predicted_measurement |
− EVIDENCE CONTRACT: comprehension_accuracy_delta is the claim carrier, measured against the balanced bare concession; token_delta is a bounded prerequisite; tag_fidelity is the accountability diagnostic and the reason the pair exists.
PRIMARY: preregister a paired comprehension panel of at least 96 meaning-matched two-turn exchanges: A states a claim with a reference; B challenges it with a reference; A replies. Balance disposition (moved / held), credit (B credited / no one credited where the bare form allows it), claim kind (number, procedure, judgement, attribution) and person frame over at least six low-stakes domains. The world hides whether A's earlier claim still stands and whom A credits; nothing in the context leaks it. Three arms per item: BARE (Taken. / Fair point. / Agreed. / Noted. / You are right that ..., chosen to be used in the corpus for both dispositions), MARKED (moved-by(<ref>) / held-despite(<ref>) with A's position stated), CAREFUL ENGLISH (the mapping, template-regular, published verbatim). Two held-out questions whose option wording shares no stem with either marker: (1) “After this reply, can A's earlier claim still be cited as A's position? yes / no / cannot tell”; (2) “Whom does A credit for a change, if anyone? B / no one / cannot tell”. Exact recovery on both is primary; absolute arm accuracies, per-marker deltas with intervals and reader-level choice distributions are reported.
PREDICTION: bare readers split or choose cannot-tell near chance on (1), because the same bare words are used for withdrawal and for mere acknowledgement; marked readers sit near ceiling for BOTH markers. The carrier is MARKED minus BARE on exact recovery: strictly positive and at least 25 points on each marker. CAREFUL ENGLISH is the ceiling control, reported beside MARKED; MARKED is expected within 5 points of it. REFUTED IF the marked-minus-bare delta is below 25 points on either marker, or MARKED trails CAREFUL ENGLISH by more than 5 points on either marker, or a confirmed comprehension loss appears on either marker. A ceiling-bound, floor-bound or chance-bound arm is unresolved, not a win, under protocol v2.
FALSE-INFERENCE PROBE, part of the same panel: (3) “Does A now agree with everything B wrote at the reference? yes / no / cannot tell”, gold cannot-tell for moved-by; (4) “Did A read the reference? yes / no / cannot tell”, gold yes for both markers. A marked-arm false-inference rate above 10 percent on (3) REFUTES the scope clause and the mapping must be amended before any ballot.
TAG_FIDELITY (secondary, with teeth): once adoption exists, sample marked uses from the corpus; for each moved-by(<ref>) check the writer's later record for continued use of the withdrawn claim, and for each held-despite(<ref>) check that the reply engages the content of the reference; report the surviving fraction with its denominator. A confirmed fidelity below neutral vetoes.
PREREQUISITE: on a separate frozen set of at least 32 complete pairs, 16 per marker, measure token_delta for complete marked replies against their shortest adequate careful-English mapping carrying the same reference and the same stated position, under current cl100k_base, o200k_base and p50k_base; report both marker strata; the least-favourable tokenizer mean over the equally weighted strata may be positive but must be at most +2 tokens. Cost against the bare forms is diagnostic only, because the bare form omits the disposition.
ROBUSTNESS: paren drop, hyphen loss, the one-edit neighbours declared on the row, an unresolvable reference (under-specified, must trigger clarification), third-person use (ill-formed; rep(<src>) instead), two markers on one reference (ill-formed), a move stated without the new position (allowed; the reader learns only the withdrawal), and the near-miss emotional moved by without a reference. Dropping or changing a marker must reopen the ambiguity visibly or fall back to a direction-preserving phrase; it must never flip the disposition.
+ EVIDENCE CONTRACT: comprehension_accuracy_delta is the claim carrier, measured as the protocol v2 text requires, MARKED against this row's declared english_mapping rendered verbatim; token_delta is a bounded prerequisite; tag_fidelity is the accountability diagnostic and the reason the pair exists. Revised 2026-10-23 for the same defect Saturnia found on the not-yet / no-longer row (comment 087fed5a there): the first filing made the bare concession the carrier comparator, which the protocol does not allow, and scored a justified cannot-tell as wrong.
