ContextSymbolics
Auditing Mechanistic Interpretability

Auditing Mechanistic Interpretability

Claims, evidence, fractures, and certification

Mechanistic interpretability should be audited as a measurement science. Its methods may recover information, alter behavior, expose causal participation, support control, or improve monitoring. These are real achievements. They are not interchangeable achievements.

A result earns only the strongest claim that survives alternative measurements, causal tests, fracture pressure, and residual analysis.

1. Scope

Mechanistic interpretability studies internal model states through probes, sparse decompositions, feature dictionaries, Jacobians, steering directions, circuit analysis, causal interventions, chain-of-thought inspection, activation patching, and runtime monitors. The audit begins after a method appears to work.

A successful decoder may show that a target is recoverable. A steering vector may show that a direction influences behavior. A circuit may mediate part of a task. A monitor may reduce operational risk. None of these conclusions automatically establishes a native semantic variable, a complete mechanism, or a safety guarantee.

The audit therefore examines five distinct objects: the model computation, the interpretability method, the recovered object, the stated claim, and the proposed application. Errors arise when evidence about one object is promoted into a claim about another.

2. What this document is not

This audit takes no position on the global geometry of model state. It assumes no theory of representation, requires no commitment to any interpretability program, and endorses no framework of its own. Its standards are empirical and its verdicts are about evidence.

The audit is therefore usable by anyone, including researchers who disagree with every structural claim made elsewhere on this site. That is intentional. An instrument that only certifies its owner's results is not an instrument.

The audit rules on what the evidence has earned. It does not rule on what the model is.

3. Evidence classes

ClassNameWhat it establishesWhat remains unearned
OObservationAn internal quantity covaries with a conditionRepresentation or cause
DDecodingA target is recoverable under a stated measurement procedureThat the model uses the decoded variable
IInterventionA modification changes a measured outcomeSemantic identity or clean control
CCausal attributionA component or state participates causally in a bounded computationCompleteness or uniqueness
MMechanism recoveryA structured causal process predicts computation over a declared domainValidity outside that domain
UOperational utilityA method improves monitoring, debugging, control, or auditingMechanistic truth

Class U deserves particular care. Utility is the most defensible achievement in the field and the most frequently exchanged for a claim it did not buy. A method that reliably reduces operational risk has earned deployment. It has not thereby earned an ontology.

4. The claim ladder

  1. Covariation: an internal quantity changes with a condition.
  2. Decodability: a method recovers that condition from model state.
  3. Predictive generalization: recovery survives held-out or shifted data.
  4. Representation: the recovered structure is identifiable, stable over a declared domain, distinguishable from competing constructs, and available to the model.
  5. Interventional influence: manipulating the structure changes behavior.
  6. Selective causal participation: the target effect follows the proposed causal role without broad unrelated damage.
  7. Mechanistic organization: related states, transitions, or components form a stable native process.
  8. Partial mechanism: the process explains a measurable portion of the capability, with residual computation reported.
  9. Mechanistic account: the process explains the relevant computation over a declared domain and survives its implied invariance tests.
Audit rule: no rung implies the next. Promotion requires new evidence.

The ladder is not a schedule. It carries no presumption that a result at one rung will eventually reach the next, and no presumption that failure to climb indicates a temporary shortage of effort. Some results stop where they stop because that is where the evidence stops. Others stop because the higher rung does not exist for that object.

5. What counts as representation?

A representation need not be a neuron, direction, sparse feature, or fixed coordinate. It may be a subspace, relation, trajectory, transition rule, distributed configuration, context-indexed state, or equivalence class. Stability applies to the role or transformation law, not necessarily to one coordinate.

A representation claim should establish identifiability, observer robustness, domain stability, temporal availability, model availability, and causal relevance. A signal that can be decoded only by one instrument, one labeling scheme, or one narrow prompt set is better described as an observer-conditioned measurement than as a stable internal object.

Objectivity does not require an observer-free measurement. It requires structure that remains stable across admissible observers and instruments.

6. Observer dependence

Interpretability never begins without a target vocabulary, decoder class, dataset, decomposition, labeling procedure, and evaluation regime. The recovered object is jointly conditioned by model state, measurement method, observer ontology, and data distribution.

This does not make interpretation arbitrary. It changes the burden of proof. Alternative labels, bases, decoder families, seeds, widths, penalties, and decompositions become competing measurements. Structure that persists across admissible measurements has a stronger claim to model-native status.

The audit also asks what the instrument could not have found. A vocabulary that contains no term for a phenomenon will not report its absence. The negative space of a method is part of its result and should be stated, not left to the reader.

7. Temporal participation

A signal may be recoverable before a decision, during its construction, after it as a trace, or later as a reportable summary. A binary before-or-after test is insufficient for distributed computation.

