PaperlayerView exact revision on arXiv

Source-linked AI summary

Making AI-Assisted Claims Independently Challengeable: Publication Authority and a Protocol for Falsifiable Publication Records

Torsten Olivi Tiltack, Yifei Dong, Kun Yu, Xu Wang, Wei Liu, Jianlong Zhou, Ren Ping Liu, Fang Chen

arXiv:2609.17631v1cs.AIcs.CRcs.CY

TL;DR

AI-assisted claims can appear more authoritative when evidence, analysis, authorization, presentation, and correction history refer to different states. The paper defines Publication Authority as a complete, exact-state publication capability and evaluates PAC-2026 SF-4; its finite models show bounded safety outcomes, not general refinement or field efficacy.

  • Problem

    Evidence, analysis, authorization, and reader-facing meaning can refer to different states, while provenance and transparency alone do not specify the publication transition governing them.

    Source: §1 · p. 1

  • Method

    The paper makes publication permission depend on fresh, exact-state passing obligations, represented by a single-use permit for one publication act.

    Source: §7 · p. 21

  • Results

    110,764 distinct reachable safe states were explored across ten configured finite models, while 76 unsafe configurations produced expected violations or declared observer blind spots.

    Source: §7 · p. 21

  • Takeaways & Limitations

    Each mandatory obligation must earn its own fresh passing result, and a materially different candidate must earn its own authority.

    Source: §7 · p. 21

    Formal evidence is confined to one factorized profile and does not prove general refinement, checker correctness, or unbounded composition.

    Source: §6.4 · p. 21

Abstract

from arXiv · show

AI-assisted claims can appear authoritative when evidence, analysis, human authorization, presentation, and correction history refer to different states. Provenance, attestation, and transparency expose history but alone do not specify the publication transition examined here. We develop Publication Authority as an exact-state, non-transferable, single-use publication capability and instantiate it in PAC-2026 (Publication-Accountability Calculus), a machine-readable AIJIM Protocol candidate. We evaluate its fourth bounded semantic freeze (SF-4), a fixed-profile specification designed for replaceable bindings. Six obligations govern evidence, runs and artifacts, measurement disclosure, authorization, surface correspondence, and lifecycle continuity. Each yields a target-bound witness, localized counterexample, or localized unverifiability; none can compensate for another. Only a fresh, complete all-pass record derives the permit consumed by one atomic publication transition. We use identity vectors, adversarial cases, finite models, and historical implementations. Ten models explored 110,764 safe reachable states; 76 unsafe configurations produced the expected violation or observer countermodel. A reader surface passing its correspondence check cannot authorize publication unless the accepted record admits that surface. SF-4 separates evidence horizon from verification time and rejects an authentic but causally invalid authorization. A historical predecessor path reproduced 17 frozen authorization-successor outcomes. A later in-house, instance-blind test of known case classes matched all 183 scored expectations; same-host package execution reproduced its 240 archived observations. Results support internal coherence, bounded safety, fault sensitivity, and limited constructibility, but not factual truth, general refinement, blind interoperability, field efficacy, or standards status.

1. Introduction

AI-assisted publication can make a claim appear more strongly supported or authorized when evidence, analysis, approval, and reader-facing meaning refer to different states. The paper frames Publication Authority as a way to govern that exact-state transition and evaluates a fixed-profile protocol candidate.

1. Introduction

  • Evidence, analysis, approval, and reader-facing meaning can refer to different versions, making a claim appear more authoritative than their joint record warrants.The paper calls this failure silent epistemic promotion; “stronger” describes recorded status, not a numerical truth scale or factual truth.Source: §1 · p. 1
  • Existing provenance, attestation, and transparency mechanisms record or transport facts, but the residual question is how they govern one exact-target publication transition.The proposed contribution is their publication-specific coupling, not novelty of the component mechanisms in isolation.Source: §1 · pp. 1–2
  • Publication Authority is a protocol-derived, non-transferable capability for one atomic publication transition involving one exact candidate and context.It asks whether this version, rather than an earlier or materially different one, may be published under the claimed status now.Source: §1 · p. 2
  • PAC-2026 SF-4 operationalizes the design through six mandatory coordinates, each yielding a target-bound witness, counterexample, or unverifiability, with failures unable to compensate for one another.The six coordinates cover evidence, runs and artifacts, measurement disclosure, exact human authorization, reader-surface correspondence, and lifecycle continuity.Source: §1 · p. 2
  • The evaluation combines canonical vectors, adversarial cases, finite-state models, and historically separated implementations, while keeping the article distinct from the normative corpus.Its evidence layers have explicit claim ceilings, and later implementation results cannot upgrade the frozen SF-4 result.Source: §1 · p. 2

