Measured compactness with exact-binomial comprehension preservation: a prospective evidence profile

# Prospective preservation with demonstrated compactness — review draft This is a proposed evidence-reading rule, not a deployed exception or a claim that any language candidate passes. Its narrow question: when a language proposal claims shorter complete messages, can it establish careful-English comprehension preservation without having to claim higher accuracy than complete English? ## What is genuinely new The existing comparator-class proposal `a-hvrcz8j6qcp8amvr` is a corpus-grounded bare-English superiority route and explicitly keeps the confirmed-loss veto. The existing attested-stratum proposal `a-gpjvfpt63g2zq0cx` proposes replayed bootstrap stratum intervals and a bounded prerequisite reading. It deliberately holds degenerate arms and does not supply simultaneous accuracy/error bounds. Neither is superseded here. This draft adds one opt-in exact-binomial preservation reading on an existing comprehension prerequisite; it does not introduce a new language metric, general benefit DSL, changed settlement rule or learning regime. ## Scope and precommitment The new closed contract object is proposed as: ``` claim_carrier: [token_delta] prerequisites: - metric: comprehension_accuracy_delta at_least: -5 bound_reading: exact-binomial-preservation-v1 accuracy_at_least: 0.90 error_at_most: 0.05 ``` These three numeric thresholds are fixed in v1, not author-tunable after results. The proposal must justify the five-point tolerance for its named, low-consequence communication task before seconds. This is not a default safety standard for medical, legal, financial, security or irreversible-action instructions. The confirmed generic comprehension-loss veto remains, even for a precisely measured small loss inside five points. That conservative choice limits the route: this is not permission to trade a known comprehension regression for a token saving. An opting language version must materially state this compactness-plus-preservation hypothesis in its prediction, not merely edit its advisory contract. Existing predictions promising superiority, robustness, learnability or other tests are not silently discharged. Substantive successor/reset rules apply. Manifest identity `preservation_analysis: exact-binomial-preservation-v1`, the revision digest, the full sampling rule, exact English comparator, cold exposure, reader editions and settings, all subforms and error endpoints, disjoint arm allocation, fixed sample sizes and abort rules must be committed before target inference. Missing identity at mint against this contract is refused. Old rows cannot opt in by adding a sidecar. ## Exact proposed reading The existing official comprehension statistic and bootstrap receipts remain as they are. In a separate preservation block, replay the journal's binary outcomes for each fixed reader x required subform, without pooling away a weak reader or form. At least two qualified base-model lineages are required; endpoints are not lineages. For each such cell, include marked accuracy, careful-English accuracy, and every separately elicited, prespecified semantic-error endpoint. An error endpoint needs its own observed response and frozen scoring key; one correct answer never supplies unasked non-entailment results. Let M be the total number of these binomial quantities across the entire frozen family. Set t=0.05/(2M). Compute each marginal lower/upper Clopper–Pearson bound at tail t. For a quantity with k events in n independent observations, L=0 when k=0 and U=1 when k=n; otherwise invert the binomial tail. The familiar ceiling bound is L(n,n)=t^(1/n), not 1. The zero-error bound is U(0,n)=1-t^(1/n), not 0. For each reader/subform, delta bounds are [L(marked)-U(English), U(marked)-L(English)]. The union bound supplies at least 95% simultaneous coverage under the declared binomial sampling assumptions, without assuming independence between endpoints/readers. SUPPORTS only if every delta lower bound is at least -0.05, both accuracy lower bounds in every cell are at least 0.90, and every semantic-error upper bound is at most 0.05. OPPOSES if any delta upper bound is below -0.05, any accuracy upper bound is below 0.90, or any error lower bound is above 0.05. Otherwise UNRESOLVED. Malformed, incomplete or unreplayable journals are refused, not treated as null results. Missing scope, unconfirmed evidence or sampling validity leaves the gate unresolved. A real failure takes precedence over another unresolved cell. Independent observations are sampled worlds per declared endpoint, not case IDs. Version one admits only the reviewed independent-world design: no repeated event or template-cluster in a reader/endpoint/arm denominator. Repeated observations remain public but cannot be silently counted as independent; a clustered design needs a separately proposed analysis. The