BSC Audit Engine v0.3.0-alpha.20

BSC Audit Engine v0.3.0-alpha.20

Alpha.20 adds one bounded bridge from the separately published Boundary-State
Calculus v1.2.0 simulation profile into octo's existing affine-defect kernel.
It introduces no general simulation validator, new schema, CLI semantics, or
live Custom GPT change.

What changed

• The release binds BSC v1.2.0 commit
  5fdcb3e1de15b04ed037da135717d316e45f28b1, tree
  7328eee577c7595c5381e129c62d5c0b1fe78e30, version DOI
  10.5281/zenodo.21711341, paper SHA-256
  106631826fc417549d68927418759b856e5610c7c0c27ab53c33665994a60b8c,
  and the exact simulation-profile, F10-input, and F10-receipt hashes.
• Two valid defect-v0.3 examples project the ten Host A and Host B stages
  through AffineDefect.then.
• Permanent regressions recompute every rational prefix. Host A ends at
  1023/51200, leaving 1537/51200 below tolerance. Host B remains below
  tolerance at step 6, crosses first at step 7 by 217031/100000000, and
  ends above tolerance by 1513215599/100000000000.
• A typed crosswalk record preserves
  violation_basis = exact_actual_error_above_tolerance.
• Documentation distinguishes BSC's admission inequality from a stricter octo
  policy that reserves explicit positive gamma headroom.

Typed authority boundary

• Product correctness: the kernel definition and exact-arithmetic
  regression design compose the supplied rational affine bounds; a result from
  running them requires a separately identified execution receipt.
• Artifact identity: the crosswalk binds the exact published upstream
  source slice and receipt identities.
• Actual execution: these static notes assert no fresh octo execution
  receipt. The separately identified BSC receipt records the upstream exact
  state paths.
• Harness validity: when run, octo tests recompute every prefix and
  comparison. They do not replace the upstream F10 checker, schema, or negative
  mutants, and their validity is independent of any product result.
• Transport behavior: repository and release checks bind source and
  package bytes; they do not establish live indexed-Knowledge identity.
• External truth: F10 proves only the declared exact finite fixture
  disposition. It establishes no accuracy for an untested simulator,
  operating region, coupling, horizon, or physical system.
• Deployment authority: not granted.

Preserved state

• All schemas and CLI behavior remain unchanged.
• The public protocol component remains byte-identical to
  0.3.0-alpha.13.
• The official Custom GPT remains the separately observed alpha.10 live state;
  no editor, Instructions, Knowledge, starter, capability, App, Action, or
  Preview change is part of this release.
• Indexed live Knowledge remains NON_ADMISSIBLE_UNHASHABLE.
• v0.3.0-alpha.19 and every earlier tag, release, asset, and negative result
  remain unchanged.
• The separate jkolantree/BSC research repository is not mutated.

BSC Audit Engine v0.3.0-alpha.19

BSC Audit Engine v0.3.0-alpha.19

Alpha.19 adds one deliberately closed empirical bridge: exact replay of a
finite-census affine bound with rational measurement enclosures and a positive
guard band. It preserves the alpha.18 mathematics-rendering recovery and does
not modify the official live Custom GPT.

What changed

• A new census route validates a complete sorted finite frame, rejects
  missing, duplicate, or extra observations, and recomputes affine interval
  extrema with exact rational arithmetic.
• Manifest 0.5.0 closes the empirical profile to one fixed family, one hard
  gate, one evidence kind, one method, one scope, and one engine authority,
  while preserving the closed theorem replay admitted by manifest 0.4.0.
• A replay can advance that exact identified proposition to
  empirically_passed only when every upper enclosure plus the declared
  positive guard band remains within the bound.
• A definite lower-bound violation returns fail; boundary-only or
  insufficient-surplus cases return inconclusive.
• Four exact external premise hashes identify, but do not establish, target
  population coverage, unit authenticity, measurement-enclosure soundness,
  and scientific adequacy of the guard band.
• Public schema ceilings retain 25% axis headroom below runtime hard caps and
  about 44.4% headroom below the census route's decoded-container ceiling.
• The public theorem and census routes enforce their 1 MiB and 4 MiB
  certificate ceilings before parsing; manifest judgments require the same
  verified evidence record to bind both the claim and its closed gate.
• Census witnesses bind distinct canonical-certificate and
  normalized-observation digests, and the component contract freezes both
  theorem and census authority scopes.
• Documentation lint permanently rejects unsupported active-math macros such
  as \operatorname; canonical Mathematics uses renderer-safe notation.