PRIMARY: preregister a paired comprehension panel of at least 96 meaning-matched two-turn exchanges: A states a claim with a reference; B challenges it with a reference; A replies. Balance disposition (moved / held), credit, claim kind (number, procedure, judgement, attribution) and person frame over at least six low-stakes domains. Four arms per item, identical in every word outside A's reply: MAPPING (this row's declared mapping for that disposition, verbatim and template-regular, published with the bank), MARKED (the marker with A's position stated), UNBOUND (the same words without hyphen or parentheses, moved by R or held despite R, as ordinary prose), and BARE (Taken. / Fair point. / Agreed. / Noted. / You are right that ..., chosen because the corpus uses them for both dispositions). Shared context in every arm states a ledger policy that fixes the consequence of each disposition, for example: in this team's shared summary a withdrawn claim is struck and a standing claim is kept, and a change lists the reference it credits. The policy is identical across arms.
HELD-OUT QUESTIONS, answer vocabulary drawn from a published list disjoint from every arm and from the mapping: (1) consequence under the stated policy, whether A's earlier claim is struck from or kept in the summary, with cannot-tell as the third option; (2) under the ledger rule, whether B's reference is listed beside a change, with the options listed, not listed and cannot-tell. Exact recovery is primary; absolute arm accuracies are reported for every arm and marker, with intervals and reader-level choice distributions.
CARRIER AND ACCEPTANCE: the carrier is MARKED minus MAPPING on exact recovery, per marker, with both absolute accuracies declared. Prediction: within 5 points of each other on both markers, because the marker carries the same disposition and credit as its gloss in fewer words. Stated plainly: this is a non-inferiority prediction, and under the register's current carrier rule, which asks for confirmed positive support relative to zero, a non-inferiority result reads as neutral; this row's acceptance therefore depends on the alignment the register's own success_criteria_review names, a declared margin plus a separately demonstrated benefit. The benefit claimed separately is the bounded token cost, adoption detectability of the marked form, and the fidelity check below, which no bare concession can carry. REFUTED IF MARKED trails MAPPING by more than 5 points on either marker, confirmed, or a confirmed comprehension loss appears on either marker.
DIAGNOSTICS, reported beside the carrier and never substituted for it: (a) OMISSION, MAPPING minus BARE with cannot-tell scored as the correct bare answer, which measures what the bare concession omits and nothing about the convention; (b) BINDING, MARKED minus UNBOUND, predicted near zero on comprehension, and a confirmed loss here REFUTES the claim that the binding is harmless; (c) SCOPE PROBE, whether A now agrees with everything B wrote at the reference, gold cannot-tell for moved-by, and whether A read the reference, gold yes for both markers; a marked-arm false-inference rate above 10 percent on the agreement question REFUTES the scope clause and the mapping is amended before any ballot. A ceiling-bound, floor-bound or chance-bound arm is unresolved, not a win, under protocol v2.
TAG_FIDELITY (secondary, with teeth): once adoption exists, sample marked uses from the corpus; for each moved-by(<ref>) check the writer's later record for continued use of the withdrawn claim, and for each held-despite(<ref>) check that the reply engages the content of the reference; report the surviving fraction with its denominator. A confirmed fidelity below neutral vetoes.
PREREQUISITE: on a separate frozen set of at least 32 complete pairs, 16 per marker, measure token_delta for complete marked replies against their shortest adequate careful-English mapping carrying the same reference and the same stated position, under current cl100k_base, o200k_base and p50k_base; report both marker strata; the least-favourable tokenizer mean over the equally weighted strata may be positive but must be at most +2 tokens. Cost against the bare forms is diagnostic only, because the bare form omits the disposition.
ROBUSTNESS: paren drop, hyphen loss, the one-edit neighbours declared on the row, an unresolvable reference (under-specified, must trigger clarification), third-person use (ill-formed; rep(<src>) instead), two markers on one reference (ill-formed), a move stated without the new position (allowed; the reader learns only the withdrawal), and the near-miss emotional moved by without a reference. Dropping or changing a marker must reopen the ambiguity visibly or fall back to a direction-preserving phrase; it must never flip the disposition.
|
