Governed candidate corpus · review pending

Mechanistic interpretability

Features, circuits, interventions, sparse representations, causal tests, and faithfulness criteria represented separately from explanatory confidence.

Primary stress point

A human-readable feature label, probe, attention pattern, or ablation effect does not by itself establish a complete or faithful causal explanation.

Canonical factory depth

11 of 30 governed factory records are active canonical releases. The candidate corpus remains below the public line until each record independently passes exact-hash review and release.

Canonical public layer

Records that passed the gate

Open JSON registry
published canonicalcomparison

Representation probing boundary

A source-bounded comparison record for representation probing boundary within mechanistic interpretability.

Inspect claims and provenance →

published canonicalconcept

Induction head circuits

A source-bounded concept record for induction head circuits within mechanistic interpretability.

Inspect claims and provenance →

published canonicalconcept

Neural feature superposition

A source-bounded concept record for neural feature superposition within mechanistic interpretability.

Inspect claims and provenance →

published canonicalconcept

Sparse autoencoder dictionaries

A source-bounded concept record for sparse autoencoder dictionaries within mechanistic interpretability.

Inspect claims and provenance →

published canonicalmeasurement

Attention pattern evidence

A source-bounded measurement record for attention pattern evidence within mechanistic interpretability.

Inspect claims and provenance →

published canonicalmeasurement

Superposition geometry

A source-bounded measurement record for superposition geometry within mechanistic interpretability.

Inspect claims and provenance →

published canonicalmechanism

Polysemantic neurons

A source-bounded mechanism record for polysemantic neurons within mechanistic interpretability.

Inspect claims and provenance →

published canonicalmechanism

SAE encoder decoder

A source-bounded mechanism record for sae encoder decoder within mechanistic interpretability.

Inspect claims and provenance →

published canonicalmethod

Causal scrubbing

A source-bounded method record for causal scrubbing within mechanistic interpretability.

Inspect claims and provenance →

published canonicalmethod

In context learning circuits

A source-bounded method record for in context learning circuits within mechanistic interpretability.

Inspect claims and provenance →

published canonicalmethod

Toy models of superposition

A source-bounded method record for toy models of superposition within mechanistic interpretability.

Inspect claims and provenance →

Below the public line

Draft inventory remains private until canonical release.

The public surface exposes aggregate capacity only. Draft identifiers, titles, routes, claims, source packets, and review blockers are excluded from crawlable pages and registries.

Withheld records

19

Withheld typed edges

22

Domain contract

The registry exposes the boundary, not just the content.

Graph records

30

Typed edges

34

Public canonical

11

Withheld

19

Read the governing architecture in the Epistemic Publication System.