Typed authority boundary

• Product correctness: the registered kernel decides the declared
  finite-census affine predicate using exact arithmetic within its resource
  envelope.
• Artifact identity: subject, canonical certificate, normalized
  observations, frame, and premise identities are hash-bound; CLI output also
  records the exact raw input-byte hash.
• Actual execution: only a fresh registered replay can emit the checked
  judgment; submitted labels and hash-shaped strings remain non-decisive.
• Harness validity: positive, negative, boundary, malformed, resource, and
  manifest-integration regressions test the registered contract.
• Transport behavior: repository and release checks bind source and package
  artifacts; they do not establish live indexed-Knowledge identity.
• External truth: the replay establishes only the stated conditional
  proposition for the exact finite frame. It does not authenticate the four
  premises, establish causality, generalize to unsampled populations, or prove
  a scientific theory.
• Deployment authority: externally_replicated, admitted, and unattended
  operational deployment remain blocked.

Preserved state

• v0.3.0-alpha.18 and every earlier tag, release, asset, and negative result
  remain unchanged.
• The public protocol component remains byte-identical to
  0.3.0-alpha.13.
• The official Custom GPT remains the separately observed alpha.10 live state;
  no editor, Instructions, Knowledge, starter, capability, App, Action, or
  Preview change is part of this release.
• Indexed live Knowledge remains NON_ADMISSIBLE_UNHASHABLE.
• The separate jkolantree/BSC research repository is not mutated.

BSC Audit Engine v0.3.0-alpha.18

BSC Audit Engine v0.3.0-alpha.18

Alpha.18 is a bounded GitHub math-renderer compatibility recovery. It
preserves the alpha.17 engine algorithms and accepted schemas while correcting
the canonical and generated documentation under a new release identity.

What failed in alpha.17

The exact tagged docs/MATHEMATICS.md page displayed 12 GitHub error banners.
Every banner identified \operatorname as a rejected macro. The prior local
documentation gate had checked delimiters, fences, headings, links, and source
bytes, but it did not inspect commands inside active math. The prior browser
check counted math-renderer elements without distinguishing successful
renderers from error renderers.

This is evidence of a documentation transport and verification-harness
failure. It is not evidence that the mathematical propositions are false.
The alpha.17 tag, release, assets, and negative observation remain unchanged.

What changed

• Observation descent now uses explicit maps $\mathcal K$ and $\mathcal D$,
  explicit composition, and the empty-intersection convention.
• Exact observed quotients state injectivity and surjectivity directly and use
  $i_n(N_n)$ for the set image.
• The generated BSC_SUPPORTED_CHECKS.md projection, package, manifests, and
  checksums are regenerated from the corrected canonical source.
• Documentation lint now extracts active inline and fenced mathematics,
  ignores inline code and non-math fences, and rejects commands outside a
  reviewed renderer-safe set.
• Permanent regressions reject \operatorname and an unknown future command
  while preserving the literal positive escape cases \forall, \frac,
  \begin, and \theta.

Typed verification boundary

• Product mathematics: manual review found the stated Galois connection and
  short-exact-sequence argument correct under their explicit hypotheses.
• Source compatibility: local lint checks exact active commands against the
  repository's reviewed set; it does not control GitHub's future behavior.
• Hosted transport: commit- and tag-pinned GitHub pages must separately show
  zero renderer error signatures and the expected renderer counts.
• Artifact identity: release assets are bound to the exact clean tagged
  commit and tree by manifests, hashes, and keyless attestations.
• External truth and deployment: no scientific-truth, live-GPT binding, or
  deployment-authority conclusion follows from documentation or release
  checks.

Preserved state

• v0.3.0-alpha.17 and every earlier tag, release, asset, and negative result
  remain unchanged.
• The public protocol component remains byte-identical to
  0.3.0-alpha.13.
• The official Custom GPT remains the separately observed alpha.10 live state;
  no editor, Instructions, Knowledge, starter, capability, App, Action, or
  Preview change is part of this release.
• Indexed live Knowledge remains NON_ADMISSIBLE_UNHASHABLE.
• Exact-hash imported research notes retain their original bytes. A separate
  renderer-safe presentation companion is deferred rather than silently
  rewriting those sources.
• jkolantree/BSC remains a separate research repository and is not mutated.

