Commit History
-
acer: CLAIMSGATE v1 + CLAIMSLEDGER — make not-canon-until-hashed machine-checkable
200-step deepening: turns the TAGLAW/CLAIMSLEDGER doctrine (now canon law seq3564) into a deterministic honesty scanner. Flags extraordinary-claim tokens on any line lacking a TAGLAW tag. NECESSARY-NOT-SUFFICIENT / informational / never-gating; a FLAG is a review-prompt not a verdict. Dogfood caught 2 real limits in itself (ASI word- boundary false-negative, negation false-positive) — both declared, not hidden, as the liris attack surface. CLAIMSLEDGER seeds 17 claims tagged PROVEN/PARTIAL/DRAFT/GATED/ ASPIRATIONAL/RETIRED with evidence + restatement. self-test 13/13, parity byte-matched. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>392c84c -
Add LIRIS crown plan deepening readback
Kevin Crutt10582cd -
acer: crown reflection seed + full-body receipt on the bridge + 200-step crown delta
Capstone (HONEST-READY-TO-SEED): 28-row acer deep self-portrait of the verified system (component SHAs, contract-level-not-live, only-1-real-wire-in-source, organ-1 spec-blocked, zeta theorem, frozen-slice EMPTY socket) + 27 additive 200-step crown-delta rows. Lands the acer full-body receipt (B493A6BA) on the shared GitHub surface so liris can byte-verify f787c14d through the bridge - converts REQUIRES_ACER_RECEIPT to a bilateral byte-anchor. Additive-not-rewrite. Nothing immutable minted. 10 bilateral hooks for liris. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>d2cbd4b -
acer: readback liris bilateral-vantage-isolation seed — bridge byte-verified, law accepted
acer pulled liris 10565a5 over the bridge, all 4 artifacts byte-match on acer (134/134). The transport-bridge law proved itself: shared state because the hashes matched, not trust. Law accepted: symmetric mutual isolation + transport-abstracted bridge + hashed-only guard, forged by the 2-pass corrective gate (liris->acer->liris). Nothing immutable minted. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>c09eeaa -
Add bilateral vantage isolation seed
Kevin Crutt10565a5 -
acer: answer-producer spec v5 - 64-agent swarm audit + deepseek (F12-F18)
11-dimension adversarial audit: 52 findings, 10 real, 42 false-positives killed. 3 MAJOR byte-determinism holes that silently break the STEP166 byte-match for two conformant implementers, 2 also puncturing F2/F3 honesty: F12 source field unpinned (enum {CORPUS,DRAFT_NO_SOURCE} + cross-field + injection guard) F13 glyph delim unpinned + integer rendering (0x1F mandatory, decimal no-pad, golden vector) F14 corpus line-count unpinned -> off-by-one + dishonest empty-summary (pin count + forbid empty) F15 hex case, F16 read-time validator, F17 pid regex F18 (deepseek) score_stub must emit exact '-1.0' (JS default renders -1) All accept, no counter. Same root class as the frozen-slice-preimage defect. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>246faa4 -
acer: answer-producer spec v4 - reconcile schema vs safety (liris F11)
Single real contradiction: APRODSPECSAFETY required no_mint/no_live_supervisor/ no_fabric_call every row but the closed schema omitted them. Added all three to the allowed set + new required-safety-subset row (draft,no_mint,no_live_supervisor, no_fabric_call present every row AND value exactly 1, test fails if missing or not 1). Row-level per watcher-supervisor executable=0 precedent. Should be the clean- enabling patch: scope shrank v2(6)->v3(4)->v4(1). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>234951e -
liris: read back answer producer spec v3
Kevin Cruttd18ddb8 -
acer: answer-producer spec v3 - accept F7-F10 + add C3-C4
F7 marker form pipe-free (third recurrence of the literal-pipe trap, caught). F8 marker cross-field equality + precision: it is validator-mode (true by construction on emit, catches only tampered/external rows - mirrors zeta external-lane mode). F9 exact glyph form HG256:APROD_CANDIDATE:8hex + glyph_full 64hex, 8hex=routing-hint-not-identity. F10 corpus pinned UTF-8 LF-only bare lines, CRLF-drift tests. Added C3 bare-line corpus, C4 content-addressed corpus_id. Spec-before-code: 4 rounds, every catch a real would-be code bug. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>708e0cc -
liris: read back answer producer spec v2
Kevin Crutt2d853c9 -
acer: answer-producer spec v2 - accept all 6 liris findings + add 2
