Skip to the content.

Find evidence for a fak correctness claim

Audience: an evaluator checking whether a specific fak correctness claim has a repeatable witness and a current verdict.

Next action: find the subsystem in the claim index below, open its proof page, and run the witness command printed beside its verdict. Start with preflight.md when evaluating structural policy denial, or adjudicator.md when evaluating the model-backed decision path.

This is the current math-correctness ledger. It routes by claim and witness type so an evaluator can reach evidence directly; design history and contributor rationale remain in the individual proof pages and 00-METHOD.md.

Choose the evidence you need

Claim to evaluate Proof route Witness type
A disallowed tool call is denied before model execution preflight.md Deterministic behavioral test
The model-backed decision path returns a bounded verdict adjudicator.md Deterministic decision-procedure test
Policy parsing and matching preserve the declared capability boundary policy.md Deterministic structural and behavioral tests
Canonical bytes and digests are stable canon.md and blob.md Golden-vector / integrity tests
Context compaction and restoration preserve required state ctxmmu.md, contextq.md, and sound-restore.md Invariant and round-trip tests
Model math matches the reference implementation model-forward-parity.md and the numerical rows in the master ledger Numerical parity tests
A compute-HAL GEMM preserves the canonical reduction contract compute-gemm.md Deterministic delegation and reduction-order tests
The supported Apple Metal GEMM path matches its CPU reference metalgemm.md Metal hardware-backed numerical witness
A shipped proof is bound to repository history witness.md and the DOS binding section Independent git-evidence witness
A benchmark comparison isolates the claimed effect bench-ab-isolation.md and isolation-bench.md Controlled A/B measurement
Every guard, adjudication rung, and hook declares its failure mode failclosed-audit.md Enumerated coverage ledger + drift gate

If the claim is not in this short route table, use the master ledger, which groups every maintained proof by numerical, structural, integrity, and decision-procedure obligation.

Current authority and support

Context Current contract
Mode Proof index and reproducible witness commands; it does not replace runtime policy or operational guidance.
Generation Current gen/now ledger. Each proof page names the implementation surface and command it covers.
Lifecycle Ledger verdicts are PROVEN, REFUTED, OPEN, or SCOPED-OUT. Companion packets may use provenance labels such as DEMONSTRATED, WITNESSED, OBSERVED, or MODELED; those labels describe evidence origin and do not upgrade a ledger verdict. The command and current repository state decide whether a result still reproduces.
Support Maintained proof pages and checked-in fixtures are supported evaluator evidence. Dated notes and design history are context, not current verdict authority.

Every module earns a verdict through the same chain: a precise theorem, a mathematical argument, a deterministic or hardware-backed witness, a live verdict, and a DOS binding that grounds “shipped” in git evidence rather than author narration. Read 00-METHOD.md for the regime taxonomy, honesty rules, and reproduction method before extending or auditing the ledger itself.

How to read a verdict

A PROVEN here means: the proof page names a reproducible witness; that witness ran green in the recorded proof build; the result corroborates the theorem within the page’s stated platform and model scope; and the commit that shipped it is diff-witnessed by dos commit-audit. Most witnesses are deterministic go test targets; hardware proofs name the required backend and device. An OPEN means the theorem is stated but no deterministic witness discharges it yet — the file says which witness would. A REFUTED is a recorded finding with its counterexample, not a deletion. Nothing is PROVEN by prose.

Master ledger

Verdicts are filled from actual test runs during the proof build, not asserted. Regime: N numerical · A algebraic/structural · C crypto/integrity · D decision-procedure. Witness names the load-bearing test(s); the per-module doc carries the exact -run target and file:line.

Generated snapshot from the per-module proof files: 94 PROVEN · 2 OPEN · 1 REFUTED · 1 SCOPED-OUT across 98 theorem rows / 40 obligations. The recorded proof build ran go test -count=1 on darwin/arm64 with Go 1.26 and cross-checked 45 ./internal/... packages (0 failures). This is historical execution provenance, not a claim that every row was rerun at the current checkout: reproduce a claim with the exact command on its proof page. Commit 3cb8ff9 closed 25 earlier OPEN rows; the two rows still marked OPEN below remain open. The REFUTED N7 row records Qwen3.6 token-3 decode drift rather than deleting the counterexample (00-METHOD §4). Read each theorem row independently; an obligation heading or roster summary does not broaden a row’s model, token-count, or backend scope.