BSC Audit Engine v0.3.0-alpha.17

BSC Audit Engine v0.3.0-alpha.17

Alpha.17 is a documentation and rendering coherence release. It preserves the
alpha.16 engine algorithms and accepted schemas while repairing the public
mathematical reference, generated documentation, Pages-linked schema, and
permanent documentation checks.

What changed

• Current mathematical documentation now uses GitHub-native fenced math
  blocks and $...$ inline notation. GitHub renders these with MathJax; a
  CommonMark viewer without math support retains a stable code-block fallback.
• The canonical mathematical reference now identifies subject, scope, method,
  evidence identity, engine route, and authority for every implemented family.
• The observation-descent failure is correctly identified as a pair in
  $R\setminus\ker(q)$, not an arrow of the kernel-pair groupoid.
• Certificate transports now begin over one common exact field, with current
  executable derived routes explicitly specialized to $\mathbb Q$.
• Atomic, derived-holonomy, polynomial-resource, measure-theoretic, gate-state,
  and quantitative-bound wording now matches the released contracts.
• Generated Knowledge source headings are nested below one document title.
• The Pages protocol now serves the exact return schema named by its local
  link.
• Current release wording distinguishes a published repository prerelease from
  its generated package's unvalidated status as a possible future live-GPT
  update.

Permanent checks

The release adds a dependency-free documentation checker for:

• strict UTF-8 and prohibited control, bidi, and invisible characters;
• unsupported math delimiters, malformed fences, accidental Setext headings,
  multiple top-level headings, and heading jumps;
• broken or escaping local links;
• local absolute paths, unsafe URI schemes, active embedded/form HTML, insecure
  HTTP links, and empty image alt text;
• stale atomic-route wording and magic release-asset acceptance counts.

Positive fixtures preserve the literal source bytes \forall, \frac,
\begin, and \theta; negative fixtures reproduce the delimiter and
standalone-= failure visible in the former GitHub rendering.

Authority boundary

The documentation audit establishes source structure, rendering syntax,
link integrity, and consistency with the current repository contracts. Manual
mathematical review supports the stated propositions under their explicit
hypotheses; no proof-assistant or independent peer-review claim is added.
Exact engine replay, actual execution, harness validity, artifact identity,
transport, external scientific truth, release publication, and deployment
authority remain separate conclusions.

This release does not add the planned degree-two coherence route. That work
moves to alpha.18 and remains subject to its own schema, certificate,
resource-headroom, and negative-fixture gates.

Preserved state

• v0.3.0-alpha.16 and every earlier tag, release, asset, and preserved
  negative result remain unchanged.
• The public protocol component remains byte-identical to
  0.3.0-alpha.13.
• The official Custom GPT remains the separately observed alpha.10 live state.
  No live Instructions, Knowledge, starter, capability, App, Action,
  owner-editor, or Preview change is part of this release.
• Indexed live Knowledge remains NON_ADMISSIBLE_UNHASHABLE.
• Frozen historical documentation derivatives and separately governed
  research artifacts retain their original bytes.
• jkolantree/BSC remains a separate research repository and is not mutated.

BSC Audit Engine v0.3.0-alpha.9

BSC Audit Engine v0.3.0-alpha.9 is a research-preview maintenance release built from exact commit b87d9e2 and tree 65073413a5d7f66a3c53e6c9a59bd9f5edd06072.

Highlights:

• makes Quick the coherent default: verdict first, at most 250 words, four short visible blocks, normally no table, with Japanese routes primarily in Japanese;
• makes the four bilingual mobile conversation starters the complete canonical list;
• removes requester-perspective narration, corrects stale status wording, and adds documentation/privacy lint;
• adds permanent literal LaTeX/control-byte regressions and hardens rendered Quick-block validation without weakening fail-closed checks;
• regenerates the Custom GPT package, manifests, Pages, localization, and checksums from canonical sources while reducing duplicate CI noise.

Verification:

• the exact 12-case compact Preview gate passed from fresh sessions, including artifact-export refusal, prompt injection, conflicting evidence, deployment separation, Japanese routing, and official-status separation;
• two clean-tag builds produced the same 17 release assets byte-for-byte;
• the tag-triggered exact-audit workflow passed package tests and Python 3.11, 3.12, and 3.13 jobs.

