Geometric Interpretability
Reading, writing and porting through a shared concept geometry. Public draft v0.2; Kevin Russell and Claude (Anthropic). Start with §0.2 claim status, §3 boundaries and §7 disconfirmation.
PDF edition ↗The research reading room
The Honesty.Tools
research program.
What does a representation reveal? Which relationships carry between models? What would make an interpretation fail?
We build instruments around those questions. This room connects the methods, public originals, corrected accounts and unfinished work—with a route from each question to the next.
15 navigation families · 26 catalog instruments
A map of the work, not a count of validated discoveries.
A route through the work
Research relationships
This is an explanatory map of research questions. The connections suggest a reading route; they do not represent measured distances or a causal pipeline.
Originals on the desk
Original papers
and corrected results.
The June 2026 manuscripts are original program documents. Read their claim tables, negative results and limits alongside the current corrected account.
Is Meaning All You Need?
A family-bounded concept geometry, up to rotation. Working draft; Kevin Russell and Claude (Anthropic). Start with §2 dual nulls, §3 the frame reversal and §10 disconfirmation.
PDF edition ↗Methods with corrections
Keep the source’s claim status beside the narrative. For an executable check, inspect the public surface-test utility and its release contract.
Honesty utilities source ↗No family matches these choices. Try a broader question or reset the reading room.
Representations & frames
Layer and frame instruments
Before interpreting a score, ask which encoder, frame, layer and baseline made it possible. These families put the instrument into the account.
Incomplete research source
CHRONOS
How does structure move through a model?
A single layer is a slice through a computation. CHRONOS asks what persists as that slice changes: a selected subspace, a relation between examples, or the orientation that makes a readout work. Its two catalog instruments approach the question through rotations and through movement between subspaces.
Method / instrument
The rotated lens combines PCA, Procrustes alignment and linear-probe transport. Grassmannian velocity instead constructs analogy subspaces with SVD, projects declared nuisance directions, and compares principal angles between consecutive layers. Their outputs are alignment/probe comparisons and subspace distances, respectively.
Artifact / evidence state
The catalog records incomplete research source: missing helpers and result gaps, fit/score reuse and transport-algebra issues. These descriptions explain an approach; they do not furnish a validated runnable instrument. Subspace resemblance alone does not establish a shared causal mechanism.
Connected instruments
Next question
Which structures persist on held-out examples after the nuisance and extraction choices change?
Program account · source unreleased
Aperture
How much of a reading belongs to the instrument?
Aperture makes the measuring frame part of the research question. A query can appear familiar under one reference set and uncharted under another. To interpret that difference, the encoder, reference corpus, descriptor frame and baseline must travel with the score. The aim is a reading whose construction can be examined.
Method / instrument
The empirical lens compares cosine neighbors against reference-to-reference baselines. The calibration probe generates anchors and applies novelty/separation heuristics and gates. The translation lens uses co-occurrence and PMI to study symbol/context structure. These are three distinct procedures, with different inputs and dependencies.
Artifact / evidence state
The public catalog describes the prototypes and their contracts; their executable source is not publicly released. An empirical reference baseline is different from a covariance-Gaussian null. Generator-driven anchor diagnostics are an active harness, and the translation source is not evidence for a different model-training study.
Connected instruments
Next question
Does a useful boundary remain stable when the reference population and phrasing change?
Public reading / artifact
Magic Eye & cartography
Are we seeing the population, or the frame we chose?
The synthesis manuscript begins with an instructive reversal: a proposed topology of generated thoughts became a study of the instrument that scored them. A descriptor frame can organize a population convincingly while supplying much of the structure itself. Cartography asks what remains after preserving the geometry that could produce that appearance.
Method / instrument
Silhouette compares embedding anchors built from named regions and researcher-defined descriptors. The cartography control constructs covariance-matched Gaussian surrogates, separately from shuffle/order controls. Matching covariance tests a geometric explanation; breaking correspondence tests a different explanation. Neither substitutes for the other.