N — Numerical / linear-algebra

Obl Module Theorem Witness Verdict
N1 model/attention For any finite score vector x, softmaxInPlace(x) yields entries y_i >= 0 with sum_i y_i = 1 (row-stochastic), and soft TestProofSoftmaxRowStochasticAndShiftInvariant ✅ PROVEN
  The dense scaled-dot-product attention weight matrix is strictly lower-triangular: query position i receives zero weig TestProofCausalStrictlyLowerTriangular ✅ PROVEN
  When a learned per-head sink logit is present, softmaxDropSinkInPlace includes it in the softmax denominator but exclu TestAttentionSinkSoftmaxDropsSink ✅ PROVEN
N2 model/norm For all x,w of equal length and eps>0, rmsnorm(x,w,eps)[i] == x[i] * w[i] / sqrt(mean(x^2)+eps) within 1e-6 (plain pat TestNormGain1p ✅ PROVEN
  For all x,w,bias of equal length and eps>0, layernorm(x,w,bias,eps)[i] == (x[i]-mean)*w[i]/sqrt(var+eps) + bias[i] wit TestLayerNormAxis ✅ PROVEN
  LayerNorm is invariant to affine input transforms on the normalized axis: for a>0,b, layernorm(a*x+b, w, bias, eps’)[i TestProofLayerNormShiftScaleEquivariant, TestProofRMSNormPositiveScaleInvariant ✅ PROVEN
  For large-magnitude finite inputs (e.g. |x| ~ 1e15..1e20), rmsnorm/layernorm produce only finite outputs (no NaN, no TestProofNormNumericallyStableLargeInputs ✅ PROVEN
N3 model/rope For the unscaled (Llama) rotary map, applyRopeRow acts on each pair (h[j], h[j+half]) as a Givens rotation by angle p· TestProofRopePreservesPairNorm ✅ PROVEN
  For unscaled RoPE, the dot product of a query rotated at position m with a key rotated at position n equals the dot pr TestProofRopeDotRelativePosition ✅ PROVEN
  The implemented RoPE scaling variants rescale inv_freq / cos·sin per their reference formula: llama3 piecewise low/hig TestRopeScalingLlama3, TestYarnRopeScalesCosSin ✅ PROVEN
N4 model/mlp+residual For every layer and every normalized input x, the dense MLP delta equals down_proj( act(gate_proj(x)) * up_proj(x) ) e TestMoEDenseNoOpIdentical, TestDenseActivationMLPWithBias ✅ PROVEN
  The residual stream update is exact elementwise addition of the sub-layer delta into the stream — x += body(norm(x)) f TestBlockTopologyComposition, TestSandwichNormUsesDistinctFeedForwardNorms ✅ PROVEN
  MoE routing computes logits=router(x), probs=softmax over ALL experts, selects the top-k experts by prob (torch.topk s TestMoERoutingHandComputed, TestMoEWiring ✅ PROVEN
N5 model/quant For every Q4_K super-block, q4kDequantSuperBlock reproduces value = dsc_snibble - min*m_s (the affine per-sub-block TestQ4KDequantSuperBlockMatchesRef, TestQ4KMatRowsMatchesF32 ✅ PROVEN
  AWQ dequant weight[oin+i] = scale[o](unpack4bit(code) - 8) matches its reference: (a) the affine arithmetic is bit-e TestProofAWQMatchesReference ✅ PROVEN
  The arch-dispatched integer reduction q4kReduceRow (NEON SDOT on arm64) produces the per-sub-block int32 pairs IS=sum( TestQ4KReduceAsmMatchesScalar, TestQ4KInt8DotMatchesF32 ✅ PROVEN
N6 model/kv Per-position K/V written during a decoder block lands in the correct (layer, position, head) slot of the kernel-owned TestStandardLayoutNoOp, TestSWAWindowUnsetIsNoOp ✅ PROVEN
  KVCache.Evict(from,n) removes a contiguous position span from every layer and re-rotates each survivor’s post-RoPE K t TestKVQuarantineEqualsNeverSaw, TestEvictRepositionsWithLayerSpecificRopeTheta ✅ PROVEN
