct RSS/OS footprint observations are recorded in trisha/audit/recursive-release.md. The other 37 programs' 186 proofs also pass: full union43 baselines/99 positives/198 verified proofs, with all34 rejection cases passing execution. Machine-readable exact fixture accounting is in trisha/audit/full-baseline-proof-coverage.json.
  • Formal audit now uses independent function obligations and injective SSA symbols, honors scalar contracts and reports incomplete analysis as UNKNOWN. Z3 failures and counterexamples cannot become SAFE. Current full inventory: 48 modules (34 core, 14 production SDK); 46 UNKNOWN files, 2 files with counterexamples, 1321 UNKNOWN functions and 3 SAT obligations. The latter exercise intentionally rejecting guards without sufficient caller assumptions. No module is claimed formally proved. Typed aggregates, interprocedural and recursive contracts, loops, RAM and intrinsic summaries remain open. See formal validation.
  • Trisha now owns a fifth, pinned Neptune consensus/RPC adapter crate. Its output constructor binds the compiled lock, canonical UTXO and addition record; complete transaction preparation fixes the caller-selected kernel and validates a real SingleProof and every output preimage. Submission uses an explicitly configured authenticated gateway with canonical bounded RPC. The generic runtime deployment capability remains false because that trait cannot carry the required transaction intent. Both genuine proof/CLI gates pass, followed by actual installed CLI submission and exact kernel admission on the pinned isolated Testnet(1) node. Changed kernel/original proof and wrong gateway credentials reject; node height remains 0 and cleanup is confirmed. See trisha/audit/neptune-local-node-validation.md. Caller-supplied transaction validation does not implement funded wallet coin selection or establish chain finality.
  • Independent adapter review found a FIFO input-open hang; the loader now rejects nonregular files before opening and its regression passes. A distinct local-testnet1 state names chain 4/testnet-1. Upstream regtest admits mock proofs only, so it cannot establish real SingleProof acceptance. The completed real-proof node gate uses fresh genesis state/timestamp, isolated network access and exact kernel presence in the mempool after RPC acceptance.
  • Mining now follows the pinned fork schedule, verifies independent PoW/MAST layouts, synchronizes CPU/GPU winner ownership and bounds nonce reservations across template updates. Eight mining tests and seven CLI lifecycle tests pass, including actual Metal oracle comparisons. Full legacy buffer construction and actual network mining remain distinct operational gates. See trisha/audit/neptune-mining-compatibility.md.
  • Updated Trisha default CPU closure passes Rust 1.89, and its all-feature Linux x86_64 cross-check passes without warnings after the new adapter/mining changes. All three arm64 Linux binaries cross-link, with an explicit Zig linker deprecation warning in that earlier cross-build. Native Rust 1.89 release builds now pass with zero warnings in the isolated Ubuntu 26.04 VM, together with installed execution checks. The full Linux proof smoke runs separately. See trisha/audit/platform-validation.md and trisha/audit/linux-runtime-validation.md.
  • Source packaging now checks every dirty source input, creates deterministic archives with normalized metadata, and includes a separately locked state fixture helper. Both packaging regressions and the helper's Rust 1.89 run pass. New worktree snapshots remain rehearsals until final coordinated commits, binary builds and installed smoke are recorded.

Final source continuation — 2026-09-12

  • Both native archive-built platforms now pass complete installed smoke, including the real recursive outer proof, JOYEXEC2, JOYZK003 and JOYST001 positive and rejection paths. Darwin wall141.50s, monitor145.634s/peak9,351,577,600bytes; Linux monitor142.342s/peak7,818,088,448bytes. Actual binary archives reproduce byte-for-byte. They retain rehearsal1 source identity; subsequent source changes below require fresh candidates. See Trisha's installed/Linux audits.
  • nox now lowers early returns in bounded loops, including nested branches, imported functions and aggregate results. Two associated scope defects are repaired: post-loop bindings cannot leak backwards into loop lowering, and sealed shadow frames cannot retain a different aggregate type layout. The original installed candidate rejects a valid reproducer; the corrected code agrees with eight actual Triton executions. Full Trident neural workspace: 858 passed, zero failed/ignored/warnings. Current Joy workspace:62 passed, zero failed/ignored/warnings, including genuine public/private loop proofs. See scope review and independent review.
  • Parallel BBG development introduced the dependency-free neuron-id crate. Preserve that work: minimal Joy and fixture-helper lock entries add the local package without changing registry versions. The 32-byte identity alias and public proof/state formats are unchanged; all three archived public fixtures verify and regenerate byte-for-byte. Current BBG default82 tests pass without warnings; all-features152 pass with three existing Fjall warnings, recorded separately. See bbg/audit/release-dependency-continuation.md. The next source closure includes Neuron; rehearsal1 still contains exactly10 repositories.
  • Release tooling now verifies complete repository/vendor inventories before and after compilation, rejects escaping symlinks and unsafe destinations, and removes partial candidates on failure. Build, smoke and binary archive bind the exact source-verification receipt; smoke also binds the executed script, which is shipped from the verified source archive. Current verifier process/mutation tests8 and binary-packager tests3 pass; synthetic packaging tests are not proof evidence. Strict committed source manifests now include the same per-file inventories as snapshots.