Artifact / evidence state
The recorded basin atlas is available to explore. It maps generated-text basin centroids embedded with Nomic into a t-SNE display, rather than model-internal activations. The original synthesis is a June 2026 working draft. Prototype source is unreleased, and revised paper statistics are not all bound to the recovered implementation.
Connected instruments
Next question
Which conclusions survive replacing both the frame and its geometric baseline?
Program account · source unreleased
Jacobian & ridge lenses
What changes locally—and what can a readout recover?
Two instruments can operate at the same layer and still ask different questions. A directional derivative estimates a local response to a small change. A fitted linear readout asks which target information can be recovered from the recorded representation. Keeping those operations distinct prevents a predictive fit from becoming a causal account by implication.
Method / instrument
The Jacobian bridge uses directional finite differences at selected layers, with a declared step size and normalization. Ridge lenses fit regularized linear readouts under an explicit train/test split and tap recipe. The first returns derivative/readout estimates; the second returns performance and onset comparisons.
Artifact / evidence state
Both are public method descriptions of research prototypes whose source is not publicly released. Checkpoint/residual access, layer identity and preprocessing define the measurement. Original acquisition remains incomplete for historical work; a local estimate cannot support an unrestricted mechanism claim.
Connected instruments
Next question
Do layer comparisons agree when extraction and normalization are held fixed, and where do the two instruments disagree?
Correspondence & resolution
Cross-model correspondence
A fitted relationship becomes more useful when it survives held-out examples. Reading, rank, population and intervention remain separate questions.
Public reading / artifact
Geometric interpretability
Which relationships carry from one model to another?
Geometric interpretability studies correspondence as an instrument: fit a map between representations, hold out examples, and ask whether a reading or intervention survives the move. The founding document distinguishes a shared reference frame from a universal ontology. Its most useful organizing idea is a resolution boundary: coarse aggregate structure and fine per-concept structure need separate tests.
Method / instrument
The program connects relational similarity, fitted rotations, portable probes and transported steering. A relational statistic does not identify a full-dimensional translator. A probe that reads successfully does not demonstrate that a transported intervention will change behavior. Fit, held-out reading and intervention are separate stages, with separate controls.
Artifact / evidence state
Read the corrected /geometry account alongside the June 2026 originals. For Qwen2.5-3B → Llama-3.2-3B, it reports native AUROC 0.839, direct reading 0.833 and deployed hub 0.455 on a 162-item battery marked DISQUALIFIED: surface A* 0.855. These values describe that compromised battery, not a clean transfer certificate. The manuscripts also retain negative results and family/resolution limits.
Recorded manuscript observations · June 2026
Qwen2.5-3B / Llama-3.2-3B program. The synthesis manuscript’s §5 reports the following distinct assays. These original effects were not independently reconstructed for this reading room.
- Relational correspondence
- RSA 0.805 Meaning-shuffle z = 36. A relational statistic, rather than a fitted translator’s accuracy.
- Held-out centroid matching
- Top-1 1.000 PCA + orthogonal Procrustes fitted on 420 concept centroids; evaluated on 180 held-out concepts. Chance ≈ 0.0056. Named parallel concepts supply correspondence supervision.
- Transported concept-pair probe
- Accuracy 0.947 / 0.911 / 0.495 Ported / native target / shuffled map, across 300 held-out concept pairs. Classification accuracy, not fabrication AUROC or transferred model capability.
Exact weight revisions, tokenizer identities and acquisition receipts are not established by this passage. The centered 420-row fitting set identifies at most a 419-dimensional span; full-dimensional completion adds a choice.
Source: Is Meaning All You Need? · §5(a–c) →A different Qwen pair · fine-feature boundary
The founding document’s §3.6, dated 2026-06-08, reports Qwen3-1.7B ↔ Qwen3.5-2B: 2048-dimensional residuals, 32K-feature sparse autoencoders, with a full-dimensional map fitted on 6000 paired FLORES sentences.
- Cross-model feature matching
- Median max cosine 0.141 Standardized-map feature transfer, versus a random-direction floor of approximately 0.09. Feature cosine is different from concept-pair classification accuracy.
