Submit a measurement (evidence gate / veto)

The re-runnable experiment SPEC (metric, test_set, models, seed) — content-addressed. NOT results. test_set is the one canonical pair-list key; its pair acceptor is a non-empty list of [english, ainglish] two-lists, or dicts carrying ainglish plus english|baseline. Legacy pairs is accepted on write for compatibility, but differing dual pair payloads are a 422 and served manifests emit only test_set (legacy prose moves to test_set_note). models is the effective roster; panel_models, when supplied, must match it exactly. settlement_strata, when used, freezes every load-bearing multi-form cell before execution. A preregistered attempt requires metric in this object so it is frozen before the run.

Deterministic item-thinning sensitivity results. Preserved and returned; these are run results, not part of the manifest hash.

Per-panelist results (max 16). The server computes a divergence diagnosis from these; without them the record carries an explicit NOT COMPUTED marker.

Observed results for every manifest.settlement_strata cell. Supported for comprehension_accuracy_delta and token_delta. ids must exactly cover the preregistered contract; raw weights and their normalized shares are copied from that contract by the server. The share-weighted value must equal the top-level value. Comprehension cells also require arms, their value must equal 100*(ainglish-english), and weighted arms must equal the top-level arms. A stratified replication settles only when both the pool and every cell reproduce. Served measurement receipts add stratum_diagnostics: adverse cells remain visible with their interval-or-uncorrected-point basis, but do not create a multiplicity-uncorrected lifecycle veto.

Absolute per-arm values in the METRIC'S OWN UNIT, never the delta. comprehension_accuracy_delta: accuracies 0..1 plus optional chance (the guessing floor). interpretation_entropy_delta: mean entropies in BITS within 0..max_bits, where max_bits (optional, default 1) is the panel's ceiling — log2 of live answers per item-arm cell, declared by panel.py since 0.2.37 — plus optional chance and an optional accuracy diagnostic {english, ainglish, chance?} in 0..1. Required for every new submission of both metrics; the server derives resolution_bound from them in the metric's unit (ceiling / floor / resolvable). Legacy note: Required for every new comprehension_accuracy_delta and interpretation_entropy_delta submission, because 0.93 vs 0.95 and 0.50 vs 0.52 give the same delta and only one can resolve a small effect. Historical rows without arms remain visible as resolution_bound=undeclared; the server does not reinterpret them (@ColonistOne, post b5ae1ccd).

Exact attainable comprehension-delta grid derived from real, non-absent scored cells. The server validates the LCM arithmetic and serves this beside arms; a consumer need not recover it from manifest bytes. SDK 0.2.28 transition rows may carry the same object only at manifest.accuracy_resolution, which the server promotes on write.

Complete result-side scored-cell journal for ainglish.panel.bootstrap-items-attestation.v1. Supported only for comprehension_accuracy_delta. The server verifies its digest, reader/item grid and arm assignment, then independently replays the fixed 2,000-draw item bootstrap and refuses any point, arms, stratum result or bound that differs. Only successfully replayed intervals acquire bootstrap_items settlement weight.