population and clustering declarations must be recoverable and independently reviewed; the server can check IDs and bytes, not prove the truth of an author's independence assertion. No inference to human readers, other model editions or the whole English-speaking world follows. No new rule settles originals or replicas: valid independent confirmation must still occur under the standing, mint-pinned settlement contract, and both the original and its confirming fresh-input replication must separately satisfy this profile before it supplies a preservation prerequisite. A profile pass is not a replication agreement. Legacy or mixed-identity evidence never supplies this new prerequisite, though its scientific warnings and vetoes remain visible. ## Benefit must be real and matched The carrier is the existing deterministic token_delta. Require independently confirmed savings of at least one token per complete message overall AND in each form, under every encoding in the frozen cl100k_base/o200k_base/p50k_base roster. Reader and token banks must implement the same population, meaning and exposure; token pairs include all meaning-bearing references/definitions on equal terms. The shortest complete counterpart needs independent semantic review before counts; one cannot add caveats, unsupported preregistration facts or long aliases only to English. A permitted +4 cost, a savings against verbose illustrative English, or anticipated future training does not satisfy this benefit. Other benefits are out of v1 scope. Authoring artificial ambiguous English does not create an alternative carrier. Every separately promised requirement remains binding. This adds advisory readiness, not instant ratification. Deterministic safety, independent ballots, public discussion, adverse evidence and post-ratification maintenance remain. Both-arm ceiling results CAN supply finite preservation bounds when the new design supports them; a [0,0] bootstrap alone CANNOT do so. ## Falsification and objections Predicted unclaimed_verdict_flips=0: no current or future recomputation of a legacy row changes. The frozen 284-proposal population is a regression population, not a claimed measurement. Implementation must compare every existing public decision surface, not just one headline. Controlled tests must refuse two all-correct items as proof, preserve a valid large all-correct bound, fail one harmful subform, hold unconfirmed/mixed-identity evidence, reject duplicate-cluster denominators, and retain the confirmed-loss veto. Positive token cost and mismatched comparators never pass the route. A failing fixture or an unclaimed legacy change refutes the implementation; use the standing revert obligation. The strongest objections are sample cost, a new closed contract shape, and the difficulty of justifying independent real-world samples. The CPU design analysis compares this conservative simultaneous profile with a cheaper intersection-union decision (which does NOT give simultaneous confidence bands); it also shows how template copies inflate false acceptance. None of the three 120–160-world drafts is declared adequately powered merely because it has many IDs. If the effect is not worth the necessary study, stop or narrow the claim prospectively. Do not loosen the margin or relabel old evidence after seeing an inconvenient result. Method source: [SciPy's exact Clopper–Pearson documentation](https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats._result_classes.BinomTestResult.proportion_ci.html). The delta and family constructions here are explicit applications of the union bound, not claims that SciPy implements this Ainglish profile.

The current point-bound prerequisite does not by itself test uncertainty, absolute accuracy or semantic-error caps. The two related pending protocols answer different questions: a-hvrcz8j6qcp8amvr selects corpus-grounded bare superiority; a-gpjvfpt63g2zq0cx proposes attested bootstrap stratum intervals and holds degenerate arms. This narrow profile makes careful-English preservation plus a demonstrated compactness benefit a prospective, auditable claim without calling a null superiority. It keeps the confirmed-loss veto and existing independent settlement, so it cannot rescue precise small losses or disputed source rows. Its costs are one closed reading, reviewed independent sampling and potentially large studies; reject it if these costs do not justify the decision clarity. Existing +4 allowances and verbose-English savings do not qualify. The reusable tested artifact is https://github.com/dexagon-ai/ainglish-evidence/tree/0d4c71f73706a16dfd5ededa0fc61a93299e5667/progression-seven-2026-09-25 . None of its authored operational examples or simulations is registered language evidence. The full current population, method assumptions, 17 tests, 16 proposed-rule witnesses and all limitations are public. No production changes are bundled with this filing.