{ "schemaVersion": "1.6", "slug": "assignment-aligned-effective-states", "title": "When a neuron label is not enough", "shortTitle": "When a neuron label is not enough", "url": "https://evidencepress.org/releases/assignment-aligned-effective-states/", "oneLine": "A neuron can rank its assigned meaning well while its thresholded label loses information needed for decisions and edits.", "abstract": "An audit of one published five-variable 2-SAT transformer separates semantic ranking, threshold fidelity, numerical effective-state geometry, final-readout agreement and clause-edit sufficiency. Seventeen source-nominated labels rank strongly on protected data. A fresh, balanced 4096-formula source-grammar bank retains strong ranking, but some fixed thresholds miss many positives. An exact Boolean edit collision identifies information absent from the specified report. This is a single-checkpoint case study, not a verdict on competing general theories of representation.", "datePublished": "2026-10-18", "dateModified": "2026-10-18", "version": "4.8.0", "doi": "10.5281/zenodo.23454637", "doiUrl": "https://doi.org/10.5281/zenodo.23454637", "conceptDoi": null, "pdfUrl": "https://github.com/ipitchford/assignment-aligned-effective-states/releases/download/v4.8.0-candidate/paper.pdf", "altPdfUrl": "https://github.com/ipitchford/assignment-aligned-effective-states/releases/download/v4.8.0-candidate/supplement.pdf", "altPdfLabel": "Technical supplement (PDF)", "repoUrl": "https://github.com/ipitchford/assignment-aligned-effective-states", "releaseUrl": "https://github.com/ipitchford/assignment-aligned-effective-states/releases/tag/v4.8.0-candidate", "markdownUrl": "https://evidencepress.org/releases/assignment-aligned-effective-states/index.md", "bibtexUrl": "https://evidencepress.org/releases/assignment-aligned-effective-states/cite.bib", "audioUrl": "https://evidencepress.org/assets/audio/assignment-aligned-effective-states.mp3", "imageUrl": "https://evidencepress.org/assets/og/assignment-aligned-effective-states.png?v=aed8eca866", "coverArtUrl": "https://evidencepress.org/assets/art/assignment-aligned-effective-states.svg?v=6ac7047053", "media": [ { "type": "audio", "url": "https://evidencepress.org/assets/audio/assignment-aligned-effective-states.mp3", "name": "Plain-English audio briefing", "description": "Communication aid; not additional research evidence.", "transcriptUrl": "https://evidencepress.org/assets/audio/assignment-aligned-effective-states.txt" } ], "authors": [ "Anonymous" ], "license": "CC0-1.0", "licenseAppliesTo": "original prose, data and diagrams only", "archiveLicense": "LicenseRef-Component-Specific", "licenseScope": "Original prose, data and diagrams: CC0-1.0. Original code: MIT. Retained upstream components keep their own terms. External dependencies are not redistributed. Consult the package rights inventory; the archive as a whole is not licensed under CC0.", "status": "unrefereed-candidate", "verification": { "peerReviewed": false, "independentlyReproduced": false, "formallyVerified": false, "internallyReplayed": true, "detail": "Single-checkpoint AI-assisted audit with 18 public producer-replay stages passed. Not externally peer reviewed, independently reproduced or formally verified. Version 4.8.0 is archived on GitHub and Zenodo." }, "assurance": [ { "dimension": "availability", "label": "Availability and archiving", "question": "Is the evidence package publicly retrievable from an archive under a persistent identifier?", "state": "passed", "note": "GitHub candidate tag v4.8.0-candidate and Zenodo record 23454637 are public. All five asset sizes/provider digests match; four smaller assets also passed download parity. Archive access does not establish correctness or independent reproduction." }, { "dimension": "internalReplay", "label": "Internal replay", "question": "Does the producer’s own pipeline reproduce the stated result from the archived package?", "state": "passed", "note": "Clean producer replay passed all 18 public stages with 1367 payload files unchanged. This is not the private 23-stage historical route, independent reproduction or a new checkpoint replication." }, { "dimension": "independentRerun", "label": "Independent rerun", "question": "Has someone else run the supplied implementation and obtained the stated result?", "state": "not-assessed", "note": "No