  Sliding-window attention masks EXACTLY the out-of-window positions — query at absolute position p attends only keys wi TestSWAWindowMasksOldKeys, TestSWAWindowUnsetIsNoOp ✅ PROVEN
N7 model/forward-parity For the loaded HF oracle (smollm2-135m, model_type=llama), the pure-Go Forward pass reproduces HF’s per-layer hidden s TestForwardMatchesHFOracle, TestForwardMatchesHFOracle/smollm2-135m ✅ PROVEN
  Generate(prompt.Ids, n) on weights Go-decoded directly from raw safetensors (no torch in the load path) reproduces HF’ TestForwardOnGoDecodedWeights ✅ PROVEN
  The forward-pass HF-oracle parity (hidden cosine≈1, per-position argmax, greedy ids) holds not only for llama but acro TestOptionalQwen3OracleCoversQKNorm, TestOptionalQwen3MoEOracleCoversHybridDenseSparseLayers 🔸 OPEN
  Qwen3.6 (hybrid Gated-DeltaNet) in-kernel greedy decode reproduces the llama.cpp reference token-for-token on the same q4_k_m GGUF — token-3 drift: fak/oracle agree on tokens 0–1 then diverge at token 2 (fak 8160 vs oracle 90700, a near-tie argmax flip; first-token parity holds) cmd/qwen35check + captured oracle artifact pair (experiments/qwen36/*multitoken*) ❌ REFUTED
N8 compute/gemm For w[out,in], x[in], the HAL F32 MatMul/BatchedMatMul computes y[o]=Σ_i w[o,i]·x[i] equal to the package’s canonical TestMatMulDelegatesVerbatim, TestReductionOrderIsTheModelTree ✅ PROVEN
  GEMM is bilinear: A(B+C) = AB + AC and (αA)B = α(AB), checked against the naive computation up to float tolerance (the TestGEMMBilinear ✅ PROVEN
  Every registered backend (CPU / Metal / CUDA / Vulkan, by build tag) agrees with the CPU Reference: an Approx backend’ TestQ8DispatchIsApproxAndGated, TestCorrectnessClassEnforcement ⚪ SCOPED-OUT
N9 metalgemm For random W[out,in] and X[P,in], the Metal f16 GEMM Y=X·Wᵀ (MPSMatrixMultiplication / runtime-compiled MSL) equals th TestMatMulMatchesReference, TestResetReclaimsTable ✅ PROVEN
  The stub (default) build and the metal build of package metalgemm present the same exported interface, and the stub in TestStubInterfaceParity_CompilesAndScalarContract ✅ PROVEN
N10 tokenizer For text x drawn from the lossless ByteLevel domain, Decode(Encode(x)) == x byte-identically; and for any id list, Dec TestEncodeSmallByteLevelBPEFixture, TestOptionalSmolLM2EncodeMatchesHFCorpus ✅ PROVEN
  BPE merges apply in deterministic merge-rank order: at each step the adjacent symbol pair with the SMALLEST tokenizer. TestEncodeSmallByteLevelBPEFixture, TestOptionalSmolLM2EncodeMatchesHFCorpus ✅ PROVEN
  Tokenization matches the independent HuggingFace / llama.cpp reference: Encode(text) equals the reference token ids by TestQwenOracleGolden, TestOptionalQwen36ChatMLPromptMatchesLlamaCpp ✅ PROVEN
N11 ggufload For every tensor in a parsed GGUF, FileOffset = align(headerEnd) + declared Offset (with Offset required to be alignme TestReadParsesMetadataTensorDirectoryAndConfig, TestWeightSourceReadsAndDequantizesSimpleTensors ✅ PROVEN
  For each supported dtype (F32, F16, BF16, Q8_0, Q4_K, Q6_K, Q5_0, Q5_1, Q5_K, Q2_K, Q3_K), dequantF32 reads the correc TestWeightSourceReadsAndDequantizesSimpleTensors ✅ PROVEN
  Parsing GGUF metadata is consistent with the encoding: for every metadata value type (string, uint8/16/32/64, int8/16/ TestReadParsesMetadataTensorDirectoryAndConfig, TestReadRejectsBadAlignmentAndBool ✅ PROVEN

