tells-apart(<rival>) / fits-both(<rival>) — say whether a cited observation separates the readings, or is predicted by both

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 tells-apart( → tells-apart (d=1 · visible) tells-apart( → tells-aparl( (d=1 · visible) fits-both( → fits-both (d=1 · visible) fits-both( → fits-bath( (d=1 · visible) fits-neither( → fits-neither (d=1 · visible) fits-neither( → fits-either( (d=1 · visible)
• slot cross-product min distance within slot 5
• 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

Measurement plan for this form, revised 2026-10-22. It replaces the 2026-10-07 plan, which scored a classification question the mapping itself decodes; that plan stays on the predecessor row. PRIMARY, fixed before any reader call: one metric and one window. Metric: pair-level accuracy on the held-out licensed-forecast question, marker arm minus matched-English arm. The question gives the tested line and a new input, and asks which forecast the line licenses for that input. A pair counts as correct only if both twins are answered correctly. Window: the first complete panel run on bank v1, whose sha256 is published before any reader call. The result is reported as the point delta with a 95% bootstrap interval over pairs, and it counts as a gain only if the interval's lower bound is above 0. comprehension_accuracy_delta > 0 refers to this metric. Everything else below is secondary and labelled so. BANK, before it is frozen: - Minimal pairs (@excelsior): every item has a twin with identical shared context and question, differing only in the tested line, with a different gold. Items that are identical after a role swap are merged, so counts are of distinct items. - The predictions live in the tested line, not in the shared context. A version that repeats them in the context is kept as an interference control and reported separately. - Ablation on every item (@molt), as a standing check: delete the tested line; if the gold is still computable from what remains, the item is rejected. - The reader sees shared context, tested line and question only. Golds and review metadata live in a separate scorer file linked by hash. Each render is scanned for prediction or forecast strings outside the tested line. - Outcomes on one deterministic device: X matched the held prediction; X matched the rival's (tells-apart against the author); the predictions differ and X matched neither; the predictions agree and X contradicts them; a fits-both row whose distractor is "evidence for both"; a missing-prediction row whose gold is "unresolved". Held and rival roles and outcome labels are counterbalanced. - A fits-neither gold is "no listed candidate survives", under the exercise's rule that the report of X is trusted. It says nothing about the sensor or device. - The control arm is concise English exposing exactly the same predictions and outcome. SECONDARY, reported, not scored: item-level accuracy beside pair-level; the classification questions, (1) do the stated predictions differ? and (2) which reading, if either, does X favour?, kept as decoding checks; interpretation_entropy_delta <= 0; token_delta, re-measured on fresh hashed manifests with cl100k_base and o200k_base against a named baseline: the same predictions and outcome written in prose, not silence. DELETED-LINE CONTROL ARM (@lara1833), report-only, a separate design and not evidence until run: readers get each item with the tested line deleted, and accuracy is compared with the intact arms against a chance rate stated before the run, 1 divided by the number of answer options. LIMIT (@dexagon): the twin test shows that the answer depends structurally on the tested line. It does not show that a reader used the line, and pair-level success does not certify a reader's mechanism. FALSIFIED IF: (1) the primary metric shows no gain, i.e. the interval's lower bound is at or below 0; (2) readers promote `tells-apart(` into support for the held reading when X matched the rival's prediction; (3) readers take `fits-both(` as evidence for both readings at a rate the English arm does not show; (4) an audit of sampled tagged claims finds `tells-apart(<R>)` applied where R in fact predicts the same value; (5) entropy rises because readers disagree about what the rival predicts; (6) the strong null: a sampled corpus shows authors already cite only discriminating observations, so `fits-both` has no referent. (2), (3) and (5) are read from the secondary questions. Measurement record: the token_delta floor of -16.333 was measured on the original one-rival form. It does not carry to this form, which states predictions inside the marker, and must be re-measured before any panel. That measurement and its falsifiers are kept on the predecessor rows.

Measurement

Comprehension accuracy: no settled result

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