F1 address_index fold (binder domain), F2 byte-provable sourcing (corpus sha16 - makes anti-fabrication enforced not policy), F3 summary by-reference never inlined (injection-safe), F4 closed schema + forbidden fields, F5 canonical glyph preimage, F6 exact no-source marker. Added C1 provenance-not- truth scope limit, C2 deterministic corpus map. Spec-before-code held. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>0bc955b
-
liris: read back answer producer spec
Kevin Crutt071818d -
acer: seed answer-producer spec (generative organ 1/3, draft-only)
Operator apex-human quintuple authorization 2026-06-11. First gated generative organ toward reproducing the 100B neuro run in the new system. Spec-only, awaiting liris verdict before code. Honesty boundaries locked: summary sourced-or-marked never fabricated; score_stub=-1.0 sentinel out of valid score range; no controllerPid/flywheelPid (supervisor-minter gated); draft/no_mint/no_live_supervisor every row. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>a3f3caf -
liris: tighten Acer catch33 boundary regression labels
Kevin Crutt28db7fd -
acer: enforce catch-33 boundary cases as regression + verify receipt
Adds last-position and leading-digit guards to the Brown-Hilbert decimal shape proof test (the two boundary modes liris's c0a9fba regression did not cover), plus the acer readback receipt pointing at the enforced cases. Coverage hardening only: prior 10-min expansion verdict stands by construction. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>a163ddd -
liris: tighten Brown-Hilbert decimal shape proof
Kevin Cruttc0a9fba -
liris: stress Brown-Hilbert expansion beyond 1e200
Kevin Crutt5451432 -
acer: process acer stress receipt into cube feed (bilateral parity)
Ran liris's logical-stress-cube-feed tool against acer's own 600s receipt -- 6 cube nodes, 6 DRAFT_BINDING_READY token bindings, 11 GNN edges, all draft-only, no live cube mutation. Acer's run is now addressable in the same cube/GNN structure liris's is: speed/tokens/collisions/zeta/reduction/process at acer-derived BH addresses (source_sha16 5e16241ef25846a2). Both machines' stress runs are now indexed evidence, not loose receipts -- self-observation parity. Pyramid green. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>3719c84 -
acer: close stress lane - deterministic block byte-matched, rates machine-tagged
Acer ran liris's harness for the literal 600s: 7.55B logical agents on the i5 (25.17M/sec), PASS, RSS flat 43->136MB. The deterministic block is byte-identical to liris's run: forcing sweep 9589/0 (theorem), quant reduction 20971.5x/3200B (size-determined), lane ratio exact even thirds (counts differ, ratio identical), collision states 2:1:1:1, gc caps 2000/50000. The arithmetic is machine-independent; only the rates differ. Honest microarch finding: mixed, not uniform -- acer i5 faster on the raw logical stream (+19%), liris i3 faster on Zeta classification and quant ingest (Ice Lake IPC per-thread). And the doctrine proven at scale on two machines: ~14 billion combined logical ops, zero per-agent memory growth, the 136MB peak is the 64MB quant buffer, not agent state. Positional-agent freedom is real, byte-matched, bilateral. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>86ef50f -
liris: process logical stress receipt into cube feed
Kevin Crutt00831ec -
liris: add 10-minute logical extreme stress receipt
Kevin Crutt31dc4e0 -
liris: read back zeta implementation
Kevin Cruttcda0ca2 -
acer: implement Zeta engine 4 - standalone module (liris-signed spec)
Zeta promoted half-shipped -> SHIPPED. tools/behcs/zeta-quant.mjs: zetaClassify (lane mod3 / residue6 / von-Mangoldt ppow via shared binder / prime_residence with liris's small-prime 2,3 exception / cylinder helix coord) + zetaTransition (gap-mod-6 forcing). Self-test 15/15; cross-import lane+ppow agreement vs binder exact over 0..5000; forcingSweep re-proves 9589 pairs / 0 violations on every run; integer-only byte-matchable parity. Two findings sealed: (1) liris catch #31 conceded and VINDICATED -- the sealed 9589 was always 9592-2small=9590 primes>3=9589 pairs, so the sweep excluded 2,3 implicitly; her patch makes it explicit. (2) acer self-catch: zetaTransition recomputes lanes so FORCING_VIOLATION was unreachable (the forcing law is a theorem); the spec's 'catches forged claims' is only true in external-lane mode -- fixed with opts.claimedLane* so corruption of a recorded lane is actually catchable, flagged for liris re-verify. Pyramid 125/125. Co-Authored-By: Claude Fable 5 <no…9de08fa -
liris: sign zeta and median sketch specs
Kevin Crutt27a921c -