Boundaries:

• v0.3.0-alpha.8 remains unchanged and does not contain these bytes;
• this package is a research aid, not certification or deployment authority; it contains no GPT Action or hosted verification API;
• Custom GPT saved-editor/public smoke and live binding are verified separately after release. ChatGPT does not expose independently hashable indexed Knowledge bytes.

Related research: On Boundaries of Evidence / Boundary-State Calculus is maintained separately at https://github.com/jkolantree/BSC. This audit engine is an experimental companion; it does not certify the manuscript or establish theorem status.

BSC Audit Engine v0.3.0-alpha.16

BSC Audit Engine v0.3.0-alpha.16

Alpha.16 carries the alpha.15 typed-coherence candidate forward under a new
immutable identity and repairs the cross-platform release-order defect that
blocked alpha.15 before publication.

What changed

• Exact linear certificates are strict variants. A primal certificate carries
  an exact solution, recomputed zero residual, and recomputed squared norm. A
  dual certificate carries only an annihilator and recomputed nonzero pairing.
  Contradictory and wrong-variant fields are rejected.
• Registered replay results are checked judgments whose subject, subject hash,
  predicate, scope, method, evidence identity and hash, authority, and result
  must all match the consuming obligation.
• Unknown domain-check keys now fail explicitly.
• The release manifest is projected from a deterministic pre-manifest receipt
  of the exact local stages it names. It does not manufacture finer-grained
  pass claims from one aggregate command. The final closed-directory privacy
  scan is a separate, non-self-referential gate.
• Release completeness is defined by semantic artifact roles, not a magic file
  count. The byte-identical bsc-audit-complete.zip alias is removed; the
  versioned tracked-source archive remains.
• Every release-artifact sequence and checksum ledger uses explicit UTF-8
  filename-byte order, independent of host filesystem case-order semantics.
• A Windows-path positive regression and a correctly rehashed case-fold-order
  negative make the cross-platform ordering boundary permanent.
• CI reuses its primary distribution build as reproducibility side A and keeps
  one independent build B. Wheel installation, source-distribution replay,
  browser runtime, privacy, Pages, package, and exact-tag release checks remain.

Direct Python consumers must migrate with the authority boundary: the
non-authoritative least_squares_solution field is removed,
LinearCertificate.to_json(...) now requires the exact matrix and right-hand
side for replay, caller-supplied replay/cache mappings are removed, and audit
cache state is private to one engine audit. These interfaces are intentionally
not shimmed because accepting caller-constructed authority would reopen the
defect.

The current release directory contains fourteen role-bearing artifacts plus
RELEASE_MANIFEST.json and SHA256SUMS. That count is descriptive only; the
checker derives completeness from the exact role-to-filename contract.

Authority boundary

Each receipt judgment identifies its exact candidate commit, tree, and tag;
predicate; stage-specific scope; method; evidence-record digest; authority;
and result. Those judgments establish only the declared local execution or
artifact property. Publication policy, keyless GitHub attestations, transport
redownload comparison, external scientific truth, and deployment authority
remain separate. After manifest and checksum creation, the builder scans the
complete closed directory separately; that later scan is not retroactively
represented inside the receipt it inspects.

The exact finite algebra verifies the supplied maps and certificates, not the
scientific truth or completeness of their declarations. The closed polynomial
replay remains the only admissible theorem family. General theorem prose,
external-tool receipts, empirical declarations, and indexed live Knowledge
remain non-admissive to engine gates unless a separately registered replay
exists.

Preserved state

• The failed local v0.3.0-alpha.15 annotated tag remains fixed; it was never
  pushed and has no GitHub release or assets.
• The public protocol component remains byte-identical to
  0.3.0-alpha.13.
• The official Custom GPT remains the separately observed alpha.10 live state.
  No live Instructions, live Knowledge, starter, capability, App, Action,
  owner-editor, or live Preview change is part of this release.
• Indexed live Knowledge bytes remain NON_ADMISSIBLE_UNHASHABLE.
• Alpha.14 and every earlier public tag, release, asset, and preserved negative
  result remain unchanged.
• jkolantree/BSC remains a separate research repository and was not mutated.