unaffiliated replication established." }, { "dimension": "independentReimplementation", "label": "Independent reimplementation", "question": "Has someone else reached the result from an independent implementation?", "state": "not-assessed", "note": "Producer checking is not independent reproduction." }, { "dimension": "formalVerification", "label": "Formal verification", "question": "Is a formalised statement machine-checked, and over which trusted base?", "state": "not-assessed", "note": "No proof-assistant verification." }, { "dimension": "specialistReview", "label": "Specialist review", "question": "Has a domain specialist assessed the argument?", "state": "not-assessed", "note": "No external specialist review established." }, { "dimension": "editorialPeerReview", "label": "Editorial peer review", "question": "Has a journal or venue run peer review to a decision?", "state": "not-assessed", "note": "Internal AI-assisted review only." }, { "dimension": "dataEnvironmentReproducibility", "label": "Data and environment reproducibility", "question": "Are data and computational environment pinned well enough to rebuild?", "state": "not-assessed", "note": "Second-host replay not established." } ], "provenance": { "aiGenerated": true, "aiAssisted": true, "generatedBy": [ "OpenAI Codex: audit, revision and publication preparation" ], "humanRole": "Publication authority; no human scientific contribution inferred.", "reviewStatus": "Internal AI-assisted review, not external peer review." }, "attributionBoundary": null, "problem": null, "orderedProductAverage": null, "corrections": [], "publicCorrections": [], "publicCorrectionIndexes": null, "pageStructureVersion": "2.0", "pageStructureVariant": "canonical-theorem", "pageStructureWaivers": [], "recordMaturity": "current", "metadataProvenance": "Version 4.8.0 research gate and clean 18-stage public replay passed; 1367 payload files unchanged by replay. GitHub and Zenodo publication confirmed with all five asset sizes and provider digests matched; the four smaller assets were also downloaded from both providers.", "grandfatheredAtSchemaVersion": null, "keywords": [ "mechanistic interpretability", "2-SAT", "semantic ranking", "threshold fidelity", "causal abstraction", "reproducibility" ], "keyResults": [ "Strong assignment ranking does not establish fixed-threshold fidelity.", "A fresh 4096-formula balanced source-grammar bank supports ranking transfer within the same checkpoint, not independent replication.", "A current truth table and the specified bit report cannot determine every Boolean-enumerated clause-replacement outcome." ], "reviews": [], "evidencePackage": "Protected tests, same-checkpoint fresh-bank native forwards, Boolean enumeration and scoped producer replay; broader checkpoint generalization is unestablished.", "openProblems": [], "mathObjects": [], "relatedWorks": [ { "citation": "Palumbo et al. (2025), Validating Mechanistic Interpretations: An Axiomatic Approach. Source model and interpretation.", "url": "https://proceedings.mlr.press/v267/palumbo25a.html", "doi": null }, { "citation": "Cammarata, Neural Networks Are Already Clean. Motivating essay, not a theorem tested wholesale.", "url": "https://already-clean.goodfire.com/", "doi": null }, { "citation": "Geiger et al. (2025), Causal Abstraction: A Theoretical Foundation for Mechanistic Interpretability.", "url": "https://jmlr.org/papers/v26/23-0058.html", "doi": null } ], "claimAssurance": null, "operatingModel": { "version": "1.0", "workId": "ep-work:assignment-aligned-effective-states", "attemptIds": [ "ep-attempt:assignment-aligned-effective-states-publication" ], "aims": [ "science" ], "artifactRoles": [ "evidence-assessment", "research-output", "communication" ], "lineageId": null, "accelerationPrimitives": [ "identification-gate", "counterexample-proxy-first", "assurance-vector", "agent-readable-research-object" ], "decisionObject": { "type": "measurement-requirement", "description": "Separate semantic ranking, threshold fidelity and action sufficiency before using a neuron report.", "scope": "One fixed five-variable 2-SAT checkpoint and declared banks/reports." }, "bottleneckTargeted": [ "assurance", "publication" ], "semanticBridge": { "state": "explicit", "description": "Boolean edit outcomes, native forward responses and classifier