# VK-IRA Run 01–33｜Validation Closure & Benchmark Matrix ## 0. Executive decision After Run 01–33, VK-IRA has enough architectural stabilization to release: $$ \boxed{\text{VK-IRA v1.0 Research Specification}} $$ but **not** enough evidence to claim: $$ \boxed{\text{Universal Error-Governance Architecture}} $$ or a statistically certified detector accuracy. The release means: > the architecture, object model, closure rules, governance rules, and cross-domain interface are stable enough for a first research specification. It does **not** mean: > universality has been proven, or benchmark performance has been established on a representative population. **作者：** Neo.K **機構：** EveMissLab／一言諾科技有限公司 **日期：** 2026-10-07（整理日期；原稿未標註日期） **版本：** 原稿未標註（本次整理版） --- # 1. Scope of the first validation campaign Numbered run entries: $$ \boxed{33} $$ The campaign covers materially different artifact families: ```text Theory / conceptual claims Mathematics Source representation Software Empirical research History Law Physics Data pipelines Machine learning Engineering / control Laboratory / biology Language / translation Cross-representation composition ``` The important point is not raw domain count. It is that the same core primitives survived very different failure topologies: - proof dependency; - software call/runtime dependency; - evidence dependency; - historical transmission; - legal amendment/authority; - physical regime; - data lineage; - ML information flow; - closed-loop feedback; - contamination/assay provenance; - semantic representation; - cross-representation adapters. --- # 2. What the campaign actually demonstrated ## 2.1 Detection is not mandatory VK-IRA successfully produced: $$ \boxed{\text{Repair}} $$ and: $$ \boxed{\text{Abstain}} $$ as distinct legitimate outcomes. An architecture that always finds a defect is not a validator. ## 2.2 A miss is part of the evidence Run 15A remains: $$ \boxed{\text{MechanismLocalization}=\text{MISS}} $$ The predicted mechanism did not match the maintainer fixes. ## 2.3 An oracle can fail Runs 13 and 14A established: $$ \boxed{ \text{CandidateOracle} \neq \text{QualifiedOracle} } $$ ## 2.4 Error existence, localization, and repair are separate $$ \text{Detect} \neq \text{Localize} \neq \text{BlastRadius} \neq \text{Repair} \neq \text{RepairEquivalence}. $$ ## 2.5 Domain validity is indexed $$ \boxed{ \text{Validity} = \text{Validity}(\text{Object}\mid \text{Domain},\text{Time},\text{Representation},\text{Assumptions}) } $$ ## 2.6 Residual origin is not its downstream shadow $$ \boxed{ \text{ResidualOrigin} \neq \text{ResidualManifestation} } $$ --- # 3. Benchmark accounting policy A single global "accuracy" number is **not reported**. The runs mix internal theory refinement, blind oracle tests, partial oracles, executable synthetic controls, historical source criticism, real engineering incidents, and negative controls. Pooling them would create a meaningless denominator. Instead VK-IRA v1.0 reports a structured ledger: ```yaml outcome_classes: pass: partial: miss: abstain: blocked: oracle_invalid: do_not_score: ``` Performance metrics should be computed only inside comparable benchmark strata. --- # 4. Important non-successes ## Run 08 Math oracle was partial. ## Run 12 No known correction was found; this is not proof of correctness. ## Run 13 Independent residuals were found, but the selected future `fix` commit was irrelevant. ## Run 14A A notebook appeared heavily changed at file-stat level, but semantic source change was not established. ## Run 15A Semi-blind mechanism prediction was wrong. ## Empirical full replay A clean public-data + public-code + frozen blind audit + qualified later correction + full executable replay case remains open. --- # 5. Release gates | Gate | Requirement | Status | Evidence | |---|---|---|---| | G1 | Cross-domain breadth: ≥ 8 materially distinct domain/profile families | **PASS** | Theory, math, code, empirical, history, law, physics, data, ML, engineering/control, lab/bio, language, XREP. | | G2 | Positive + negative behavior: Must support repair and abstention | **PASS** | Multiple paired positive/negative controls; explicit P_abstain since v0.4. | | G3 | Failure preservation: At least one genuine miss retained | **PASS** | Run 15A semi-blind mechanism localization miss retained with score 0. | | G4 | Oracle governance: Reject invalid oracle instead of retrofitting success | **PASS** | Runs 13 and 14A rejected candidate oracles from scoring. | | G5 | Executable validation: At least three domains with executable replay | **PASS** | Code, data, ML, control, language-frame, XREP include executable checks. | | G6 | Cross-representation composition: At least one end-to-end representation-chain test | **PASS** | Run 32–33 localizes Math→Code adapter failure and validates equivalent unit systems. | | G7 | Residual causality: Distinguish origin from inherited manifestations | **PASS** | v0.14 first-failure rule + inherited manifestation state. | | G8 | External empirical positive full replay: Public data/code, blind audit, later qualified correction, full replay | **OPEN** | Still not cleanly completed; access/oracle issues and Run 15A miss remain important. | | G9 | Independent third-party reproduction: Architecture run by independent teams/models with frozen protocol | **OPEN** | Current evidence is generated inside this research program. | | G10 | Statistical performance benchmark: Large preregistered corpus with precision/recall/FP/FN estimates | **OPEN** | Heterogeneous case studies are not a statistically representative benchmark. | --- # 6. v1.0 release interpretation Passing G1–G7 is enough to stabilize the **research specification**. Open G8–G10 prevent stronger empirical claims. ```yaml specification_maturity: v1.0 domain_general_candidate: strong universal_claim: false statistical_performance_certified: false independent_reproduction_complete: false ``` --- # 7. Strongest current claim $$ \boxed{ \text{VK-IRA is a strong domain-general error-governance architecture candidate.} } $$ Supported meaning: > a common validity-kernel / residual / typed-causal-graph / adapter / closure framework has remained useful across materially different domains and can represent repair, abstention, negative controls, misses, oracle uncertainty, dynamic propagation, and cross-representation failure localization. Not supported: > it detects all meaningful errors; > it is optimal; > it is universal; > it has a known precision/recall on the space of real-world errors. --- # 8. Core architecture at closure $$ \boxed{ O_t \rightarrow \text{Parse} \rightarrow \mathfrak D \rightarrow \text{RepresentationAdapters} \rightarrow K \rightarrow R \rightarrow G_{\mathrm{typed}} \rightarrow BR/DBR \rightarrow \text{Decision} \rightarrow \text{Patch/Abstain} \rightarrow \text{Sandbox} \rightarrow \text{Closure} \rightarrow \text{OracleGovernance} \rightarrow \text{Certificate} \rightarrow O_{t+1} } $$ --- # 9. First campaign conclusion The campaign began from: $$ \text{Can a theory be repaired without destroying its valid core?} $$ It ended at: $$ \boxed{ \text{Can validity, residuals, provenance, causal propagation, representation changes, repair and abstention be governed by one common architecture?} } $$ Run 01–33 provide substantial evidence that the answer is **sometimes yes across many different domains**. That is enough for a v1.0 research specification. It is not enough to close the research program.