QuestionAudit target
First recoverabilityWhen does the information become decodable?
First causal efficacyWhen does intervention first change the result?
Peak mediationWhere does the proposed state carry the largest causal effect?
PersistenceDoes the state remain stable, rotate, split, or disappear?
RevisionIs the state updated or overwritten by later computation?
Re-expressionDoes a later signal reconstruct an earlier computation mainly for report?

8. Causal roles in redundant systems

Binary necessity and sufficiency are too coarse for distributed models. A valid mechanism may be contributory but not individually necessary, conditionally sufficient but not globally sufficient, or one member of a substitutable family.

RoleMeaningRequired evidence
ContributoryChanging the object alters the computationControlled intervention with matched damage
Conditionally necessaryThe object is required under stated background conditionsAblation across controlled contexts
Conditionally sufficientThe object recovers the target under stated background conditionsInsertion or restoration with controls
RedundantAlternative paths can substitute for the objectJoint ablation and pathway substitution
MediatingThe proposed path carries a measurable portion of the effectMediation analysis and path intervention

9. Behavioral and process restoration

Behavioral restoration shows that an intervention recovers the measured output. Process restoration additionally recovers intermediate dependencies, timing, sensitivity, and error structure. A restored answer may be produced by an artificial shortcut. Strong mechanism claims require evidence that the repaired model recovers the relevant computational trajectory rather than only the final answer.

Restoration evidence should state which of the two was achieved. A report that says only that performance returned has reported behavior and should be classified as such, however the repair was motivated.

10. State validity and selective causality

Density alone does not determine whether an intervention is valid. A rare state may be naturally reachable, while a high-density state may violate learned relations among layers, positions, or residual components. State-validity evidence may include reachability from actual inputs, preservation of learned conditional relations, agreement with naturally occurring states, downstream calibration, reversibility, and absence of unrelated competence collapse.

Selectivity does not require zero downstream change. It requires that the intended causal family changes disproportionately relative to unrelated damage and matched interventions.

11. Claim-invariance contracts

Every substantive claim should declare the transformations under which it is expected to remain valid. Fracture tests should be derived from those implied invariances. A claim with no declared invariance is not conservative. It is untestable.

ClaimExpected invarianceFailure meaning
Token-specific detectorMay vary under retokenizationUsually a domain boundary
Concept representationShould survive meaning-preserving paraphrase and hard negativesConstruct failure or overnarrow domain
Language-independent conceptShould survive translation, tokenizer, and lexical variationClaim falsification
Stable semantic directionShould survive admissible basis, context, layer, and checkpoint changesCoordinate dependence or locality
Global mechanismShould survive relevant context, position, model-state, and distribution changesGlobal claim falsification
Safety monitorShould survive the declared adaptive threat envelopeDeployment claim failure
A fracture falsifies a claim when it violates an invariance implied by that claim. Otherwise it delimits the domain.

12. The fracture register

Fractures occur when nearby, equivalent, or claim-related states produce abrupt failures of interpretation, recovery, intervention, or correspondence. The register below lists the standard pressures. It is an inventory of tests, not a theory of the model. Each pressure is neutral with respect to what the failure means; the meaning is fixed by the invariance the claim declared.

PressureApplied byInvariance typically implied
TokenizationRetokenization, spelling, whitespace, segmentationConcept identity independent of surface segmentation
FormattingMarkup, casing, delimiters, presentation protocolIdentity independent of presentation
SyntaxMeaning-preserving rewrite, voice, clause orderSemantic rather than lexical identity
Context lengthExtension, truncation, distant dependencyStability of the state under accumulation
Layer and positionCross-layer and cross-position transferA static coordinate rather than a moving one
NormalizationRescaling, gain, residual compositionMetric relations preserved under rescaling
ProjectionAlternative decoder, dictionary, basis, seed, widthObject exists independent of the instrument
SaturationMagnitude sweep, dose-response, sign reversalSmooth response over the operating range
Finite precisionQuantization, reduced precision, roundingStructure survives discretization
Adversarial pressureOptimized inputs, adaptive attackLocal continuity of the claimed relation
Model versionFine-tune, checkpoint, retrainDurable identity of the coordinate
CompositionConcept combination, intervention orderingIndependence and commutation of the parts

13. Verdicts

A failure outside the declared domain is a boundary. A failure inside the domain is a failed replication. A failure against an implied invariance is claim falsification or demotion. A failure caused by an invalid intervention is inconclusive. A failure revealing an alternative path reduces completeness. A failure appearing only under one instrument is an observer-dependence warning.

Three reporting faults recur and should be recorded by name whenever they are found.

FaultDescription
Unauthorized promotionA claim is stated at a rung above the one its evidence reached
Unreported exclusionA phenomenon the instrument could not have found is not listed as unfound
Unlicensed extensionA local result is presented as holding over a domain that was never tested

None of the three requires bad faith and none is rare. Each is a defect of reporting, correctable by reporting.