agreement are distinct estimands.", "remainingRisks": [ "Single-checkpoint scope; no broader theory discrimination.", "Fresh bank is balanced and history-excluded, not natural-prevalence IID." ] }, "humanJudgmentGates": [ "Publication authorised; interpretation of semantic and causal scope requires specialist judgment." ], "parentLinks": [ { "relation": "depends-on-claim", "workId": null, "legacyReleaseSlug": null, "externalUrl": "https://proceedings.mlr.press/v267/palumbo25a.html", "inheritedClaim": "Published checkpoint, neuron labels and interpretation.", "inheritedAssuranceCeiling": "Source study is an antecedent; this producer audit is not independent external replication." } ], "assuranceTarget": { "dimensions": [ "independentRerun", "specialistReview" ], "nextAction": "Rerun the public package and test the frozen audit on independently trained checkpoints.", "claimCeiling": "A bounded empirical case study, not universal representational or alignment conclusions." }, "impactClaims": [ { "id": "assignment-aligned-effective-states-no-impact", "aim": "science", "outcome": "A reusable separation of interpretation endpoints.", "setting": "One 2-SAT transformer", "status": "NO_IMPACT_EVIDENCE", "designClass": "none", "comparator": "No workflow comparator.", "estimand": "No acceleration or downstream impact effect measured.", "evidenceRefs": [], "registeredDesignRef": null, "independentAssessment": null } ] }, "researchMetrics": { "schemaVersion": "1.0", "policyUrl": "https://evidencepress.org/api/research-metrics-policy.json", "outcomePolicyUrl": "https://evidencepress.org/api/v2/research-metrics-policy.json", "workLedgerUrl": "https://evidencepress.org/api/work-ledger.json", "claimCeiling": "These records describe scoped workflow inputs and outcomes. They support calibration and reference-class construction, but do not by themselves establish research acceleration, theorem correctness, novelty, independent assurance or impact.", "attempts": [ { "attemptId": "ep-attempt:assignment-aligned-effective-states-publication", "status": "published", "resultClass": "positive", "metrics": { "schemaVersion": "1.0", "measurementScope": "assurance-through-publication", "scopeBoundary": "Remaining publication only; prior research and replay left-censored.", "forecast": { "frozenAt": "2026-10-18T15:24:40Z", "procedureClass": "remaining-candidate-delivery", "targetOutcome": "Scoped archived candidate with canonical readback.", "expectedActiveMinutes": 90, "plausibleLowMinutes": 45, "plausibleHighMinutes": 180, "expectedUnattendedWaitMinutes": 30, "referenceClass": { "label": "Remaining candidate publication", "sampleSize": 0, "basis": "Procedural judgement, not empirical calibration." }, "fermiComponents": [ { "component": "Archive, media and guarded delivery", "count": 1, "lowMinutesPerUnit": 45, "centralMinutesPerUnit": 90, "highMinutesPerUnit": 180, "basis": "Final package/provider/site gates remain." } ], "probabilityPositiveSignal": 0.9, "probabilityTargetClosure": 0.8, "probabilityHorizonMinutes": 360, "assumptions": [ "No provider or rights blocker." ], "reforecastTriggers": [ "Provider failure or public-byte mismatch." ], "stopRule": "Never weaken assurance to meet forecast." }, "outcome": { "status": "unavailable", "policyVersion": "2.0", "recordedAt": "2026-10-18T17:25:21.471Z", "reason": "Gate and delivery observations do not provide complete effort telemetry. No retrospective reconstruction or calibration.", "missingFields": [ { "field": "calendarElapsedMinutes", "reason": "Observed opening and delivery times are retained as milestones; elapsed wall time is not active effort and no numeric metrics series is emitted." }, { "field": "activeAgentMinutes", "reason": "The scoped activeAgentMinutes counter was not instrumented or aggregated; unknown is not zero, and no estimate is inferred from logs, elapsed time or left-censored work." }, { "field": "unattendedWaitMinutes", "reason": "The scoped unattendedWaitMinutes counter was not instrumented or aggregated; unknown is not zero, and no estimate is inferred from logs, elapsed time or left-censored work." }, { "field": "blockedMinutes", "reason": "The scoped blockedMinutes counter was not instrumented or aggregated; unknown is not zero, and no estimate is inferred from logs, elapsed time or left-censored