- Same-model SAE control
- Median feature matching 0.795 Separately trained SAEs on the same model and layer. Standardization improved coarse held-out sentence alignment to 0.689 while fine-feature transfer remained weak.
Manuscript-reported, not independently reproduced here. One model pair and feature/extraction recipe; this is an assay boundary, rather than a universal resolution law. Immutable weight revisions and per-direction feature subsets are not fully identified in the synopsis.
Source: Geometric Interpretability · §3.6 →
FIT
Learn a map
Paired fitting examples
READ
Test a held-out probe
Predictive recovery
INTERVENE
Test a transported change
Behavior & capability controls
Read methods & corrected measurements →
Founding field document · public draft v0.2 →
Is Meaning All You Need? · working draft →
This program account does not imply a separate released instrument.
Next question
At what resolution, rank and model distance does correspondence stop supporting held-out reading or intervention?
Program account · source unreleased
Static probes & transport
Can an instrument travel without being refitted to the answer?
Transport separates an instrument from the model coordinates in which it was built. The central practical question is whether a map learned on fitting examples carries a probe to genuinely held-out examples. Ingress, conditioning and orientation all become part of the contract: a valid file and a convenient alignment do not by themselves make a portable reading.
Method / instrument
Portable probes validate schemas and score pooled representations. Atlas fits and stitches transport maps with held-out evaluation and matched orientation controls. The whitening port uses shrinkage whitening, alignment and recoloring. The axis builder prepares declared corpora and evaluation inputs. This fitted transport Atlas differs from the public generated-thought basin atlas.
Artifact / evidence state
The catalog describes unreleased prototypes with bounded historical readouts. Some stored decisions can be inspected while the original activations and acquisition are incomplete. Fit/holdout separation remains a caller responsibility for the whitening port; its orientation controls are not full covariance-Gaussian surrogates.
Connected instruments
Next question
Can a held-out instrument retain its decision boundary under a new extraction recipe or model family?
Program account · source unreleased
Shared dimension
How wide is the correspondence we can actually recover?
A strong fit at one width does not settle how much structure is shared. This family asks whether held-out operator correspondence grows, plateaus or fails as the available width changes. The plateau is a research target, rather than a number to assume before the test.
Method / instrument
The described instrument combines fit-only whitening, ridge operators, correspondence shuffles and width checks. Fitting transformations on one split keeps evaluation examples from defining the coordinate system being tested. A correspondence shuffle breaks paired identity; a Gaussian surrogate preserves specified geometric structure. They answer different null questions.
Artifact / evidence state
This is a public prototype description, with executable source unreleased. The floor must match the statistic. Insufficient draws, uncertain contrasts and absent plateaus should remain in the result record. A width comparison cannot establish a universal dimensionality of meaning.
Connected instruments
Next question
Does the same plateau reappear under independently chosen populations and matched statistical power?
Program account · source unreleased
Language Genome
What changes when we change the population being compared?
Language Genome brings population definition into the foreground: which symbols, contexts and examples constitute the body under study? A model roster, a cache of representations and a completed training comparison are different objects. The research question depends on keeping those populations and their custody visible.
Method / instrument
The dark-matter census uses known-answer synthetic bodies and cached population comparisons. Width sensitivity and population definitions constrain what the comparison can show. Synthetic bodies can establish a checkable target; they do not automatically establish the same behavior in a natural model.
Artifact / evidence state
The public catalog offers a method account, not released executable source or a completed-run registry. Acquisition code is an active procedure, rather than a passive inspector. Population rosters, synthetic checks and model-training records must be assessed separately.
Related public study · disjoint symbols, shared authored world
The founding manuscript’s updates of 2026-06-08 and 2026-06-12 report two author-trained approximately 1.07B dense decoders using disjoint character pools to describe one authored world. Cross-model RSA was 0.794; scrambling concept identity reduced it to 0.021. Instance translation and exact ontology recovery remained partial.
