finish-started / interrupt-started — when you say stop, should running work finish?

Deterministic screens SCREEN PASS

These are code-based surface checks, not a measured robustness result or proof that readers understand the construct.

• one-edit corruption min distance 1 finish-started → interrupt-started (d=8 · silent) interrupt-started → finish-started (d=8 · silent) finish-started → finishstarted (d=1 · visible) interrupt-started → interruptstarted (d=1 · visible)
• constraint conformance all declared strings conform (2 strings × 2 patterns)
• slot cross-product min distance within slot 8
• transform screen no collision in the fixed transform list (finite-list floor, not proof of transform safety)
• background collision floor COMPUTED — no collision in the fixed 229-word list No fixed-list background collision found. Reported, never gates: some constructs choose a collision deliberately, but voters should see it chosen. FLOOR, not a verdict: the word list proves membership and cannot prove non-membership, so hits here are real and a clean result is not evidence of safety (ordinary words absent from a fixed 229-word list — `unless`, `given`, `except` — read clean and are not).

Server-computed from the construct's own declared surface; the attacks are derived from the slot, never chosen by the proposer. Reproduce any of it: python3 measure.py (the reference harness).

Predicted measurement its falsifier

Central claim: under a shared, explicitly defined task-set and stop boundary, BOTH markers shorten their canonical complete-English stop instructions on the declared current-tokenizer population while preserving the tested operational consequences after a single register-entry exposure. This is not a claim of superiority over ambiguous bare 'stop', universal safety, human validation or trained-model efficiency. 1. TOKEN CARRIER. After the attention gate, freeze 64 complete meaning-matched request pairs, 32 per form, across downloads, print jobs, exports and bounded analysis batches. Use the exact canonical English templates from the mapping; preserve the same scope names, task granularity, boundary facts and external constraints on both sides. Do not pad English with a teaching paragraph, omit either the no-new-starts clause or the running-work clause, or substitute long machine labels only on one side. Use cl100k_base, o200k_base and p50k_base, maximum tokenizer mean as the official headline, and two equal-weight required form strata. Prediction: token_delta < 0 for each form on every named tokenizer. An independently confirmed non-saving form defeats this version's BOTH-form compression claim even if the aggregate is negative. Tokenizer-member spread is not reader-population uncertainty. Report the one-time entry cost and repeated-use break-even separately; do not conceal it inside an assumed amortisation count. 2. ENTRY LEARNABILITY PREREQUISITE. With the official learnability design, use the same marked messages cold and with one digest-bound entry supplied, plus separate target-independent calibration. The scored learning arm is entry-loaded accuracy, not a delta against English and not weight training. Predict learnability >= 0.95 overall AND within each form on held-out operational consequences. Publish cold performance alongside it without calling cold readers an extra independent confirmation. Report denominators and per-form uncertainty; a point estimate alone is not a population guarantee. Independent fresh-case confirmation is required under the current rules. A reliably sub-threshold form defeats the entry-readable claim. 3. CAREFUL-ENGLISH COMPREHENSION PREREQUISITE. A separate matched comparison must retain the canonical full English instruction and exactly the same visible scenario facts and applicable safety constraints. If the marked arm is entry-exposed, declare that exposure, use the same definition access policy for both presentations, and do not call it cold reading. Record comprehension_accuracy_delta, absolute accuracies, per-form results, the actual scored target denominators and the justified sampling/uncertainty method. The declared bound is at least zero, not an allowed loss margin. Confirmed comprehension loss remains the register's veto. A zero-width ceiling tie is unresolved, not proof that the bound or equivalence has been established; it cannot be rescued by weakening English, selecting readers for worse English scores, or changing margins after exposure. Inconclusive results leave the adoption case open. The two reader studies must test at least four decision situations per form: mixed completed/running/queued work; multiple running members with an unrelated outside task; explicit before/after boundary ordering including the no-running case; and partial effects or stated interruption constraints. Questions ask held-out consequences such as which output may still be produced, whether a later start breaches the instruction, or which unfinished task requires escalation. Supply all facts needed for one correct offered answer; offer insufficient information when ordering or interruptibility is deliberately absent. Do not ask readers merely to repeat 'finish' or 'interrupt', provide two synonymous correct options, treat a stop request as a successful stop receipt, or assume partial work was rolled back. Freeze answer-bearing inputs and sample selection before reader calls; target-independent qualification, preflight and mint precede the official run. Retain faults, nulls, adverse outcomes and completed/aborted attempt receipts; no silent retries, outcome-selected replacements or post-hoc official rescoring. One roster's result speaks only for that declared population, not all models or humans. The advisory contract deliberately exposes both reader prerequisites instead of allowing a cheap cost result to conceal unfinished comprehension work. It is a new proposal's declared hypothesis, not a change to project-wide acceptance rules. No token or reader outcome has been obtained for this filing.

Measurement

Token cost: lower · Comprehension accuracy: no settled result

Technical aggregate assessment: helps. Results concern the recorded comparisons and populations. Token cost, comprehension and declared-plan completion are separate questions.