Reader & instrument registry

Which systems actually produced the evidence? This inventory is derived from every public measurement row. It preserves exact roster identifiers and shows when a manifest binds a model digest. It does not turn similar names into families or repeated appearances into independent voices.

Generated 2026-08-28 15:46 UTC · machine-readable: GET /api/v1/readers

How to read the inventory

• An entry is an exact roster identifier, not an inferred model family and not an independent vote.
• A digest-pinned receipt establishes the served model artefact named by that measurement; it does not establish training-data composition or absence of Ainglish exposure.
• Provider, model and precision values are submitter-declared manifest fields. The registry does not infer ownership or operator independence from them.
• Measurement-row appearances are uses of an instrument, not independent samples and not a quality ranking.

Reader-panel panel_neff is still a declared, unvalidated decorrelation judgement; the registry does not “correct” it from model names. See the measurement methodology and inspect the underlying rows before treating a panel as diverse.

Model readers

Language-model instruments used to answer comprehension, ambiguity, robustness, learnability or fidelity tasks. 99 public measurement rows use this axis.

Tokenizers

Deterministic encodings used to price token cost. Exact legacy roster ids remain visible even when they used an obsolete provenance suffix. 308 public measurement rows use this axis.

Corpus slices

Frozen text samples used as the instrument for background-collision measurements. 3 public measurement rows use this axis.

Protocol re-runners

Named implementations or principals used to re-run project-machinery blast-radius checks. 44 public measurement rows use this axis.