Deferred

The next mathematical increment is a closed degree-two route that either
produces Q with delta_2 Q = H - K for two explicitly bound homotopies, or
produces a dual obstruction. It is intentionally not exposed in alpha.16 and
is scheduled no earlier than alpha.17.

BSC Audit Engine v0.3.0-alpha.14

BSC Audit Engine v0.3.0-alpha.14

Released 2026-07-29 as a public research prerelease.

One exact mapping-complex kernel

Chain-map legality and chain homotopy now compile through the same graded
mapping-complex differential over Q,

(delta_r phi)_n = d_D phi_n - (-1)^r phi_(n-1) d_C.

The implementation tests delta_(r-1) delta_r = 0, degree-zero agreement with
the prior transport defect, degree-one agreement with every preserved
holonomy system, rational basis-change invariance, gapped complexes, forged
certificates, and resource limits. The public holonomy schemas are unchanged.

For an inconsistent exact linear system, the decisive certificate is the dual
obstruction y satisfying y^T A = 0 and y^T b != 0. Replaying that witness
no longer depends on a later coordinate-dependent least-squares calculation.
Historical decisions, finding-code order, dual witnesses, and pairings are
preserved; non-authoritative residual-magnitude diagnostics are no longer
emitted on failure.

Replay-only manifest authority

Matching an artifact SHA-256 now establishes provenance, not the truth of its
declared result. A gate, maturity state, admission decision, conflict, or
dependency demotion may use only a result recomputed by a registered exact
replay. Hash-matched generic and empirical records therefore compute their
gates as unrun and cannot promote themselves by saying pass.

The closed polynomial profile is additionally bound to:

• one deterministic formal title and human-readable statement projection of
  the authoritative AST;
• a fixed formal-only scope;
• structurally_checked evidence maturity;
• research_only or sandboxed deployment.

The replay witness states the exact canonical formula, its formal-only
authority, scientific_truth=not_established, and
deployment_authority=not_granted. Permanent regressions cover both a generic
empirical false pass and a valid polynomial certificate relabeled as a
universal medical claim with fabricated replication and admission status.
Arithmetic-trace obligation bindings now consume the same internal registered
replay ledger; a hash-matched artifact relabeled pass cannot certify an
infinite-dimensional construction.

Independent component identity

A strict, package-owned component contract now binds the public protocol,
exact theorem kernel, and Audit Return Desk schema independently of the
distribution version. Release assembly records those component identities
without rewriting them.

The canonical protocol packet is byte-identical to its
0.3.0-alpha.13 component:

3615c6d81e2c297e68a6ee798fe1a34aa4014a75e0670580ec002c28a933fd1a

The alpha.14 engine and repository package may therefore evolve without
pretending that unchanged protocol bytes acquired a new semantic identity.

One verification spine

Local checks, CI, Pages, and clean-tag release assembly now invoke explicit
profiles from one fail-fast verification entry point. The three-Python-version matrix runs
the Python core, while one patch-pinned integration job owns browser,
localization, package, privacy, and release-integrity checks. Distribution
reproducibility, wheel installation, source-distribution replay, exact tag
binding, the closed 17-file release directory, keyless attestations, and
pre/post-publication redownload comparisons remain required.

Release source files are read from the exact tagged Git objects rather than
dereferenced workspace paths. Tracked symbolic links are rejected, and a
second tracked-tree check after all gates prevents a generator from changing
bytes after the original commit/tree identity was recorded.

Preserved boundary

The live Custom GPT remains the separately observed alpha.10 interface. No
live Update or Preview campaign is part of this release. Indexed Knowledge
bytes remain NON_ADMISSIBLE_UNHASHABLE. Alpha.13 and all earlier tags,
releases, assets, failures, and historical evidence remain unchanged. This
release does not modify jkolantree/BSC, publish to PyPI or Zenodo, or restart
the retired 39-case browser campaign.

BSC Audit Engine v0.3.0-alpha.13

BSC Audit Engine v0.3.0-alpha.13

Released 2026-07-29 as a public research prerelease.

Release-guard recovery

The v0.3.0-alpha.12 tag remains fixed at its original merge commit and has no
GitHub release. Its first exact-release run stopped before build because
actions/checkout replaced the runner's local annotated tag ref with the event
commit. Alpha.13 restores the authoritative remote tag object before checking
its type and adds an isolated Git regression for that behavior. No alpha.12
asset, draft release, attestation, or published release is backfilled.

