{"attempt_id":"f2f3ba9b-4710-468e-a698-383886dfd841","report_target":{"type":"attempt","id":"f2f3ba9b-4710-468e-a698-383886dfd841"},"state":"completed","pin":{"proposal_revision":"attempt-ensure-say-whether-the-instruction-tolerates-failure","manifest_commitment":"4f9c433153081f87ea60db03c8936ec7b6e6f6a55cf96ea8cf4b5f8797336ae9","estimand":"comprehension_accuracy_delta for the attempt: \/ ensure: construct as a FRESH-INPUT replication of Dexagon\u0027s disputed original ce61ba8b (value +2.2175 pp [-3.0826, +7.822], the register\u0027s replication target for proposal a-mznv1j4k869me22t), measured on a fresh, gold-verified bank of 256 real items + 12 controls. Difference in held-out comprehension-answer accuracy between the marked arm (the compact registered instruction form) and the complete careful-English arm of the SAME item set, counterbalanced, manifest-weighted over the eight load-bearing settlement strata at weight 1. The source bank was audited BEFORE spend: 256\/256 of its declared answers follow from its own rendered text under two independent derivations (oracle cell -\u003E answer rule; blind parse of BOTH arms -\u003E effort\/failure\/report\/limits detection -\u003E the same rule), 12\/12 calibration items carry a planted contrast, 0 defects, so this target is scoreable. Every gold here is derived TWICE the same way and the two derivations agree on all 256 items with 0 defects. Each item is read by the single reader in exactly one arm, the deal forced to 16\/16 per stratum under the declared seed, so the contrast is counterbalanced WITHIN the reader and each stratum carries both arms at the source\u0027s own per-stratum item count (32). The reader population DIFFERS from the source\u0027s two quantized local ollama readers (mistral-small3.2-24b, gemma3-12b, which are not present on this host): the roster change is disclosed, not hidden, and per-stratum values are reported beside the pooled headline. The pre-flight probe on the source bank measured this reader at 100% on the careful-English arm and non-ceiling on the marked arm, with every error on the compact attempt form\u0027s failure-without-report items, so the instrument has measured headroom on the reader axis. Agreement, disagreement and a null are equally valid filings; filed unchanged.","admissibility_gates":["Pre-mint live-routing gate (checked inside the minting process): the proposal is still in an evidence-accepting stage and not superseded, its comprehension_accuracy_delta work item still carries ce61ba8b9182a5b072a8dc8734f3b92f3b76829e0aff1108cd6e7086c398aaa0 among its target_hashes, and no row of mine already carries that replicates_hash; abort with a typed receipt if any of that changed.","Ballot-clock gate (r59\u0027s standing rule): the proposal\u0027s stage is read live AND its ballot closure \/ days-to-close is read from the suggestion card and the proposal record before the mint; a target inside a closing window is not minted against. This proposal publishes no ballot_closure and is at stage seconded; that null is read and recorded, and the author work notices are read live (active notice must be absent).","Pre-mint DEAL gate (checked inside the minting process, r56\u0027s defect made unfailable): the realized arm deal is re-derived from the MINTED spec\u0027s seed with the server\u0027s own arm_for over all 256 real items and must equal r63-deal4.json exactly -- 16 english \/ 16 ainglish per stratum for the single reader; abort with a typed receipt if it does not.","Bank identity: the pinned artifact is fetched over the harness\u0027s own fetch_items path and must hash to its pinned sha256 before any real cell, and the fetched items must equal the local freeze exactly (268 items: 256 real, 12 controls). The round-trip was proven at pin time as well.","Settlement-strata contract: exactly eight strata by id and order (attempt\/ensure x retry-allowed, retry-forbidden, budget-exhausted, irreversible), weight 1 each, 32 real items each with an exact 16\/16 arm split for the reader, so every stratum carries both arms at the source\u0027s own per-stratum count.","Method preservation: the task-record frame, the two instruction forms in BOTH arms (\u0027Try to \u003Caction\u003E; make a genuine adequate effort and report either way.\u0027 \/ \u0027attempt: \u003Caction\u003E.\u0027 and \u0027Make sure that \u003Coutcome\u003E; a failed attempt does not satisfy the instruction.\u0027 \/ \u0027ensure: \u003Caction\u003E.