We build instruments to inspect representations, compare models, test interpretations, and keep the evidence attached.
A representation is a reading taken through an instrument. A correspondence is a map to be tested. A convincing result is the beginning of the control work.
Research relationshipsExplanatory map / select a program
Conceptual connections, without measured coordinates or a universal chronology.
Correspondence / portable reading
Which relationships carry between models?
Geometric interpretability compares relational organization and asks what a fitted map carries from one model into another. Fit, held-out reading and intervention are different tests; a shared frame can help organize a comparison without identifying a universal map of meaning.
Layer readings, cross-model maps, composed systems and audit tools meet here. Follow a question into its method, an inspectable artifact, and the control that could change the conclusion.
I
What does a representation reveal?
CHRONOS and Aperture ask how layer, subspace and reference-frame choices shape a reading. Jacobian lenses estimate directional change; ridge lenses fit a readout. These instruments make different measurements even when they examine the same model. An extraction recipe, a normalization choice and a held-out split belong to the scientific object.
Geometric interpretability, static transport and shared dimension study correspondence without assuming identical coordinates. A fitted translator may carry a probe while losing fine structure; an aggregate comparison may survive a route that a deployed readout does not. The important question is what transfers, under which extraction and test contract.
Residual feedback asks about an engineered return path; receiver integration asks about combining components in a bounded instructed study. Calibration, control and self-report assays can answer particular engineering questions. Their interpretation still depends on shortcut tests, component comparisons and broader capability checks.
A benchmark floor asks whether a surface-only classifier can recover the labels. A calibration contract asks whether a decision rule behaves as specified. A receipt asks whether bytes and accounting match. Photon puts grounding, readouts and verification into a composed system. Each property needs its own test and its own claim.
The June manuscript reports held-out concept matching and a transported concept-pair probe between Qwen2.5-3B and Llama-3.2-3B. A separate grounding assay reveals a benchmark confound and a weaker deployed route.
Three tests / explanatory schematic
01Fit
Estimate a correspondence from the declared fitting material.
→
02Read
Evaluate the transported readout on a separate test set.
→
03Intervene
Test a behavioral effect and the capability cost.
Same coordinates?Not required
Same relational organization?A measurable question
Portable causal control?A separate experiment
The reported fit on 420 concept centroids and matching on 180 held-out concepts test a concrete correspondence. A separate 300-pair probe assay asks whether an instrument travels with it.
Fitting success can coexist with poor generalization. A portable probe can coexist with weak steering. Excessive steering can reduce capability; a reading result does not resolve that tradeoff.
Parallel named concepts supply correspondence supervision. The centered fitting set identifies at most a 419-dimensional span. This is a coarse centroid result under one concept set and extraction; fine-feature transfer is a separate test.
Source: Is Meaning All You Need?, §5(a–b). First-party report; exact weight revisions and original acquisition receipts are not established here.
A transported concept-pair readout
Reported manuscript result
Qwen2.5-3B / Llama-3.2-3B program · 300 held-out concept pairs
source probe→ fitted R →target readout
Concept-pair classification accuracy (0–1)
Readout / control
Reported accuracy
Ported probe
0.947
Native target probe
0.911
Shuffled map control
0.495
The readout remains useful after transport on the reported concept-pair assay; the shuffled map is near chance. This is classification accuracy, not fabrication AUROC or a transfer of general capability. Pair reuse and per-pair test-instance counts are not fully specified by the passage.
Corrected public account · familiarity-reading AUROC (0–1)
Reading route
Reported AUROC
Native readout
0.839
Direct pairwise map
0.833
Deployed hub route
0.455
Surface A* 0.855; battery DISQUALIFIED. The input surface carries label information. Only 1 of 18 cells separates after correction. These are historical route comparisons with limited evidential interpretation, distinct from the concept-pair accuracy result.