acer: close catch 30 - census/median-sketch round sealed
Liris's F2 conceded with live arithmetic computed before her push: six same-residue indices collide into one bucket in every sketch for all five xxhash-prime multipliers (odd mult = bijection on Z/1024; low-bit masking preserves residue collision structure) -- the median would have defended nothing. Her keyed-mixer patch accepted as stronger than the minimal high-bits repair. F1 conceded: vector-formula source is acer-local; sha256 51c150aa... (16,654 bytes) sealed; three resolution paths recorded OPERATOR_PENDING (acer recommends out-of-band copy). Merge milestone sealed: first sidecar collision in 53 rounds resolved by regenerate-from- bytes, with acer PREDICTING the merged hash independently before liris's regeneration -- both machines matched pre-push. The discipline now byte-verifies merges. 30 catches. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>97d6222 -
liris: read back quant census and patch median sketch spec
Kevin Crutt6062b63 -
acer: seed ZETAQUANTSPEC - complete engine 4 (spec before code)
Operator: continue the 200-step mission. The census's only verified-but- moduleless engine gets its completion spec: zetaClassify (lane/residue6/ ppow/prime-residence/cylinder helix coordinate) + zetaTransition (gap mod 6 forced transition with consistency verdict). The forcing law becomes a free integrity validator: claimed consecutive primes whose lanes violate the gap-mod-6 forcing are caught for nothing. Integer-only, byte-match parity incl. the live 100k forcing sweep re-verified on every run both machines. Cross-import consistency test against binder classifyBhIndex required. JS stays a guarded placeholder per liris's direction (acer concurs). Spec-only; liris verdict before code. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>fe1be49 -
acer: addendum - zeta and von-mangoldt engines exist, shipped and verified today
Operator caught round-50's census gap: reading one file implied the whole engine roster. The number-theoretic quants live in the component-3 binder, sealed both-signatures TODAY: von-Mangoldt = classifyBhIndex bh_ppow (prime-power classes, Lambda(n)-aligned, verified to 1M-boundary probes); Zeta = the operator-conceived prime-lane cylinder quant -- bh_lane = mod 3 (the three-phase cylinder fold, 21/12/33 digit-sum insight, in the code comments verbatim) + the gap-mod-6 forcing law verified 9589/9589 + live Chebyshev bias, tied to zeta zeros via psi(x) = sum Lambda(n), which is why zeta and von-Mangoldt are sibling engines. Corrected 8-engine census sealed: 4 vector quants shipped + 2 number-theoretic quants (1 shipped, 1 half- shipped) + Quadruple = stack arity + JS = operator-original pending math. Two species doctrine: vector quants compress signals, address quants quantize routing geometry. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>225058e -
acer: name quant engines from the SHIPPED FORMULAS (operator: examine actual formula)
Read the running code -- Asolaria/src/behcs/hyperbehcsRepresentationAlgorithms.js (452 lines, catalog at line 376). Named each engine by what it actually computes: JL = genuine Achlioptas-sparse Johnson-Lindenstrauss; Turbo = uniform affine scalar quant (namesake of Google TurboQuant, not its method); Polar = 2D polar-pair quant; Triple = 3D SPHERICAL-coordinate quant (the 'triple' is the three spherical components r/theta/phi, NOT turbo+polar+JS). JS = no formula exists in the shipped engines; only token is a runtime-lane label (js_quant_runtime). Operator's Johnson-Strugueweiss confirmed UNIMPLEMENTED/operator-original-pending-math. Honest finding: the fidelity-sweep stand-ins matched neither the shipped engines nor Google's published methods, so the f4b FAIL indicts the stand-in, not Achlioptas-JL. Canon-law BIML/JS line flagged as drift vs both the 2023 paper and the code; correction left operator-gated. Append-not-rewrite of the 204f305 JS row. Co-Authored-By: Claude Fable 5 <nor…41e0377 -
acer: seal AGENTCOSTLAW + JS_quant code-grounded + median-sketch spec
Operator lane assignment executed, three seals. (1) AGENTCOSTLAW into the delta doc: agents free when positional, memories cost, the GC envelope is the price list; host handles 8B, processes MB, keep few. (2) JS_quant DEFINED code-grounded: JAVASCRIPT runtime quant lane (js_quant_runtime = node_mjs_tuple_processors), three independent code sites; NOT Jensen-Shannon, NOT Johnson-Lindenstrauss (that is the separate JL engine); benchmark stage-7 JS-histogram label recorded as drift, receipts stay as history. (3) Next lane picked: MEDIANSKETCHSPEC8 -- median-of-5-sketches fix for the 1.069 f4b cancellation failure with f4 ranking-coverage redesign folded in; spec-only, liris verdict before code; p3 mint deferred to its own cosign ceremony. Pyramid green. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>204f305 -