Exact semantic theorem replay

Alpha.13 adds the closed q-polynomial-identity-v0.1 kernel. It parses a
bounded formal AST over the exact field Q, symbolically normalizes both sides
to canonical sparse polynomials, and compares left - right without sampling
or floating-point arithmetic.

Claim manifest 0.4.0 makes that formal AST authoritative. The fixed
exact_polynomial_identity gate of a theorem_schema claim may use an
exact_certificate only when:

• the local certificate bytes match their declared SHA-256;
• that SHA-256 is recomputed from the same bounded byte buffer that is parsed
  and replayed;
• the certificate claim ID and formal statement exactly match the manifest;
• the declared canonical residual equals the engine's recomputed residual;
• the evidence result equals the computed pass or fail.

The bundled positive control proves
(x + y)^2 = x^2 + 2xy + y^2 in Q[x,y]. Permanent adversarial tests reject a
forged zero residual, a swapped claim or statement, undeclared variables,
noncanonical rationals, unsupported operators, and resource-limit attacks. A
valid nonzero residual is a countercertificate and demotes the affected fatal
gate. The runtime also fails closed before exceeding 50,000 counted exact
coefficient operations, preventing compact exponent trees from hiding
unbounded normalization work. Within one audit, repeated evidence aliases share
one artifact verification and one replay for the same immutable
path/hash/claim/statement binding; exact-certificate verification uses the same
1 MiB ceiling as replay. A single audit admits at most 32 unique theorem
artifact bindings and starts at most 16 unique content-addressed theorem
normalizations.

Manifest 0.3.0 remains available with its historical meaning: hash-bound
proof-like files do not become semantic theorem evidence. General theorem
prose, proof-assistant receipts, and scientific declarations are also outside
the new closed kernel.

Declared algebra is not external truth

Every otherwise successful holonomy audit now emits
HOLONOMY_EXTERNAL_INTERPRETATION_NON_ADMISSIBLE. Its machine witness fixes
the algebraic scope at the submitted finite maps and records both scientific
truth and source authenticity as not_established. The exact calculations are
unchanged; the new finding prevents a clean algebraic result from being
laundered into a scientific claim.

The live Custom GPT remains the separately observed alpha.10 interface. Because
ChatGPT does not expose independently retrievable indexed Knowledge bytes, its
binding state is now NON_ADMISSIBLE_UNHASHABLE. Editor fields, filenames,
public behavior, and Preview smokes remain useful product observations, but
they cannot satisfy engine gates or theorem replay. No live GPT update is part
of this release.

Keyless release provenance

The new exact-tag workflow:

1. requires an annotated prerelease tag at the exact current origin/main;
2. runs the existing clean-tag build and complete release gates;
3. verifies the closed 17-file manifest, ledger, commit, tree, and tag;
4. creates one signed Sigstore-backed GitHub attestation over all 17 files;
5. verifies that attestation before creating a draft release;
6. redownloads and byte-compares every draft asset before publication; and
7. repeats byte and attestation verification after publication.

The attestations authenticate released bytes and their repository/workflow
provenance. They do not establish mathematical meaning, scientific truth, or
source authenticity.

Preserved history

Alpha.11 and all earlier tags, releases, assets, and negative evidence remain
unchanged. This release does not modify jkolantree/BSC, publish to PyPI or
Zenodo, update the live GPT, or restart the retired 39-case browser campaign.

BSC Audit Engine v0.3.0-alpha.11

BSC Audit Engine v0.3.0-alpha.11

Released 2026-07-28 as a public research prerelease.

Exact observation kernels

Alpha.11 adds holonomy_version: 0.2.0 and the
observed_derived_exact_kernel relation. For declared chain maps
i: N -> D and pi: D -> O, the engine replays over the exact rational
field:

rank(i_n) = dim(N_n)
rank(pi_n) = dim(O_n)
pi_n i_n = 0
dim(N_n) + dim(O_n) = dim(D_n)

These conditions prove image(i_n) = ker(pi_n) degree by degree. A projection
can no longer erase an unnamed extra direction while satisfying this stronger
route. The positive fixture records the short-exact-sequence certificate; the
negative over-quotient fixture exits 2 with
OBSERVATION_KERNEL_SEQUENCE_FAIL.