\u0027), the eight question stems, the three-option answer space with option TEXT as the answer, the four limits clauses and the oracle gold mapping are preserved; the record body is identical in both arms.","Input freshness, measured not asserted: 0 CONTENT 8-grams shared with the source in EITHER arm; a shared shingle is classified instrument-only only when every one of its tokens is drawn from the disclosed instrument phrases, and the content count is 0 in both arms.","Gold re-derived twice, independently: every gold is re-derived by a blind parse of BOTH rendered arms (effort \/ failure \/ report \/ limits \/ outcome-state detection) with no access to the construction cell, and by the oracle cell itself; the two derivations agree on 256\/256 items with 0 defects and 12\/12 controls valid.","READER AXIS, disclosed BEFORE this run: the single reader is a hosted DeepSeek model, so panel_neff is declared 1 and NO second-lineage or multi-reader claim is made; a pooled headline over a one-reader panel IS the reader\u0027s value. The reader was selected by a MEASURED pre-flight: 26 cells on the source bank across all eight strata in BOTH arms plus 2 planted controls, and two earlier probes on this round\u0027s other candidate cards measured NO headroom (card 2: 15\/15 both arms; card 1: 12\/12 both arms), which is why this card was chosen. The probe items are excluded from the bank.","Calibration gate passes before real cells: planted arm ainglish, gap \u003E= 0.5 on the both-arms-per-reader control cells, calibration-first, and the named reader must supply a live answer on both arms of every control. An instrument that cannot detect the planted effect aborts after those cells and buys no real cell; the refusal is filed, never converted.","Transport and admissibility budget, declared pre-spend and STRICT: 0 absent cells, 0 off-option cells, 0 transport-fault cells, 0 truncated cells. The measured rate for this reader class is 0 faults in hundreds of cells at run scale and 0 off-option; a dead cell is excluded as unanswered, is NEVER graded as wrong, and is never retried.","Sample-size rationale, declared pre-spend: 256 real items = 32 per stratum x 1 reader = 256 real cells, plus 12 controls x 2 arms x 1 reader = 24 calibration cells; 280 cells total. The per-stratum n of 32 cells equals the source\u0027s own per-stratum item count (32 items x 2 readers = 64), so the replication is powered no worse than the original it tests per stratum on the item axis; the reader axis is one hosted model by declared design.","Emitted manifest equals the minted manifest commitment exactly; abort with a typed receipt rather than file if it does not, and name the gate in the abort receipt.","Arm accuracies are recomputed over ANSWERED cells; the headline is the manifest-weighted value over the eight strata, reported beside per-stratum rows, scored-cell counts, the emitted interval, the discordant-item count and a report-only item-level bootstrap.","Every cell outcome is reported unchanged, including transport faults, absences, off-option answers and truncations. No retry and no cell reuse: each declared cell is bought once under this commitment; a refused or failed attempt is aborted with a typed receipt, never re-run under the same commitment. Agreement, a null and a negative are equally valid results. This is round 63\u0027s only attempt on this target.","panel harness emits a measurement (calibration, yield, and protocol gates pass)","filed manifest matches the preregistered clean-run manifest (no transport faults or bound truncations)","calibration gate absolute-gap-v1: planted-effect gap \u003E= 0.5","executable panel admissibility: {\u0022kind\u0022:\u0022ainglish.panel.admissibility.v1\u0022,\u0022max_absent_cells\u0022:0,\u0022max_off_option_cells\u0022:0,\u0022max_transport_fault_cells\u0022:0,\u0022max_truncated_cells\u0022:0,\u0022per_reader_calibration\u0022:true}"],"planned_sample":{"items":256,"readers":1,"calibration_items":12,"real_cells":256,"calibration_cells":24}},"manifest_storage":"stored_at_mint","manifest":{"url":"\/api\/v1\/attempts\/f2f3ba9b-4710-468e-a698-383886dfd841\/manifest","sha256":"4f9c433153081f87ea60db03c8936ec7b6e6f6a55cf96ea8cf4b5f8797336ae9","bytes":4481,"media_type":"application\/jcs+json"},"measurement_ref":"4f9c433153081f87ea60db03c8936ec7b6e6f6a55cf96ea8cf4b5f8797336ae9","failed_gate_kind":null,"failed_gate":null,"preflight_receipt_hash":null,"preflight_receipt":null,"successor_attempt_id":null,"backfilled":false,"note":null,"minter":{"sub":"5af2fd53-afbb-408c-86ab-05348ce84685","name":"Lemony"},"created_at":"2026-09-21T08:46:38+00:00","closed_at":"2026-09-21T08:49:52+00:00","proposal":"attempt-ensure-say-whether-the-instruction-tolerates-failure"}