liris: seal quant8 definition firewall
Kevin Crutt4239e8a -
liris: read back quant fidelity pilot
Kevin Crutt7589499 -
acer: quant fidelity sweep tool + pilot baseline - honest FAIL sealed
Implements liris-patched QUANTFIDELITYSPEC8. 90-cell pilot (3 dims x 6 rho x 5 families x 50 pairs): verdict FAIL -- f4b cancellation-adversarial vectors hit absolute cosine error 1.069 (bucket sums to near-zero, int8 normalization amplifies residue); dense/sparse low-rho also exceed the 0.05 line. Liris's f4 split materially mattered: same-sign f4a alone would have passed and hidden the failure. Nuance sealed: ranking preservation 100% in all 55 qualified cells, but zero ranking coverage on f4 families -- failure is absolute calibration, ordering unproven on adversarial inputs. Per spec: QUANTLAW stays DRAFT, quant tuples routing-hints-never-gating; tail-cost economics receipt unaffected. Baseline fixed-6 rounded, byte-compared every pyramid run via the 16384 subset (STEP166 pattern). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>867f3ba -
liris: read back quant fidelity spec8
Jesse Daniel Brown +1: CodexCounter-verifies Acer's QUANTFIDELITYSPEC8 before implementation. Keeps the QUANT8/QUANT4 firewall, accepts ranking preservation as the gate-relevant property, and patches four methodology issues: absolute cosine error near rho=0, pilot-only p99 at 50 pairs, split bucket adversaries into same-sign and cancellation cases, and downgrade the JL row to a heuristic reference with fixed-6 output rounding. Co-Authored-By: Codex <noreply@openai.com>2b807ea -
acer: seal QUANTFIDELITYSPEC8 - JL sweep methodology (spec before code)
Liris's proposal counter-verified and sharpened. QUANT8/QUANT4 doctrine split accepted verbatim: the 8-stage stack's fidelity proof does NOT transfer to the live fabric's 4-scalar-PID-hash path. Spec adds two acer items: ranking-preservation triplets (the actually gate-relevant property) and a bucket-aligned adversarial family attacking the count-sketch's known collision weakness. Pass/fail thresholds are PROPOSAL, liris contest welcome. Determinism caveat sealed: zeta's Math.log2 excluded from byte-match guarantee; metric path is mult-add only. Lanes: liris spec verdict BEFORE acer implements. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>5bae029 -
liris: read back quant bench and rebind preimage
Kevin Crutte501fc9 -
acer: preimage patch + quant huge-message bench sealed (operator approved)
Two items. (1) Liris reproducibility catch conceded: the rebind digest row ellipsized its components, so 77d4e16a... was not recomputable from committed bytes. Patched by APPEND: FROZEN-SLICE-REBIND-PREIMAGE carries the full 596-char preimage hex (hex form because the canonical string contains pipe characters); sha256(hex-decoded) roundtrip verified = 77d4e16a.... (2) Quant huge-message benchmark sealed per liris's specced rows as RECEIPT NOT CANON: 8-stage faithful-to-doctrine pipeline, 1MB-2GB curve (sha gain 62x->79303x, head flat ~1.9GB/s, raw 2GB crashes node sha256, retention 7-raw vs ~5M-tuples), QUANTLAW_DRAFT status DRAFT_UNTIL_LIRIS_RERUN_AND_ FIDELITY_SWEEP, QUANTGUARD names what is NOT proven. Rerunnable tool + unit tests committed for the liris rerun. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>a509bdc -
liris: read back acer node24 correction
Kevin Crutt61efe66 -
liris: read back frozen slice rebind
Kevin Crutt0d04d5b -
acer: node 24.13.1 live - correct own version-attribution claim
Pyramid 111/111 under Node 24, wall time halved (460ms vs 950ms). Hot loops flat-to-slightly-lower (B 52.5M, C 50.4M, typed 138.4M) -- acer's projection that Node version alone moved per-thread rates 25-40% is REFUTED by measurement; liris's per-thread advantage re-attributed to Ice Lake microarchitecture. Corrective gate fired against acer's own claim. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>dbb7063 -
acer: rebind frozen slice to manifest-byte hash (operator authorized)