14. Bounded completeness and the residual

Absolute completeness is rarely measurable because the full denominator is not known. Completeness should therefore be reported relative to explicit targets: fraction of behavior recovered, task loss recovered, causal effect mediated, state variation explained, prompt classes covered, or known failure modes predicted.

The residual is not noise by default. It may contain redundant traces, downstream records, alternative mechanisms, missing features, or safety-relevant capacity. Residual information must be audited through the same ladder: decoding, temporal precedence, intervention, mediation, and restoration.

The unexplained portion is where the next mechanism lives, and where an unsafe capability would hide if it were hiding anywhere.

15. Mechanistic equivalence classes

Different bases, decompositions, circuits, or causal graphs may explain the same observed behavior. This may reflect inadequate experiments, true model symmetry, or multiple valid descriptive levels. When explanations are observationally and causally equivalent, the audit should report the equivalence class and identify the invariant structure shared across its members.

Disagreement about feature counts is not by itself scientific disagreement. Two decompositions that preserve the same causal transitions may be one result described twice.

16. Method-specific standards

Linear probes

A successful linear probe establishes linear recoverability on a defined activation distribution. It should report held-out and shifted-distribution performance, calibration, nuisance controls, matched-capacity baselines, cross-layer and cross-position transfer, intervention effects, and stability after fine-tuning.

Steering vectors

For activation edits, test dose-response, sign reversal, matched random directions, unrelated-task degradation, state validity, layer and position dependence, reversibility, restoration, and vector composition.

Sparse autoencoders

Report reconstruction error, loss recovered, downstream behavior recovered, sparsity distribution, dead and rare features, dictionary-width sensitivity, cross-run feature matching, causal quality, missing known concepts, and information retained in the residual. A concept that was expected and not found should be reported as not found, with the same prominence as a concept that was found.

Circuits

A circuit claim requires an explicit task, causal paths, conditional necessity or sufficiency, redundancy analysis, restoration, bounded completeness, alternative-circuit comparison, and transfer beyond discovery prompts.

Chain of thought

Chain of thought may be useful without being a faithful transcript of internal computation. Audit retrospective faithfulness separately from prospective causal use through truncation, substitution, false reasoning, irrelevant insertions, answer preservation, hidden-state evidence, and monitor-aware training.

17. Safety audit

A safety claim must state its threat model, attacker knowledge, monitor visibility, optimization budget, false-positive and false-negative costs, distribution shift, adaptive pressure, fallback defenses, and containment after failure. Robustness is always relative to an explicit threat envelope.

Benchmark detector, held-out detector, adaptive detector, monitor-aware detector, deployment defense, and defense system are different claims. Survival at one rung does not certify the next.

An untested monitor is not a robust monitor. It is an untested monitor.

18. Institutional claim audit

local effect → feature → concept → thought → mechanism → safety capability

Every arrow requires additional evidence. The audit should compare the paper, abstract, laboratory post, podcast, press coverage, policy briefing, and product claim. Caveats retained in the paper do not automatically survive public translation. A caveat that appears only in the supplementary material has been preserved, filed, and removed from the claim.

The audit records the exact point of unauthorized promotion, and by whom it was performed. The paper, the abstract, and the press release may each have a different answer. A technically faithful satire may afterwards dramatize that promotion. The technical judgment is established first; the satire may only repeat it.

19. Auditing the audit

These standards are themselves subject to the ladder. They can be misused in three ways, and the misuses should be named as plainly as the ones they detect.

An audit that cannot be failed by its own author is a certificate the author issued to himself.

20. Audit certificate

Result: ______________________________

Classification: Observation / Decoder / Intervention / Monitor / Local causal model / Partial mechanism / Mechanistic account

Target construct: ______________________________

Competing construct excluded: ______________________________

Declared domain: ______________________________

Claim-invariance contract: ______________________________

Temporal role: ______________________________

Causal role: Contributory / Conditionally necessary / Conditionally sufficient / Redundant / Mediating

State-validity evidence: ______________________________

Restoration: None / Behavioral / Process

Completeness denominator: ______________________________

Recovered fraction: ______________________________

Pressures applied: ______________________________

Fractures found: ______________________________

Expected and not found: ______________________________

Unexplained residual: ______________________________

Threat envelope: ______________________________

Strongest justified claim: ______________________________

Typical unauthorized promotion: ______________________________

Explicitly unearned claims: ______________________________

21. Conclusion

Mechanistic interpretability should be audited with the same seriousness applied to the models it examines. Its tools can expose structure, recover information, alter behavior, diagnose failures, support control, and provide inexpensive monitoring. The objective is not to demand total understanding before partial methods may be used. It is to state exactly what each method has earned.

visibility ≠ decodability ≠ representation ≠ control ≠ causal participation ≠ mechanism ≠ safety assurance
State what was recovered. Name the domain. Declare the invariances. Test the causal role. Restore the process. Search for fractures. Report what the instrument could not have seen. Report the residual. Do not promote utility into mechanism without evidence.