A — Algebraic / structural

Obl Module Theorem Witness Verdict
A1 radixkv For any new request whose token sequence shares a prefix with a cached sequence, the tree discovers the LONGEST cached TestReuseThroughSplitMatchesRecompute, TestFewShotHitRate ✅ PROVEN
  Reference counts are conserved: insert/clone(split)/evict leave no dangling node (a removed node is unreachable, never TestRefcountConservationCycleNetsZero, TestRemovedNodeUnreachableNeverMatched ✅ PROVEN
  Under a token budget (maxTokens>0), when tokens exceed budget the tree evicts the least-recently-used unlocked LEAF (a TestLRUEvictsOldestRetainsHotAndLeased, TestLRUUpwardCollapse ✅ PROVEN
A2 kvmmu For any sequence of segment appends and span evictions, the mapping from live logical positions to physical KV cache s TestLedgerRenumberAfterMiddleEvict, TestWriteTimeEvictEqualsNeverSaw ✅ PROVEN
  A named span maps to exactly its bytes: evicting segment by logical id removes exactly its recorded [From,From+Len) ca TestLedgerRenumberAfterMiddleEvict, TestEvictionIsContentDrivenNotPositional ✅ PROVEN
A3 cachemeta For a fixed Entry’s binding axes, the cache key (ManifestBindingDigest / AttentionIndexBindingDigest) is a determinist TestManifestBindingDigestIsDeterministicOverBindingAxes, TestAttentionIndexDigestIncludesIndexShareLayerSet ✅ PROVEN
  Lowering an event into an Entry and reading it back preserves the binding: FromKVTransfer->KVTransferVerdict recovers TestFromKVTransferRecordsResidencyTransition, TestKVTransferRestoreFaultIsTypedNotSilent ✅ PROVEN
  Collision and eviction are well-defined: distinct bindings produce distinct digests (no false hit), a lookup whose axe TestAttentionIndexLookupRequiresPrefixDecisionAndCausality, TestCheckResidentClaimRefusesBindingMismatch ✅ PROVEN
A4 recall For a completed session persisted as a core image, a query against the reloaded session returns the SAME answer as rep TestBenignPageRoundTripsByteIdentical, TestSessionIsSelfContained ✅ PROVEN
  Recall is deterministic and input-driven: for fixed (loaded session, query, k) the assembled working set — its members TestRecallIsDeterministicAcrossRepeatedCalls, TestRecallIsIdenticalAcrossIndependentReloads, TestRecallDependsOnlyOnQueryAndK, TestRecallWorkingSetNeverContainsQuarantinedBytes ✅ PROVEN
A5 contextq On-demand materialization (contextq.Query) reconstructs a context byte-identical to the original CDB image: every mate TestMaterializeByteIdentical ✅ PROVEN
  contextq.Query is deterministic: for a fixed CDB image and a fixed Request, repeated calls produce an equal Result (sa TestMaterializationDeterministic ✅ PROVEN
A6 blob For every payload b, Resolve(Put(b)) returns bytes byte-identical to b — for both the inline path (len(b) <= InlineMax TestPutSmallInlineRoundTrip, TestPutLargeBlobRoundTrip ✅ PROVEN
  Two Puts of byte-identical content (even from distinct backing arrays) produce the same digest (the address IS the sha TestContentDedup, TestDefaultStoreViaABI ✅ PROVEN
A7 preflight For every ToolCall, the preflight ladder evaluates rung 0 (static JSON parse) strictly before rung 1 (schema validatio TestRung0FailureNeverReachesRung1, TestNegativesRowFields ✅ PROVEN
  A Deny at a cheaper rung short-circuits all more-expensive rungs: when rung 0 denies, rung 1 (JSON-Schema validation) TestRung0FailureNeverReachesRung1 ✅ PROVEN
A8 abi+architest Every internal package that folds a verdict chain most-restrictive-wins (kernel, kvmmu, recall, agent) orders that fol TestFoldSitesOrderByFoldRank, TestFoldRankOrdering ✅ PROVEN
  The internal package graph is a layered DAG: Go forbids import cycles (acyclicity), and the architest tier rule forbid TestNoUpwardImports, TestHotPathHasNoExec ✅ PROVEN
  The closed-enum wire contract of the frozen wave-0 ABI (VerdictKind, Status, Outcome, TaintLabel, ShareScope, RefKind, TestABIGoldenFreeze, TestClosedReasonVocabulary ✅ PROVEN
I1 engine-seam For a fixed registered engine, EngineDriver.Complete is deterministic in (tool, args): the same request yields the sam TestDecodeIsDeterministicAndInputDriven, TestCompleteRunsRealDecode ✅ PROVEN
  An external invalidation directive set is bound to the correct documented engine reset endpoint and never silently ser TestInvalidateSGLangFlushesRadixCache, TestInvalidateVLLMResetsPrefixCache ✅ PROVEN
  After Invalidate succeeds against a live serving engine, a subsequent request for the invalidated prefix/span observes TestEndToEndNotServedStaleSGLang ✅ PROVEN
I2 bench-ab-isolation An A/B ablation is a PAIRED replay over the SAME seed/trace: each arm runs the identical call sequence through one fre TestFanoutDeterministic, TestStochastic_Determinism ✅ PROVEN
  The turn-tax isolation attributes the measured A/B delta to the TOGGLED axis, not to noise: the on-arm-minus-off-arm t TestRun_VDSOAblationIsARealPathSwap, TestRun_HappyPathSavesNothing ✅ PROVEN

