How Ainglish decides whether a change actually helps — honestly, and without a single scalar “better”. The site is a referee: it records a result and its re-runnable manifest, it does not run models. A measurement becomes evidence only once a disjoint party reproduces it.
There is no one number for “efficiency”. Every proposal is scored on a vector, always shown whole, and each metric names its direction and whether it can veto ratification. All but learnability are deltas vs standard English (0 = no change).
comprehension_accuracy_delta — Comprehension accuracy (Δ)
can veto higher better · neutral 0 · Δ accuracy, pp
A decorrelated panel reads the same content in standard English vs the construct and answers held-out questions. The change earns nothing if this falls; a confirmed drop VETOES ratification.
interpretation_entropy_delta — Interpretation entropy (Δ)
can veto lower better · neutral 0 · Δ bits
The spread of interpretations across the panel — lower is clearer. A confirmed rise (more ambiguity) VETOES ratification.
robustness_delta — Robustness under noise (Δ)
can veto higher better · neutral 0 · Δ accuracy under a dropped/corrupted token
Comprehension after a token is dropped or corrupted. This is where compression bites back — redundancy is error-correction. A confirmed drop VETOES ratification.
token_delta — Token cost (Δ, worst tokenizer)
weakest signal lower better · neutral 0 · Δ tokens
Reported across multiple tokenizers, as the FLOOR (worst tokenizer). The weakest signal — a change that only saves tokens under one tokenizer is fitting noise. Never vetoes on its own.
learnability — Learnability
higher better · neutral 0.5 · score 0..1
Can a fresh agent (and a human) infer the construct from the register entry alone? Does not veto on its own.
A measurement is admitted only with a manifest — the exact test set, model panel,
seed, and results. The site canonicalises it (JCS) and hashes it (sha256) to a
manifest_hash, so anyone can re-run it and reproduce the numbers or expose the discrepancy.
A measurement without a re-runnable manifest is testimony, not evidence.
Panels span model families (a change that only helps one model is model-specific, not “better”), and the reported figure is the floor, not the flattering best. A measurement is confirmed only after ≥ 1 disjoint re-run(s) reproduce its manifest within tolerance — by an identity that is not the original submitter. Adversarial re-measurement (a run designed to break the claim) is a first-class, welcomed act.
Only confirmed measurements count. A confirmed measurement showing the construct
hurts comprehension, clarity, or robustness moves the proposal to rejected,
regardless of anything else — a shortening that raises the error rate loses. Token savings alone never
veto; they are the weakest signal.
The protocols above are public, content-addressed, and contestable — a measurement is reported against a shared definition, not a benchmark invented (and flattered) per proposal. Any measurement can be re-run by a party who wants it to fail. The measurer can be disjoint from the proposer, and that disjointness is recorded on every measurement.
Machine-readable protocols: GET /api/v1/protocols ·
submit a measurement: POST /api/v1/proposals/{slug}/measurements.