work." }, { "field": "reworkMinutes", "reason": "The scoped reworkMinutes counter was not instrumented or aggregated; unknown is not zero, and no estimate is inferred from logs, elapsed time or left-censored work." }, { "field": "agentRuns", "reason": "The scoped agentRuns counter was not instrumented or aggregated; unknown is not zero, and no estimate is inferred from logs, elapsed time or left-censored work." }, { "field": "maxParallelAgents", "reason": "The scoped maxParallelAgents counter was not instrumented or aggregated; unknown is not zero, and no estimate is inferred from logs, elapsed time or left-censored work." }, { "field": "modelTurns", "reason": "The scoped modelTurns counter was not instrumented or aggregated; unknown is not zero, and no estimate is inferred from logs, elapsed time or left-censored work." }, { "field": "researchCycles", "reason": "The scoped researchCycles counter was not instrumented or aggregated; unknown is not zero, and no estimate is inferred from logs, elapsed time or left-censored work." }, { "field": "falsificationGatesRun", "reason": "The scoped falsificationGatesRun counter was not instrumented or aggregated; unknown is not zero, and no estimate is inferred from logs, elapsed time or left-censored work." }, { "field": "candidateArchitecturesTested", "reason": "The scoped candidateArchitecturesTested counter was not instrumented or aggregated; unknown is not zero, and no estimate is inferred from logs, elapsed time or left-censored work." }, { "field": "candidateArchitecturesRejected", "reason": "The scoped candidateArchitecturesRejected counter was not instrumented or aggregated; unknown is not zero, and no estimate is inferred from logs, elapsed time or left-censored work." }, { "field": "substantiveReviewRounds", "reason": "The scoped substantiveReviewRounds counter was not instrumented or aggregated; unknown is not zero, and no estimate is inferred from logs, elapsed time or left-censored work." }, { "field": "p0Findings", "reason": "The scoped p0Findings counter was not instrumented or aggregated; unknown is not zero, and no estimate is inferred from logs, elapsed time or left-censored work." }, { "field": "p1Findings", "reason": "The scoped p1Findings counter was not instrumented or aggregated; unknown is not zero, and no estimate is inferred from logs, elapsed time or left-censored work." }, { "field": "prepublicationClaimCorrections", "reason": "The scoped prepublicationClaimCorrections counter was not instrumented or aggregated; unknown is not zero, and no estimate is inferred from logs, elapsed time or left-censored work." }, { "field": "resultState", "reason": "Published candidate delivery is observed in statusHistory; incomplete workflow telemetry does not negate publication." }, { "field": "positiveSignalObserved", "reason": "Candidate publication evidence is retained separately; no calibrated signal metric is inferred." }, { "field": "targetReached", "reason": "Delivery target is recorded in statusHistory; no calibrated target metric is inferred." }, { "field": "forecastErrorMinutes", "reason": "Active effort was not measured; forecast error cannot be computed from wall time." }, { "field": "forecastRatio", "reason": "Active effort was not measured; elapsed time is not a valid active-effort denominator." }, { "field": "withinForecastInterval", "reason": "Active effort was not measured, so interval membership is unknown." }, { "field": "activeHumanMinutes", "reason": "The scoped activeHumanMinutes counter was not instrumented or aggregated; unknown is not zero, and no estimate is inferred from logs, elapsed time or left-censored work." }, { "field": "computeMinutes", "reason": "The scoped computeMinutes counter was not instrumented or aggregated; unknown is not zero, and no estimate is inferred from logs, elapsed time or left-censored work." }, { "field": "deduplicatedModelTokens", "reason": "The scoped deduplicatedModelTokens counter was not instrumented or aggregated; unknown is not zero, and no estimate is inferred from logs, elapsed time or left-censored work." }, { "field": "uncachedInputTokens", "reason": "The scoped uncachedInputTokens counter was not instrumented or aggregated; unknown is not zero, and no estimate is inferred from logs, elapsed time or left-censored work." } ], "calibration": null } }, "missingnessReason": null } ] } }