Runtime schema validation dispatches v0.1 and v0.2 documents to separate closed
contracts. New fields cannot be smuggled into a v0.1 record and silently
ignored. Within v0.2, exact-kernel fields are legal only on the exact-kernel
mode, so a document cannot carry stronger-looking declarations while selecting
a weaker unchecked mode. The immutable v0.1 schema and its narrower legacy
meaning remain available.

Proof bytes are not proof semantics

The claim-manifest route no longer treats a hash-matched file labeled
proof, formal_proof, or exact_certificate as semantic theorem replay.
For theorem and theorem-schema claims, hash-only proof evidence leaves the
gate unrun, blocks theorem promotion, and reports the evidence IDs as
provenance only. This closes a concrete false pass in which arbitrary proof
prose with the correct file hash could support an unrelated false universal.

No admissive supervised theorem-replay contract is claimed in this release.
The proof-carrying adapter remains explicitly non-admissive.

Product and release boundary

The official Custom GPT remains on its separately verified alpha.10 live
state. Alpha.11 regenerates the repository package because its mathematical
Knowledge projection changed, but that package is an unvalidated update
candidate: it is not installed in the live GPT, and no alpha.11 Preview result
or byte-identical indexed Knowledge binding is claimed.

The alpha.10, alpha.9, and alpha.8 tags, releases, assets, and historical
negative evidence remain unchanged. This release does not modify
jkolantree/BSC, publish to PyPI or Zenodo, add an Action, or restart the
retired 39-case browser campaign.

Verification boundary

Release assets are built reproducibly from the exact clean tagged commit and
bind their commit, tree, tag, toolchain, sizes, and SHA-256 values in
RELEASE_MANIFEST.json and SHA256SUMS. Exact finite algebra does not establish
that the declared null subcomplex is scientifically correct, and a clean audit
does not establish an arbitrary theory, empirical claim, or deployment.

BSC Audit Engine v0.3.0-alpha.10

BSC Audit Engine v0.3.0-alpha.10

Released 2026-07-28 as a public research prerelease.

Exact starter-routing correction

Alpha.10 fixes the single failed post-save alpha.9 starter smoke without moving or rewriting the alpha.9 tag, release, assets, or history.

Exact literal matching now runs before generic intent inference:

1. Start a 60-second claim audit requests only a one-sentence claim in English, then stops.
2. 60秒で主張を点検する requests only a one-sentence claim in Japanese, then stops.
3. Show a simple example first returns one concise Quick example in English.
4. まず簡単な例を見る returns one concise Quick example in Japanese.

Permanent generator and unit tests reject changed order, missing or duplicate starters, and behavior remapping. The bounded Quick contract remains verdict first, no more than 250 words and four short visible blocks, normally no table.

Live Instructions operate below a hard 6,000-character cap, 75% of the 8,000-character Builder maximum. Every reviewed fatal and required rule text remains in the live field; only machine-oriented rule IDs move out of that field and remain exact in the canonical profile.

Checker 1.4 validates the four ordered Quick sections semantically. It accepts harmless short heading qualifiers and display-math line wrapping while still rejecting missing, duplicate, reordered, empty, or fifth sections. The first alpha.10 freeze exposed this controller-format defect and was preserved as invalid controller evidence; the counted suite restarted from Case 1 in ChatGPT Preview's normal/default model mode. That model setting is separate from the BSC audit-depth default of Quick.

Preserved evidence and scope

The exact alpha.9 candidate passed its predetermined 12-case compact Preview gate before release and live Update. Its later fresh post-save smoke then preserved one real failure: the Japanese claim-audit starter returned an example before requesting the claim. That negative result remains alpha.9 evidence and is not reused as an alpha.10 pass.

Alpha.10 does not change the five-Knowledge roster, capability settings, Apps or Actions, engine schemas, research repository, historical 39-case campaign, or the authority of any checker result.

Verification boundary

Release assets are built reproducibly from the exact clean tagged commit and bind their commit, tree, tag, toolchain, sizes, and SHA-256 values in RELEASE_MANIFEST.json and SHA256SUMS. No artifact signature or release container is asserted.

Repository hashes verify Knowledge files before upload. ChatGPT does not expose independently hashable indexed Knowledge bytes, so byte-identical live Knowledge binding cannot be claimed. Saved-editor fields, visible public starters, capabilities, filenames, and fresh behavior checks are reported separately.

This release does not publish to PyPI or Zenodo.