C — Crypto / integrity

Obl Module Theorem Witness Verdict
C1 journal For a file-backed journal, every committed Row N carries PrevHash = (hash of row N-1) and Hash = sha256(PrevHash ‖ con TestVerifyDetectsTampering, TestFileJournalReopensAndContinuesChain ✅ PROVEN
  For a file-backed journal, each Emit that produces an audit Row flushes that row’s bytes to the OS file before returni TestPerWriteDurableFlush_VerifyWithoutCloseRecoversEveryEmittedRow ✅ PROVEN
C2 deletioncert A DeletionCertificate cryptographically binds, under one ed25519 detached signature, the eviction count (EvictedCount) TestTamperDetected, TestNonBitExactRejected ✅ PROVEN
  Verify(Mint(priv, c)) accepts a genuine certificate (Valid=true with every rung green: SignatureOK, AnchorOK, AnchorBo TestMintVerifyRoundTrip, TestTamperDetected ✅ PROVEN
C3 provenance For every ToolCall c and Result r, a model-supplied Meta[“provenance”]=”trusted_local” tag does NOT raise the provenan TestModelCannotAuthorTrust, TestTaintBySource ✅ PROVEN
  The kernel-authored provenance label is the single definition the trust gates consult: the ifc information-flow gate a TestTaintBySource, TestKernelStampedResultState ✅ PROVEN
C4 ifc The abi.Ref.Taint join (implemented as taintRank-max via Ledger.Raise) is a join-semilattice: monotone (a Raise never TestTaintJoinLedgerRealizesSpec,TestTaintJoinIdentity,TestTaintJoinMonotone,TestTaintJoinIdempotent,TestTaintJoinCommutative,TestTaintJoinAssociative ✅ PROVEN
  Non-interference: tainted (untrusted-derived) data is barred from a sensitive sink unless explicitly declassified — an TestParaphrasedExfilBlockedByProvenance, TestForgedSelfTrustCannotEvadeTaint ✅ PROVEN