Typed-entry release continuation — 2026-09-12

FINAL3 now has native macOS and Linux full installed proof smoke, reproducible binary archives, fresh extracted verification and actual Linux Neptune node admission with its exact binary. Evidence: candidate report. A subsequent correctness review found uninitialized typed Triton entry parameters. The live complete input adapter is implemented with actual VM ordering, aggregate, narrow-type and genuine proof/claim-mutation tests. It requires compiler API3; API2 descriptors are rejected before foreign compilation/dispatch. This fixes source semantics in addition to the already sound assembly/claim binding. FINAL3 remains unchanged and must not be described as ready after this finding. The next installed smoke includes this bug and nox bounded loop/scope regressions. Source/package helper regression suite:14 tests pass in16.575s on macOS (/tmp/cyber-final4-packaging-tests.log). This is tooling evidence, not proofs. The remaining formal, general nox, self-hosting, FHE, external assurance and committed/publication gates above remain open until independently satisfied.

nox source-entry and scalar formal continuation — 2026-09-12

Additional accepted-source defects were reproduced and fixed: nox entry arity, U32/Bool range admission, flat aggregate marshalling, imported array dimensions, and deep noun read/write paths. Native45-test source surface and3 Joy proof tests pass. Exact evidence.

Scalar formal audit now models bounded branches, early returns and each path's postconditions with target-correct Bool/Field truth. Twelve real CLI/Z3/native checks and83 verification unit tests pass. The complete library inventory still reports46 UNKNOWN modules and2 counterexample modules; no whole-library SAFE claim is made. Formal evidence.

Current Triton compiler output matches all84 compared native program hashes for both FINAL3 and the retained original proof-generation binary;266 baseline files match FINAL3. Historical proof logs lack full proof-time input inventories for all198 runs. Preserve that provenance limitation rather than retrospectively claiming signed or cryptographically bound historical source identities. Continuity analysis.

Generic and terminal-return continuation — 2026-09-12

Further source review found defects outside entry marshalling: imported generic bodies were not concretely checked, constant/layout ownership could be lost, Triton struct literals could reverse named fields, and nox discarded terminal conditional values. These are being repaired before the next source freeze; the earlier84-program continuity comparison predates these changes.

The nox terminal fix passes the complete48-test native source surface and two Joy proof tests, including a genuine private STARK and claim mutations. Evidence. Ordinary return types and concrete generic bodies now undergo checking; self-hosted compiler unit helpers explicitly discard their intended unused Field results. Broad suites are rerunning against the coherent changes. New installed smoke includes imported generics, named struct layout, terminal branches and invalid return-type rejection.

Strict committed-source packaging was exercised with11 isolated fixture Git repositories, retaining all2442 archived source entries and557 vendor entries. Packaging evidence. Fixture commits are not production release commits. A fresh full baseline gate will bind actual proof verifications to exact candidate and source inventories.

The coherent compiler checkpoint passes884 Trident tests (49 actual nox surface), 448 all-feature Trisha workspace tests and67 Joy tests, with zero Rust warnings. Five explicit Trisha proof/transaction tests remain excluded from that ordinary suite and retain separate proof receipts; they are not counted as passing here. All-target/all-feature checks pass in all three workspaces. Release tooling passes23 Python tests. Logs: /tmp/{trident,trisha}-final4-workspace-reviewed.log, /tmp/joy-final4-workspace.log, /tmp/{trident,trisha,joy}-final4-check.log, /tmp/trisha-final4-tooling-tests.log.

The Trisha broad run first exposed four intermediate-branch width failures and two stale generated-label assertions. The former are fixed by preserving value context while discarding intermediate branch tails; the latter now execute concrete/inferred generic calls on the actual VM. No diagnostic guard was weakened. Field literals also now emit canonical target residues, matching the native nox and corrected scalar formal semantics; U32/dimension bounds remain integer checks. Two native Triton regressions pass for these literal cases. The updated installed smoke includes the canonical Field constant case.

FINAL4 builds on both native platforms but fails the new installed generic loop case: nox did not expose an unrolled immutable loop index to array access lowering. FINAL4 is retained as failed; no binary package or full baseline proof gate ran for it. Scoped bindings now retain each iteration's known index, preserving nested/shadowed/sealed frames and U32 limits. The complete Trident suite passes888 tests, including53 native nox source tests.

The exact failed fixture now passes both profiles and both inputs803/809 through Joy, including genuine private proofs. New evidence totals8 public certificates and4 private STARKs with input/output mutations rejected: joy/audit/loop-index-smoke-review.md. A separate development CLI replay of the remaining installed-smoke section passes. It also caught and fixed a smoke-script field lookup (claim.public_input, not a top-level proof field); expected claims and modified-proof rejection now execute. This replay is not a complete installed release receipt. The next frozen candidate must repeat the full gate.

Graph