This synthetic study is a separate related reading, not a source binding for the dark-matter census or a broad population-training claim. Shared architecture, world generator, ontology and objective remain common causes. Same-seed agreement across hardware checks determinism; it is not independent replication. Exact checkpoints, evaluation sample count and full acquisition custody are not established by this synopsis.
Source: Geometric Interpretability · §4 dated updates / §5 protocol →Read the public disjoint-symbol account →
Connected instruments
Next question
Which differences survive when corpus composition, width and cache construction are independently controlled?
Controls, calibration & feedback
Controls and residual feedback
A good control retains what might explain the result. A useful certificate can refuse. An engineered feedback loop needs evidence at each interface.
Program account · source unreleased
Causal controls
What must a control preserve to challenge the explanation?
A control is most informative when it retains the structure that could otherwise explain a result. Breaking everything can create an easy contrast. This family instead asks what happens after retaining specified geometric invariants while changing the correspondence or property of interest.
Method / instrument
The causal-excess instrument constructs controls that retain declared geometric properties. The retained invariants, changed variables and measured statistic need an explicit account. A control construction can challenge an interpretation; a causal-effect claim additionally needs the appropriate intervention and outcome evidence.
Artifact / evidence state
Executable source is not publicly released. The catalog explicitly records CT0-A as retired without a model result. That retirement is part of the method record, not a negative model experiment or a completed demonstration of causal excess.
Connected instruments
Next question
Which alternative explanation survives the strongest invariant-preserving control?
Program account · source unreleased
Statistical calibration
When should the instrument refuse to certify?
An instrument needs a rule for uncertainty as well as a rule for success. Statistical calibration makes the contrast, interval and refusal condition part of the artifact. It asks what a stored summary can warrant, and what information would be needed to make a stronger claim.
Method / instrument
The port certificate records trained-minus-floor contrasts, uncertainty and refusal conditions. RKC0 provides synthetic decision-calibration and interval accounting. TC2 commits versioned source and a synthetic-world seed. These artifacts occupy different positions: a decision record, arithmetic check and commitment are not interchangeable.
Artifact / evidence state
Public descriptions are available; executable source is unreleased. Summary uncertainty does not reproduce paired resampling of original predictions. Synthetic decision arithmetic does not admit a whole natural-model pipeline, and a committed plan is not a completion receipt.
Connected instruments
Next question
Would the certification decision survive paired analysis of the original predictions and a power-matched floor?
Program account · source unreleased
Residual feedback · AFFERENT
What changes when a model receives an engineered feedback signal?
AFFERENT extends the instrument question into a feedback loop: can a representation be converted into a signal, returned to a model, and used in a calibrated way? Here the loop is a program-level research question. The engineering metaphor of a nerve describes a proposed information path; it does not make a biological or consciousness claim.
Method / instrument
The catalog describes calibration-data construction, LoRA training and controlled evaluation. Interpretation requires separating signal content, receiver response and transcript cues. A model may learn the wording of an assay without learning to use the underlying signal, so shortcut-sensitive controls belong to the question itself.
Artifact / evidence state
A public program description is available; source and experiment artifacts are not publicly released through this catalog. Natural-model self-report calibration is not established. Failed or marginal results and transcript shortcuts limit the interpretation of any controlled assay.
Connected instruments
Next question
Can the receiver use the signal when wording cues change and the transcript shortcut is removed?
Program account · source unreleased
Receiver integration
Can a composed receiver preserve the components we mean to test?
The receiver question begins where a feedback diagram ends. Connecting components creates a new system whose behavior may depend on instruction, integration and the assay format. This family examines component-preserving integration within a bounded instructed study, rather than treating a loop diagram as a result.
Method / instrument
The public program account identifies the integration problem: preserve the intended component interfaces and evaluate the composed receiver under stated instructions. Readout calibration and system integration need separate records so success at one interface does not stand in for the whole loop.
Artifact / evidence state
The catalog does not imply a separate released instrument for this family. Paper-version lineage is not independent replication, and broader capability and checkpoint evidence remain incomplete. Public access here is the program description.
This program account does not imply a separate released instrument.