2. Literature and Standards Context

The literature supplies mechanisms for evidence, provenance, attestation, authorization, and lifecycle records, but does not by itself specify their semantics at a complete reader-facing publication transition. PAC-2026’s residual contribution is their coupling, with the scope and novelty claims bounded by unresolved prevalence and comparison questions.

2. Literature and Standards Context

  • Existing standards and systems provide provenance, research-object, attestation, authorization, and lifecycle mechanisms; the comparison asks how they couple at publication.The section distinguishes the mechanism layer from publication-specific semantics for a candidate acquiring recorded status.Source: §2 · p. 3

2.1. Evidence, computation, and accountable claims

  • Claim-level evidence gaps and computational reproducibility concerns support explicit evidence and run closure, while accountability research motivates inspectable rules and later verification.These sources do not establish prevalence of the complete multi-state publishing failure or validate PAC-2026.Source: §2.1 · p. 3

2.2. Strong adjacent standards and systems

  • Provenance, attestations, transparency, and lifecycle systems can represent or transport record components, but SCITT leaves application meaning to profiles and relying systems.PAC-2026 neither replaces SCITT’s transparency service nor requires it as a substrate.Source: §2.2 · p. 3
  • Scholarly-agent protocols and recent authorization proposals cover research-object verification, evidence-gated claims, versioned approval, or exact-action controls, but govern different units than the complete reader-facing publication transition.LEDGER is described as an upstream complement for inspectable evidence and run records, not publication-admission semantics.Source: §2.2 · pp. 3–4

2.3. Publication-transition coordination

  • The residual design question is whether a composition explicitly couples exact-candidate admission, separate outcomes, non-compensation, revision-exact authorization, surface correspondence, and lifecycle effects.No item in that list is claimed novel alone; the candidate contribution is their coordination at one transition.Source: §2.3 · p. 4
  • The six-coordinate coupling is not claimed universally minimal or empirically prevalent, and global priority, search completeness, and the absence of a stronger composition remain open.Stale authorization and reader-surface drift remain analytically grounded hypotheses whose prevalence and field effect are unmeasured.Source: §2.3 · p. 4

3. Design Science Method and Requirements

The section retrospectively audits artifact-first requirements and develops Publication Authority in PAC-2026 as a fixed-profile protocol for one exact publication transition, requiring fresh, complete, exact-target verification and authorization, separate obligation outcomes, and a single-use permit. Its falsifiable contracts define a bounded conformance scope that does not establish factual truth, lifecycle completeness, or general refinement.

3.1. Research design and unit of analysis

  • PAC-2026 was designed before its literature matrix, which retrospectively traces requirements to external sources, failure classes, obligations, criteria, and falsifiers while leaving retrofit risk.The design unit is one fixed-profile publication-accountability transition, not a deployed product; implementation results and finite models cannot establish protocol theorems or field utility.Source: §3.1 · pp. 4–5

3.2. Failure classes and requirement admission

  • Five failure classes motivate eight requirements, admitted when omission leaves a declared failure possible and the requirement has an observable obligation and direct falsifier.The requirements are not claimed universally minimal, jointly sufficient, or novel; the mapping to five laws, obligations, propositions, and evaluation criteria is not one-to-one.Source: §3.2 · p. 5

3.3. Build–evaluate learning cycles

  • Counterexamples drove the build–evaluate cycles: SF-2 could not express a stale predecessor result, SF-3 separated target assessment, and SF-4 separated evidence horizon from verification time.These cycles are design-learning evidence, not a preregistered experiment; SF-4 also preserved explicit causal ordering.Source: §3.3 · pp. 5–6

