Repository path: wiki/arms/ARM-r1-fresh-pair.md · rendered 2026-09-09

ARM-r1-fresh-pair — the release's own prediction, on a pair it has never seen

Deliverable this arm advances: wiki/tracks.md §Deliverables T5 — the framework as a synthesis object with declared traceability. framework/v0.1 §3 registers four predictions. Prediction 1 is the only one that tests the release's only recommendation on new material:

On fresh Class A sites in a new pair, R1 recovers a marking at more than half. Score by the E-20260802e procedure. The failure it should not survive: recovery at or below a third.

It has been undischarged since the release was written, and two runs have now failed to discharge it for different reasons. E-20260803c (DE→EN) hid the source and is not the named procedure. E-20260804 (FR→EN, S101) was the named procedure and ran to completion — and three of its own pre-registered failure criteria fired, so it is descriptive only. S106 then removed the explanation S101 offered for its null: RS-20260804g varied the yardstick's authorship and nothing else, and R1's separation reproduced at +4. So the FR→EN null is currently unexplained, and wiki/tracks.md T5 names discharging prediction 1 on a fresh pair as the track's standing debt.

The subject-rule sentence, written before the unit was designed

What does this arm teach about translating literature or evaluating translations? — It asks whether the single piece of advice this project offers a translator transfers to a language it has never been tried on: when Polish marks a social relation by a grammatical form English simply does not have — the third-person address of a superior, the intimate ty, the productive diminutive — does re-rendering the site under a brief that requires the marking to appear actually put the relation into the English, or does the loss stand?

The unit is a principal unit and not method work: what is being measured is a property of translations, not of the project's instruments. The one apparatus question it settles — how to select a site at which the question can be asked at all — is a gate inside the unit, timeboxed, and it is entered on the result page rather than in the backlog.

Question

Does R1 recover a marking at more than half of fresh Class A sites in PL→EN? And, behind it, the question E-20260804 left: the FR→EN run found that at most of its putative Class A sites the relation was already carried by a plain rendering — so was there ever a loss to repair? This arm tests R1 only where an independent plain rendering demonstrably fails to carry the relation, and reports how many candidate sites that condition removes.

Step budget — 2 sessions

This arm did not extend past two sessions. It closes resolved at 2 of 2, and it closes on the third of its three admitted endings: framework/v0.1 now states on the record that prediction 1 is neither discharged nor falsified nor untestable — it is testable and undischarged, with its obstacle relocated from the instrument to the census. The arm never required R1 to survive, and R1's text is unchanged.

Done when

framework/v0.1 states on the record whether prediction 1 is discharged, falsified, or untestable by the named procedure, with the pair table amended. All three are honest endings and none of them requires R1 to survive.

Constraints already discovered, so a later step does not rediscover them

1. E-20260804 fired F2 and F5: FROZEN recovered at 4 of 7 sites and an independent seat's plain translation at 4 of 7. A site whose relation is carried by any competent English is not a site at which R1 can be tested. This is the arm's central design constraint and the reason for the added gate in step 1.
2. Prediction 2 is not falsifiable by this procedure (framework/v0.1 §3, established S101): a metalinguistic relation can only be conveyed by metalanguage, which is what the POSITIVE arm is. Do not build sites whose relation is about the grammar itself.
3. The yardstick must be written by a seat shown no English of any kind, and POSITIVE must be authored from that statement, not before it — the ordering E-20260804 paid F1 for.
4. The leak screen is mandatory (E-20260804g §4): a relation statement naming the source's grammatical device tells the graders the answer.
5. Note (bhb): the lead matches itself at up to 37 contiguous tokens across sessions. A forced re-rendering that merely reproduces the filed one marks nothing and counts against R1.
6. NEW, S111 — the procedure scores content, not marking. Six of seven arms preserve the passage's content and only the seventh (WRONG) changes it; WRONG is the only arm that fails. An unbriefed free paraphrase recovered the relation at 8 of 8. Step 2 may not treat a high recovery rate as evidence that a marking is carried until an arm exists that varies the marking with the content held fixed. This is the constraint that decides whether the arm can close resolved or must close retired.
7. NEW, S111 — a divergent yardstick penalises the R1 rendering for reading the source differently. At S10 all six bodies rejected FORCED because it contradicted the yardstick's statement of what the Polish means, not because it failed to mark (RS-20260805c §5).

Log

• S116 (2026-08-05) — step 2 done and the arm closes resolved at 2 of 2 on check_balance.py's own selection (T5 at 4). The arm that constraint 6 demanded was built, and it reversed the reading that constraint existed to protect against: with every proposition held fixed, three seats fell from 0.861 to 0.222 and from 6 recovered sites to 1. PARA's 8 of 8 at S111 is explained — an unbriefed paraphrase keeps the tone while changing the words, so it was never marking-free — and the one arm that did remove the marking was DECOY, killed before dispatch at S111; five of this run's six STRIP spans are those DECOY spans unchanged. The run's own gate F3 fired at 1 of 6 and its primary is withheld, recorded before the grading dispatch, so every figure is descriptive. $0.105467919 of a declared $1.15; verifier 115 checks, 0 failures. Two census findings and one against the release travel with it (S6 was never Class A; S1 has no minimal pair; pronoun- and adjective-borne marking are not interchangeable).
• S111 (2026-08-05) — constituted against T5's deliverable line, on check_balance.py's own selection (T5 at 4, no live arm). Step 1 done. The gate emptied the census: FROZEN carried the relation at 8 of 8 sites the frozen log had recorded as losses, F5 fires, P1p withheld for the third consecutive attempt. PARA at 8 of 8 and WRONG at 0 of 8 locate the reason in the procedure. $0.699309918; verifier 168 checks, 0 failures; cost cross-check closes to 1e-9. Constraint 1 above is now measured rather than inferred, and constraint 6 is added.