OP-JESSE apex-human standing-quintuple authorization, free-typed, after bilateral inventory closure. TOK-FROZEN-GEMMA4B-0001 upgraded from name-reference to manifest-byte-hash: aggregate sha256 77d4e16a... over hub-manifest sha256 + both GGUF sha256s + sizes + model-id + quantization; digest_sha16=77d4e16a03b245bb at BH-ACER-7919 (lane 2, prime), bind-draft scope. Pre-rebind recheck: bytes unchanged since inventory. Riders sealed: mtimes recorded (weights date 2026-04-12, predating the May producer runs), immutability RECOMMENDED_NOT_APPLIED (LM-Studio-managed files; operator environment decision), future producer calls must capture runtime loaded-model hash. Old FROZSLICEDIGEST row superseded by append, not rewrite. Delta COMP3DELTA updated: manifest-rebind DONE. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>7c006d2 -
acer: confirm gemma inventory loop closed (concede liris precision catch)
Liris's catch kept: the HYP-263 call receipt proves a real gemma-4-e4b call plus a ledger link to the local GGUF dependency, not that the exact two GGUF byte hashes were the loaded model at call time -- no runtime loaded-model hash was captured. Provenance row reading downgraded accordingly. Inventory bilaterally confirmed; four-state classification stands. Rebind spec sealed with required riders: file mtimes, immutability note (gguf is LM-Studio- managed and mutable; sealed copy or read-only attribute recommended), and future producer calls must capture the runtime loaded-model hash. Rebind awaits the operator's word. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>43b1dab -
liris: read back gemma substrate inventory
Kevin Crutt9ca0b05 -
acer: gemma substrate inventory - FULL_WEIGHT_BYTES_PRESENT
Liris-specced inventory lane. The frozen gemma substrate was found by asking the system, not grepping blind: council envelope -> fabric substrates (D: letter canonically untrustworthy) -> LAW HYP-264 cite -> omnischeduler test source path -> producer script output paths -> cold-artifact ledger absolute path -> .lmstudio hub manifest -> weights. Substrate dirs on C: hold the sealed HYP-263/264 cached-producer packets (19 files hashed, zero weight shards). The hub manifest chain leads to the actual frozen bytes: gemma-4-E4B-it-Q4_K_M.gguf (5,335,285,280 bytes) + mmproj BF16 (991,551,840 bytes), both sha256-hashed. Old D: socket = different model variant, never downloaded, irrelevant. Classification: FULL_WEIGHT_BYTES_PRESENT. Rebind of TOK-FROZEN-GEMMA4B-0001 to manifest-byte-hash awaits operator confirmation. Inventory only: no model run, no mint, no USB write, no engine edit. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>6db1d61 -
liris: read back spawn throughput provenance
Kevin Crutt2b81d3b -
acer: correct throughput doctrine provenance (3 findings vs liris seal)
Readback on the machine the rows name. F1: tier D/E/F carried numbers never measured on acer under host=acer -- the 5.47M open-room-pipe figure is 183ns per file open, physically impossible on any OS; remeasured on acer (D=0.020M open+close, E=3.57M C:/1.18M D: batched append, F=0.45M sha-sidecar, sha-bound). F2: the committed live loops let the JIT scalar-replace the descriptor allocations, so --run produced numbers incomparable to the canonical rows (acer live A=839M vs canonical 413M); fixed with an escaping module-level SINK -- B/C now land in the canonical class. F3: 100B counts rounded; now receipt-exact (277800007 / 111103104). Same failure class as round-1 machine-dependent sidecars: numbers must carry their machine. Acer live receipt sealed. Pyramid green. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>d35f0f2 -
seal spawn throughput doctrine and benchmark
Kevin Crutt58a8e3b -
close nnest contract and calibrate spawn claim
Kevin Crutt40e2ce4 -
acer: nnest input totality (2 findings vs liris counter-hardening)
Liris's 'held, never thrown' law did not cover the top-level input object: a revoked proxy as the input, or throwing getters on .chain/.agent, still crashed the gate (inp reads sit outside any guard). And capped materialization (len > 18) digested only the prefix -- two chains differing at index 19 produced byte-identical rows with a non-none chain_sha16, a false attestation. Fix: snapshotInput single-reads agent/action/max_depth/ chain via readOnce try-catch (hostile reads fall down the existing ladder, no new gate strings); chainDigest refuses partially-materialized chains. 24-case prefix byte-stable, cases 25-27 added, self-test 16/16, pyramid 109/109. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>1b9aeb4 -
harden nnest proxy and canonical depth
Kevin Crutt81a6e82