3.4. Evaluation strategy and strongest rival

  • The primarily artificial evaluation tests semantic coherence, bounded safety, targeted fault sensitivity, and constructibility, but does not evaluate utility or establish novelty.The historical novelty mapping is conditional and cannot show that a coherent composition cannot substitute; the separately preregistered successor challenge remains unexecuted.Source: §3.4 · p. 6

4. Publication Authority Design Knowledge and PAC-2026

  • Publication Authority is a derived capability for one exact publication act; PAC-2026 SF-4 operationalizes it as a fixed-profile specification whose governed artifacts control executable semantics.The antecedent RM-2025 conceptual architecture represented related components but did not supply PAC’s executable publication-transition semantics.Source: §3.3 · p. 5, §4 · p. 6

4.1. Five laws for an accountable publication transition

  • Five laws require typed accountability closure, fresh exact-target authorization, non-amplification across evidence coordinates, registered-surface correspondence, and continuing challengeability.Surface correspondence covers registered machine-readable observables, not unrestricted language equivalence or reader comprehension; lifecycle status must expose unverifiability when it cannot be reconstructed.Source: §4.1 · pp. 6–7

4.2. Fixed-profile instantiation

  • SF-4 keeps six non-compensating judgments distinct and requires a fresh, complete evaluation before an atomic act consumes a single-use permit and records publication.The evidence horizon fixes the candidate’s evidence cutoff, while verification time governs freshness and current-status checks. R1 does not guarantee replay, and frozen SF-4’s measured branch retains an unresolved HC-0024 binding requirement; no incorrectly issued permit was recorded.Source: §4.2 · pp. 8–9

4.3. Separate outcomes and no silent promotion

  • Each obligation returns PASS with a target-bound witness, FAIL with a localized counterexample, or CANNOT_VERIFY; only a valid, fresh, complete all-pass record can support publication.A separate gate blocks promotion when execution, identity, or freshness binding is inadequate; a negative VerificationRecord remains portable diagnostic evidence, not a publishable record.Source: §4.3 · p. 9, §4.2 · p. 9, Figure 1 · p. 10
  • Each obligation’s outcome and promotion authority remain separate: verified NOT_MEASURED may pass disclosure checks without establishing measurement, truth, or scientific sufficiency, while CANNOT_VERIFY blocks authority derivation.No coordinate can promote another or contribute to a global score.Source: §4.3 · p. 11, Figure 2 · p. 11, Table 3 · p. 11

4.4. Deriving and consuming Publication Authority

  • Only a complete, fresh record with exact active authorization and every Core obligation passing substantively and at the gate can derive a permit bound to the candidate and publication context.AUTHORIZED records A1 pass only; it is neither a permit nor an alternative path to Publication Authority.Source: §4.4 · p. 12
  • The permit is single-use: one atomic publication consumes it and creates one immutable record, authenticated PUBLISH event, and initial lifecycle prefix.Verification and publication must satisfy the profile’s time and validity checks; later status is reconstructed from the supplied authenticated prefix, which may omit later events.Source: §4.4 · p. 12

4.5. Falsifiable propositions and conformance boundary

  • Four falsifiable contracts cover exact permit use, coordinate-local promotion, stale-authorization non-inheritance, and integrity of current status within the admitted lifecycle view.A trace violating any stated condition falsifies its corresponding proposition; the status contract does not establish discovery of events absent from that view.Source: §4.5 · p. 13
  • These protocol contracts are assessed through schemas, validators, adversarial vectors, and finite models, but do not establish checker soundness, factual truth, or universal correctness.Clock authenticity, common-mode input faults, and discovery by a non-executing observer remain outside the checked boundary.Source: §4.5 · p. 13
  • Substrate neutrality remains a design objective: technical bindings must preserve protected protocol meanings, and SF-4 does not establish production binding or general refinement.Conformance is attached to a versioned role and its required inputs, outputs, and fail-closed behavior, not to a brand or possession of one valid record.Source: §4.5 · pp. 13–14

5. Evaluation and Results

The evaluation assesses SF-4 against four criteria using executable artifacts, finite models, adversarial cases, and historical implementation paths, with explicit limits on what each evidence layer supports.