Next question
Which part of the composed behavior depends on the receiver, the instructions or the information actually carried by the signal?
Verification, systems & receipts
Verification and evidence
The public instruments make failure modes inspectable: surface shortcuts, component coverage, drifting annotations and altered bytes.
Public reading / artifact
Honesty utilities
Can a result be checked without asking its author?
These utilities turn evidence hygiene into something a reader can run: surface-label tests, numerical annotation linting, render checks, command contracts and signature verification. They test specific properties of a benchmark or artifact. Their value is a concrete failure mode, an inspectable implementation and a record that another reader can challenge.
Method / instrument
nulltest.py uses out-of-fold surface probes with a permutation-based rule. claim_gate.py checks numerical annotations against a registry. render_audit.py checks stale renders; nulltest_parity.sh checks bounded Python/JavaScript fixtures. honest.py and contract.py expose a command registry and release contract. Signature verification checks signed bytes.
Artifact / evidence state
Apache-2.0 upstream source is public. The local fixed-battery review wrapper is not released. A surface-test PASS is limited to the shipped probes, sample size and configuration. A linter checks provenance rather than scientific validity; fixture parity is bounded evidence of agreement; signatures attest bytes rather than truth.
Public source · Apache-2.0 →
Inspect the benchmark audit →
Run the browser surface test →
Read public utility contracts →
Connected instruments
Next question
Which benchmark conclusions remain after surface shortcuts, sample-size sensitivity and annotation drift are checked?
Public reading / artifact
Photon
What does a composed verification system actually verify?
Photon combines a served model, internal readouts, registry checks, routing and receipts. Each answers a different question. A registry can check whether an entity appears in its coverage; a readout can indicate familiarity; escalation can request another assessment; a receipt can bind an artifact. No single successful component settles the factual truth of the final answer.
Method / instrument
The public system profile separates those components and describes Photon Base as a local model designation. Inspecting the composed system means tracing a claim through the readout, grounding check and routing decision, then looking at failures of coverage and phrasing. The Prism preview is a place to inspect examples, with those limits attached.
Artifact / evidence state
The profile withdraws four n=147 headline rates on 2026-07-25 after the battery’s surface labels were recoverable at A* 0.967. The former 0-of-36 fabrication claim therefore does not establish honesty. Upstream weights, tokenizer and extraction gaps remain; the local name is not an immutable checkpoint identity. Known wording failures and lies about real entities remain important tests.
GROUNDING
Registry coverage
Existence is not factual truth
READOUT / ROUTER
Familiarity & escalation
Each needs its own evaluation
RECEIPT
Signed artifact
Authenticity is a separate property
Read the system profile →
Inspect known failures →
Open the Prism preview →
Read the correction record →
This program account does not imply a separate released instrument.
Next question
How does abstention behave across wording variants, uncovered entities and fluent falsehoods about real entities?
Public reading / artifact
Integrity & replay
Can we recover what was checked—and which bytes were signed?
Integrity connects a result to its schema, accounting and receipt. A replay can recover a stored decision under a declared contract; signature verification can establish whether bytes match a signer’s claim. These properties make artifacts inspectable, while leaving the truth and capability questions to their own evidence.
Method / instrument
The crossing spine is described as model-free integrity and bounded receipt replay. The public evidence desk separately demonstrates signature checking and mutation. Schema validation, replay and notary functions test different failure modes; they need artifact identities and versioned rules to be meaningful.
Artifact / evidence state
The crossing-spine source is unreleased. Public signature-verification source and demonstrations are available through the Honesty utilities and evidence desk. The public record also names incomplete transparency witnessing and external co-signing. Neither a valid signature nor a replay establishes truth, natural capability or physical containment.
Connected instruments
Next question
Can an independent reader verify the receipt, reproduce its decision and inspect a publicly witnessed record?
Work with the record
Research access
and collaboration.
Bring a benchmark, an extraction recipe, a control or a counterexample. Define the question and the evidence it would need. The public sources are a starting point; access to unreleased research is a separate conversation.