Explore a retained thought-basin snapshot: search a concept, inspect its record, then follow its recorded nearest neighbors. The map gives the archive a readable surface.
647displayed basin records
10,963generated thoughts mapped
768D → 2DNomic centroids → t-SNE display
The displayed basins have at least seven thoughts, selected from a recorded pool of 12,237 basins. These are embeddings of generated text; they are distinct from the model-internal activations used in transport research.
A basin is an archive record.
Thoughts sharing an anchor are averaged into a 768-dimensional centroid. A two-dimensional projection places selected records on the screen; position and neighborhood in that display depend on the projection. Recorded citation and novelty fields are metadata, not verified truth or discovery scores.
Could a classifier recover the labels without seeing a model? The surface floor turns that question into an inspectable test. Our withdrawn headline is one of the examples.
Withdrawn / July 25, 2026
0 / 36The former Photon fabrication headline
Historical lexicality A* 0.967
The recovered 147-claim battery was DISQUALIFIED: surface-only probes recovered its labels. That old headline did not survive the benchmark control. The value here is the fixed public report; a browser recomputation is labeled separately.
The public browser NullTest refits surface probes under label permutations. Its verdict is limited to the probes that can run and the battery's power. A PASS cannot rule out an untested feature, and small samples can hide large artifacts.
The public ledger retains the withdrawn headline, a tie-break fix, sample-size comparability limits, and the difference between a rounded AUROC of 1.000 and a stored value of 0.9997. Each correction changes what the original claim can support.
Receipt verification checks an artifact under a key and format contract. A one-byte mutation can invalidate a signature while a faithfully signed claim can still be wrong. Source custody and scientific validity require their own evidence.
AFFERENT studies engineered residual feedback: a channel that supplies structured information about a measured state, and a receiver that must decide how to use it. Its public record describes methods; detailed source and result access remain separate.
Feedback method sketchSchematic / no experiment runs
01 / MeasureRead a state
Declared instrument & tap
→
02 / ReturnEncode a signal
Specified feedback path
→
03 / ReceiveIntegrate a component
Bounded receiving contract
← compare outcomes, controls and capability ←
The instrument defines the reading.
Specify the layer or tap, reference points, normalization and calibration. Changing the instrument can change the apparent state. A reading is a measurement under that contract.
Control question: does the readout survive a different reference or a shortcut test?
“Nerve” is an engineering metaphor for a feedback path. This illustration supplies no biological or consciousness inference.
Read an original, follow its corrected account, inspect an artifact, or choose an instrument. The public shelf contains 26 catalog entries with different access and evidence states.
nulltest.py and the local fixed-battery reviewerCan benchmark labels be recovered from the input surface?Public utility; private review wrapperPublic upstream source; wrapper not released
Inputs
A JSON battery with text, binary labels and optional strata; no model access needed.
Method
Out-of-fold surface probes with a permutation-based decision rule.
Output
Surface-floor verdict, effect and test diagnostics.
Prerequisites
Public nulltest source, Python standard library and an authorized labeled JSON battery.
Limit
A PASS is limited to the probes, sample size and test configuration. The local review wrapper is not itself a published release.
Signed verification / broader charterWho signed these bytes, and what does the tool contract cover?Public verification utility and proposal ledgerPublic source; additional tools vary
Inputs
Exact signed bytes, signature/public key and governing verification spec.
Method
Signature verification plus a public ledger of available and proposed tools.
Output
Authenticity checks and release/proposal status.
Prerequisites
Versioned verifier/specification, signed payload and the issuer public key.
Limit
Signatures do not establish truth. Transparency witnessing/co-signing and proposed tools are separate maturity questions.
Keep methods and versions attached. Preserve negative findings. Choose controls that test the statistic in use. Read first-party reports as first-party reports until independently reproduced.
Collaboration
Source review and reproduction
Name the artifact or instrument you want to examine, the result you want to reproduce, or the control you think is missing. The collaboration page explains access and reproduction.