5.2. Criterion 1: semantic coherence

  • SF-4 rejected a pre-decision authentication proof as FAIL/A_TEMPORAL_ORDER_INVALID; all 23 frozen temporal cases matched declared tuples, with 17 SF-3 cases retained as passing regressions.The correction is relative to the protocol’s predecessor, not an external system or a general temporal theorem.Source: §5.2 · p. 15

5.3. Criterion 2: bounded safety

  • 110,764 safe reachable states were explored across ten models, while 76 deliberately unsafe configurations produced expected violations or an observer countermodel.The 76 unsafe configurations are expected witnesses, not real-world failures; the models are bounded to their named configurations.Source: Table 5 · p. 16, §5.3 · p. 16

5.4. Criterion 3: targeted fault sensitivity

  • Fault tests localized nonpasses and rejected locally valid but unbound substitutions; declared NOT_MEASURED or CANNOT_VERIFY states required witnesses, while unverifiable M1 blocks authority.The controls establish witness requirements for declared absence, not measurement validity, and do not resolve HC-0024’s separate evidence-binding issue.Source: §5.4 · pp. 16–17

5.5. Criterion 4: bounded constructibility

  • Historical constructibility included a 373-test SF-3 path, RM-2025 mutation and observer cases, and a separate consumer’s expected outcomes, but not independent-customer evaluation.The SF-3 path reproduced all 17 frozen successor observations; the separate consumer was not an independently developed AIJIM implementation.Source: §5.5 · p. 17

5.6. Synthesis and post-evaluation status

  • A later SF-5 binding study matched all 24 expected outcomes, with 22 negative cells rejected before permit derivation, but supports only its specified bindings and corpus.It neither retroactively implements SF-4 nor establishes deployment, production readiness, or independent reproduction.Source: §5.6 · p. 17
  • The 92 expected declarations across 161 role invocations were a disclosed-case calibration, not blind testing; shared AI-assisted integration’s contribution to agreement was unquantified.Later source handoffs recover pre-integration identities, not a complete account of development responsibilities or information exchange.Source: §5.6 · pp. 17–18
  • All 183 scored G7-B instances matched expectations across four pairings: 75 expected acceptances and 108 typed expected rejections.Acceptance is the analyzer’s operational category, not publication permission; 15 valid-verification-record cases retain localized FAIL or CANNOT_VERIFY.Source: §5.6 · p. 18
  • Same-host package execution reproduced all 240 archived G7-B observations, supporting repeatability of the completed study rather than an additional blind experiment.The package’s evidence layers remain distinct and do not establish complete conformance, independently evaluated interoperability, or field effects.Source: §5.6 · p. 18

6. Discussion

The discussion frames Publication Authority as a checkable consequence of a complete, exact publication state and assesses the evidence for PAC-2026 SF-4 within bounded criteria. It distinguishes the protocol’s diagnostic scope from claims about truth, general refinement, independent evaluation, or field utility.

6.1. Answers to the research question and subquestions

  • Publication Authority derives from a fresh, complete record for the exact candidate, with separate outcomes for each obligation and no compensation across failures.A material non-equivalent successor must be re-admitted; the design does not establish whether the underlying claim is true.Source: §6.1 · p. 19
  • SF-4’s evidence supports semantic coherence, bounded safety, and targeted fault sensitivity, while constructibility is supported indirectly rather than by an SF-4 production binding.Ten finite models explored 110,764 safe reachable states and 76 deliberately unsafe configurations; later SF-5 and G7-B evidence applies to their own evaluated roles and paths.Source: §6.1 · p. 19

6.2. Contribution hierarchy and relation to standards

  • PAC-2026 SF-4 contributes a fixed-profile instantiation, while its added design knowledge concerns dependencies among checks rather than the checks’ individual existence.Its joint obligations distinguish authentic approval for an earlier version, portable negative verification, and registered surfaces that mismatch the accepted record.Source: §6.2 · p. 19
  • Comparative superiority and general refinement remain unestablished, and replaceable binding is a design objective rather than demonstrated substrate independence.The later two-binding results provide bounded refinement evidence for those paths, not a production binding for SF-4 or a general refinement result.Source: §6.2 · p. 19