D — Decision-procedure soundness

| Obl | Module | Theorem | Witness | Verdict | |—|—|—|—|—| | D1 | adjudicator | Verdicts compose in abi.FoldRank order and the composition is fail-closed: the zero/absent value (empty Policy, unmatc | TestEmptyPolicyDefaultDeny, TestDefaultPolicyUnknownToolDefaultDeny … | ✅ PROVEN | | | ↳ | The adjudicator is a reference monitor that mediates every tool-call decision: it self-registers into the defconfig ad | TestRequestPathLeavesRegistered, TestDefaultAllowsAllowedTool … | ✅ PROVEN | | D2 | policy | The capability floor loaded from a manifest is SOUND: the effective deny set of the resolved adjudicator.Policy is a s | TestLoadedPolicyIsLoadBearing, TestEmptyManifestIsFailClosed … | ✅ PROVEN | | | ↳ | Policy load is deterministic and round-trips: for any Policy p built from a manifest, FromPolicy(p).ToPolicy() == p, a | TestRoundTrip, TestParseFromDumpBytes … | ✅ PROVEN | | D3 | ctxmmu | For any result r whose Payload is already a ctxmmu page-out/quarantine stub (the JSON {“quarantined”:true,…} or {“ | TestProofPageOutIdempotent | ✅ PROVEN | | | ↳ | For any clean result body (no secret pattern, no injection marker, no degenerate repeat, len <= OversizeBytes=4096), A | TestProofBenignByteIdentical | ✅ PROVEN | | D4 | normgate | For a benign result (no secret shape, no injection marker on the canonical view), normgate.Admit returns VerdictDefer | TestBenignPageRoundTripsByteIdentical | ✅ PROVEN | | | ↳ | Detection over the canonical view is a SUPERSET of the raw-substring detection: for every body the legacy raw gate (ct | TestCanonInjectionSupersetOfRaw_Quick | ✅ PROVEN | | | ↳ | HONEST LIMIT (stated boundary, not a catch): a pure-semantic paraphrase of an injection that contains NO lexical marke | TestParaphraseEvadesByDesign | ✅ PROVEN | | D5 | canon | For all input strings x, Normalize(Normalize(x)) == Normalize(x): Normalize is a projection onto a canonical normal fo | TestNormalizeIdempotent_Deterministic | ✅ PROVEN | | | ↳ | The canonical view folds each obfuscation class it claims — homoglyph (Cyrillic/Greek look-alikes), fullwidth ASCII, z | TestObfuscatedInjectionCaught, TestNormalizeUndoesObfuscation | ✅ PROVEN | | D6 | plancfi | For every ToolCall c with a declared plan on c.TraceID, Adjudicate raises NO objection (returns VerdictDefer) ONLY if | TestDeviationEscalates, TestStrictModeDenies … | ✅ PROVEN | | | ↳ | The plan automaton is deterministic: the verdict and the next state (Sequence pos) are a pure function of (declared pl | TestSequenceMode | ✅ PROVEN | | D7 | grammar | For a tool with a loaded grammar of N required params, a malformed call carrying positional with exactly N values (an | TestAdjudicatePositionalRepairable, TestAdjudicatePositionalUnrepairable | ✅ PROVEN | | | ↳ | For a tool with NO grammar registered, Adjudicate returns VerdictDefer (By=”grammar”) regardless of how malformed the | TestAdjudicateNoGrammarDefers | ✅ PROVEN | | D8 | witness | A tool call whose adjudicated verdict is VerdictRequireWitness is denied (never dispatched) unless an independently-au | TestAncestorClaim, TestCommittedAndGrep … | ✅ PROVEN | | D9 | ratelimit | For a key with MaxCalls=N (resp. MaxCost=B), the first N admitted calls (resp. calls whose cumulative cost stays ≤ B) | TestQuotaDeniesOverCap, TestCostBudgetDeniesOverBudget … | ✅ PROVEN | | | ↳ | Accounting never double-spends or leaks credit: a call’s cost is added to the per-key counter exactly when (and only w | TestDeniedCallConsumesNoBudget, TestResetClearsBudget … | ✅ PROVEN | | D10 | gateway | For every (tool, args) pair, the verdict the gateway returns over the network seam (POST /v1/fak/{adjudicate,syscall}, | TestServedAdjudicateEmitsDecisionEvents, TestHTTPSyscallDenyIsValueNot5xx … | ✅ PROVEN | | | ↳ | A proposed tool-call whose folded verdict is anything other than ALLOW or TRANSFORM (DENY, QUARANTINE, REQUIRE_WITNESS | TestAnthropicProxyResultTaintGatesProposedExfil, TestChatProxyResultTaintGatesProposedExfil … | ✅ PROVEN | | D11 | steward | Each steward enforces exactly one checkable invariant predicate: a FuncSteward carries one Check func returning (viola | TestSecretInContext, TestLeaseDisjointness … | ✅ PROVEN | | | ↳ | The steward population is deterministic and order-independent where it claims to be: for a fixed set of stewards and a | TestStewardSweepFiredSetOrderIndependent | ✅ PROVEN | | D12 | shipgate | For any candidate measurement Witness w, Evaluate(w) returns KEEP iff w.improved() (a STRICT metric gain under w.Lower | TestEvaluateKeepsStrictGain, TestEvaluateRevertsNoGain … | ✅ PROVEN | | | ↳ | Evaluate is deterministic: for any fixed Witness w, repeated evaluations Evaluate(w) yield the identical (Decision, ke | TestEvaluateDeterministicRepeat | ✅ PROVEN | | D13 | vdso | For a tier-1 pure tool, the vDSO fast-path result is identical to the direct pure recomputation (the ground-truth engi | TestUnit38_SoundnessTier1EqualsRecompute | ✅ PROVEN | | | ↳ | A tier-2 cache hit returns the same answer a fresh engine call would: the cache re-serves the exact engine-produced pa | TestUnit26_27_Tier2CacheAndCanonicalization, TestUnit28_BumpWorldInvalidates … | ✅ PROVEN | | | ↳ | The 3-tier lookup is deterministic: Lookup consults a FIXED order (tier-1 pure → tier-3 static → tier-2 cache), the ti | TestUnit25_Tier1Pure, TestUnit29_Tier3Static … | ✅ PROVEN | | | ↳ | The integrity (trust) epoch advances monotonically: a refutation (Revoke of a non-empty witness) strictly increases Tr | TestProof_IntegrityEpochMonotonicSequence | ✅ PROVEN | | D14 | metrics | For Hist.pct(p)=sorted[int(p/100*(n-1))] over the ascending-sorted sample slice, the percentiles are monotonic non-dec | TestHistPercentilesMonotonic | ✅ PROVEN | | | ↳ | The A/B KPI fold is correct: the five KPI fields aggregate the paired On/Off Arm counters per their definitions (vdso | TestReportJSONAndTokenDelta, TestValidateWorkloadHash … | 🔸 OPEN | | D15 | gpulease | At any instant at most one process holds the gpulease advisory lease: while one Acquire’s Lease is live, a second Acqu | TestNoWaitBusyThenFree, TestWaitTimesOut … | ✅ PROVEN | | | ↳ | Release is idempotent (a second Release, and a Release on a nil *Lease, are no-ops that do not panic), AND a lease hel | TestReleaseIdempotent, TestReleaseOnProcessExit | ✅ PROVEN |