6.3. Implications for research and practice

  • The protocol’s proposed practical value is diagnostic: it can report which declared obligations passed, failed, or could not be verified, but does not warrant truth or appropriate reliance.Human review records responsibility, not truth; a commitment to bytes alone proves neither their existence at the asserted time nor their authenticity or public inspectability.Source: §6.3 · p. 20

6.4. Limitations and threats to validity

  • The six-coordinate design is not field-validated, may omit or over-separate relevant coordinates, and has not been assessed for naturalistic effects or broad independent conformance.Formal evidence is confined to one factorized profile, while durable, distributed, concurrent, and crash-consistent permit consumption remains untested.Source: §6.4 · p. 21

7. Conclusion

The conclusion presents Publication Authority as a checkable consequence of the exact state being published and PAC-2026 SF-4 as an executable fixed-profile instantiation. Its evaluation supplies bounded evidence for the specified cases, while independent conformance and field utility remain open.

7. Conclusion

  • Publication Authority makes permission to publish a checkable consequence of the complete, exact state, and a materially different candidate must earn its own authority.Each mandatory obligation requires a fresh passing result and promotion witness; an at-most-once permit represents authority for one publication act.Source: §7 · p. 21
  • PAC-2026 SF-4 makes six obligations executable within one admitted Core Profile, separating individual object validity from the conditions for joint publication authorization.The obligations cover evidence, analytical work, measurement disclosure, human authorization, registered presentation, and lifecycle continuity.Source: §7 · p. 21
  • Ten configured finite models preserved invariants across 110,764 reachable safe states, while 76 unsafe configurations produced expected violations or an observer blind spot.The validator also rejects individually valid replacement surfaces and lifecycle plans outside the evaluated record and rejects an authorization construction accepted by the SF-3 predecessor.Source: §7 · p. 21
  • The paper presents a testable account of publication-authority relationships, but independent conformance and field utility require further evidence.Shared integration, known case classes, and bridge-derived classifications limit the independence inference from G7-B; local reproduction establishes repeatability, not external evaluation.Source: §7 · p. 21

CRediT authorship contribution statement

  • The contribution statement assigns conceptualization and supervision across several authors, with the lead author also credited for methodology, software, validation, formal analysis, and investigation.The listed roles also include writing, visualization, project administration, and data curation.Source: CRediT authorship contribution statement · p. 22

Funding; Declaration of competing interest

Declaration of generative AI and AI-assisted technologies in the manuscript preparation process

  • The authors report using Codex and Claude Code under human supervision for literature discovery, manuscript drafting and revision, and research-code development.The lead author reviewed and revised AI-assisted material, and the authors take responsibility for the article.Source: Declaration of generative AI and… · p. 22

Data and artifact availability

  • The versioned public artifact collection includes the SF-4 freeze, historical SF-3 snapshot, later binding-refinement, calibration and G7-B materials, and a combined reproduction record.Restricted RM-2025/SF-P0 material and historical separate-consumer deposits were not included or rerun.Source: Data and artifact availability · p. 22

Supplementary Material: Making AI-Assisted Claims Independently Challengeable: Publication Authority and a Protocol for Falsifiable Publication Records

The supplement retains audit detail while distinguishing its material from the normative protocol source. Its authors are affiliated with the University of Technology Sydney.

Supplementary Material

  • The listed authors are affiliated with the University of Technology Sydney across its Data Science Institute, Global Big Data Technologies Centre, and School of Computer Science.Source: Supplementary Material: Making… · p. 25
  • The supplement preserves crosswalks, bounded counterexamples, model inventories, historical predecessor cases, and artifact identifiers for audit.It does not restate or replace the normative PAC-2026 SF-4 machine-readable freeze.Source: Supplementary Material: Making… · p. 25
  • The supplement’s retained details are not the normative protocol source.Source: Supplementary Material: Making… · p. 25

1. Authority and Reading Boundary

The supplement identifies the governed machine-readable package as the semantic authority for the evaluated SF-4 layer. Its schemas and validators are referenced by package identity rather than duplicated in prose.

1. Authority and Reading Boundary

  • The governed AIJIM-PAC-2026-SF-4 package, version 0.4.0-sf.4, is the semantic authority for the evaluated SF-4 layer.Source: §1 · p. 25
  • SF-2 and SF-3 remain immutable historical predecessors, while supplement tables and prose are projections of identified objects rather than authority to modify them.Source: §1 · p. 25
  • Full schemas, reason-code registries, canonical vectors, validators, and model configurations are referenced by package identity to avoid a second source drifting from executable bytes.Source: §1 · p. 25

2. Formal and Registered Identifier Detail

The section retains article-level shorthand for typed identity and registered obligations, while the governed Registry, Profile, schemas, and test vectors remain controlling.

2. Formal and Registered Identifier Detail

  • The article retains shorthand for the frozen typed-identity operation over an object type, required context, and body.The passage introduces these inputs but does not provide the operation’s full expression.Source: §2 · p. 25
  • The shorthand uses admitted canonicalization K0 and admitted SHA-256 H0 without replacing the governed Registry, Profile, schemas, or test vectors.Source: §2 · p. 25
  • The article expands obligation labels into registered identifiers, while the machine-readable Core Registry remains controlling.The table distinguishes article labels from registered identifiers.Source: §2 · pp. 25–26

3. Versioned Design and Risk Crosswalks

The crosswalks distinguish conceptual mappings and design rationale from executable conformance authority, and keep the Core’s enforceable responsibilities separate from claims it does not establish.

3. Versioned Design and Risk Crosswalks

  • The six atomic Core Registry obligations remain the executable conformance authority; architectural-family mappings indicate conceptual refinement, not identity or retrospective evidence for PAC-2026.Source: Table S1 · p. 26
  • The obligation-removal scenarios explain the responsibility lost without a check, but are design rationale rather than executed ablations or proof of universal minimality.They also do not establish that the other five coordinates remain conformant.Source: Table S2 · p. 26
  • The risk synthesis distinguishes what the fixed PAC-2026 Core Profile can enforce from residual deployment and scientific risks.Source: Table S3 · p. 27
  • Each observer-local obligation identifies a record transition, divergence class, falsification witness, and stronger claim it does not establish.Source: Table S4 · p. 27

4. Finite-Model and Fault-Injection Detail

Finite-model checks explore bounded configurations: safe reachable states and unsafe mutation runs are reported separately, and exhaustion does not establish unbounded safety or cross-model refinement.

4. Finite-Model and Fault-Injection Detail

  • The ten finite models factor fixed-profile obligations into bounded state spaces, with safe reachable states distinct from unsafe configurations.Unsafe counts represent expected violations or declared blind spots, not additional safe reachable states; exhaustion supports only encoded configurations and invariants.Source: §4 · p. 27, Table S5 · p. 28
  • Each unsafe configuration maps to its expected invariant in the frozen TLC-check script, with model paths and invariant names bound in the evaluation SSOT.Source: §4 · p. 28
  • The factorized checks do not establish a cross-model refinement proof or unbounded safety.Source: §4 · p. 28

5. Historical SF-3 Reference-Path Evidence

The historical material separates predecessor-path regression evidence from superseded novelty evidence and limits each conformance class to its stated responsibility.

5. Historical SF-3 Reference-Path Evidence

  • The historical SF-3 cases document how a private reference path handled authorization successors, as regression evidence rather than SF-4 conformance or independent implementation evidence.The frozen tables report expected observations from an immutable predecessor vector and bounded record/permit consequences; the collected suite passed source, extracted-sdist, and isolated-wheel runs, but node counts are engineering coverage.Source: §5 · p. 28, Table S6 · p. 29, Table S7 · p. 29

6. Historical, Superseded Rival Procedure

  • The rival analysis used a corpus closed on 27 July 2026 and is retained for historical traceability, not current novelty evidence or the later composition challenge.It withheld credit for newly written integration rules, and a later audit found relevant pre-cut-off omissions.Source: §6 · p. 30, Table S9 · p. 30

7. Conformance and Refinement Boundary

  • A conforming class licenses only its stated responsibility; complete protocol conformance, independent interoperability, production readiness, and standard status require additional open gates.Source: §7 · p. 30, Table S10 · p. 31

8. Artifact and Evidence Index

The index distinguishes scored evaluation results, reproduction evidence, and bounded implementation claims, while documenting unresolved ambiguities and access limits. Its artifacts support repeatability and bounded constructibility, not complete conformance or broader interoperability.