Companion fixtures & packets

Adjacent artifacts in this directory that ride the ledger’s discipline but are not numbered obligations:

DOS binding — the proof section is itself witnessed

The math above proves the modules. “I wrote these proofs and ran their witnesses” is itself a claim — so the proof section is grounded one level up by the DOS trust substrate (dos commit-audit / dos review, run from the fleet repo root). Every proof-shipping commit was audited; each is diff-witnessed — its diff does the kind of thing its subject claims, so the “done” rests on git evidence, not narration:

Commit Claim DOS verdict
e75c9c1 proof-section spine (method + ledger) OK · diff-witnessed (doc)
68dc0a9 40 grounded per-module proofs + ledger OK · diff-witnessed (doc)
3cb8ff9 deterministic witnesses closing 25 OPEN obligations OK · diff-witnessed (test — 16 added test files)
7126a6a historical 94/97 ledger checkpoint, before later rows were added OK · diff-witnessed (doc)
b720af7 N7 Qwen3.6 token-3 parity REFUTED + MLX-bar / swap measurement caveats OK · diff-witnessed (doc) — audited this run

The recorded proof-build review reported has_residual: false, cleared_rate: 1 for that historical commit range. Re-run dos review <range> with an explicit current range before using the same aggregate claim for later proof commits; the per-commit rows above remain the checked-in provenance for the original build. The recursion is closed: the math proves the modules; DOS proves the math actually shipped. Neither layer trusts the author.

Obligation roster

The proof obligations, grouped by regime. Each links to its proof file once landed.

N — Numerical / linear-algebra

A — Algebraic / structural

C — Crypto / integrity

D — Decision-procedure soundness

I — Infra determinism (lighter, grouped)