Artifact inventory

  • The executable corpus includes ten TLA+ models, 76 expected unsafe configurations, and 110,764 distinct reachable states across safe configurations.The corpus also includes schemas, reason codes, identity vectors, temporal vectors, and validation artifacts.Source: §8 · p. 32

Historical implementation

  • All 17 historical SF-3 authorization-successor cases produced their expected normalized outcomes, with 373-test suites passing from source, source distribution, and installed wheel.The evidence is bounded to a private, unmerged predecessor path and establishes neither complete conformance nor general refinement.Source: §8 · p. 32

Cross-path evidence

  • A bounded binding-refinement package rejected 22 negative cells before permit derivation, while developer calibration covered 23 public cases without establishing interoperability.The refinement package reports 24 unchanged result cells; calibration used visible cases and expected projections.Source: §8 · p. 32

Evaluation boundaries

  • G7-B’s scored outcomes have qualifications: known case classes and integration against expected results constrain claims of implementation and evaluation independence.Twelve scored archetypes used terminal-only A1 targets whose compared Rust terminal came from the bridge operation table; withdrawn and repaired cases also bound interpretation.Source: §8 · pp. 33–34

Evaluation results

  • 183/183 scored instances matched the registered endpoint; raw analyzer matches (210/240) and four-way equality (219/240) answer different questions.The reported 75 expected acceptances are not all-coordinate passes or publication permissions.Source: §8 · p. 33

Unresolved findings

  • HC-0024 leaves M1 acceptance unresolved, whereas HC-0062 leaves authoritative-time admission and rejection localization ambiguous; neither case showed an incorrectly issued permit.The index also records withdrawn cases, specification clarifications, untested anchor ranges, and harness-conversion limits.Source: §8 · p. 34

Reproduction

  • On a same-host reproduction, G7-B passed all eight required reproduction checks; this was neither an external replication nor another blind study, and the restricted precursor deposits were not rerun.The optional online beacon re-fetch alone was skipped in offline mode; stored beacon and time proofs were checked.Source: §8 · p. 35

Availability and scope

  • The archived evidence collection preserves source records and reproduction materials, but some precursor and separate-consumer deposits remain restricted and excluded.The rival-source audit supports only the bounded finding NO_COMPLETE_SUBSTITUTE_IN_BOUND_CORPUS, not a systematic or human-independent absence proof.Source: §8 · pp. 35–36

9. Bounded Correction Candidate After the SF-4 Evaluation

The post-evaluation correction proposes bounded changes to commitment binding and time handling while preserving the distinction between two unresolved SF-4 ambiguities. Its local validation is additional specification evidence, not regenerated SF-4 evaluation or complete conformance.

Correction design

  • M1 and R1 bind canonical disclosure and run–artifact sets to the candidate, checking measured-entry assignments and artifact membership without requiring referenced value bytes to be supplied.Missing entries, substituted commitments, or missing value references yield localized failure; reference membership does not establish byte availability.Source: §9 · p. 36

Time and adjacent requirements

  • The candidate binds time-bearing evaluation objects to one parsed instant, rejects substituted instants at the bound layer, and makes selected context and lifecycle-payload requirements executable.It preserves HC-0024’s acceptance-relevant M1 ambiguity as distinct from HC-0062’s time and rejection-localization ambiguity; neither original case showed an incorrectly issued permit.Source: §9 · p. 37

Validation and limits

  • 49 local correction cases were validated alongside 17 SF-3 regression and 23 SF-4 temporal cases, without rescoring G7-B’s 183 scored instances among 240 generated instances.Porting the correction to Core and evaluating a regenerated successor corpus remain incomplete; the repair is bounded, not complete conformance or field validation.Source: §9 · p. 37

Related papers

Loading 2609.17631v1…