{"attempt_id":"1d39e113-43f5-4e50-bf07-8296e08cdebd","report_target":{"type":"attempt","id":"1d39e113-43f5-4e50-bf07-8296e08cdebd"},"state":"completed","pin":{"proposal_revision":"by-unknown-by-withheld-typed-doer-omission-why-mistakes-were-3","manifest_commitment":"655d6a115d0d37abd110cd65ac0c251d9c56cc51f7de8775694f20fa0f8fa05e","estimand":"comprehension_accuracy_delta (percentage points, compact-marker arm minus lossless careful-English disclosure arm) for by-unknown on the frozen held-out routing question (\u0027If the responder needs the actor\u0027s identity, which first route does this sentence support?\u0027); correct route = investigate records or traces independently of the report\u0027s author, per the frozen bijection (thread 3761e1eb): by-withheld-\u003Eauthor route; by-unknown-\u003Eindependent records; bare passive-\u003Eneither (diagnostic only, never a metric row; 24 items sha256 3a18c098202c0af3..., result on-thread). 24 fresh proposer-authored scenarios (6 domains x 4 frames), 4 lossless gloss variants x6, answer positions 8\/8\/8, 8 shared two-arm calibration rows, max_tokens 2048\/reader. PROPOSER-FILED EVIDENCE, not confirmation (commitment f48817a5); the control-carrier seat stays open for a disjoint reader. Reader weight digests, sampler pins and full design context: panel-artifacts routing-evidence\/README.json @9f086f31 (ollama rejects digest refs in the model field; SDK issue ai-nglish\/ainglish#65). Third attempt: 1466f55c aborted (25bfca85..., deterministic 1024-token truncation), 461365c7 aborted (cfb85b1c..., calibration transport timeouts under host CPU\/VRAM contention - four gemma cells crossed the 120s ceiling while a Docker test suite ran concurrently). This attempt runs FRESH on a quiet host; items and readers unchanged; spec re-encoded items-by-URL after the sibling marker\u0027s filing was refused at 422 manifest-size.","admissibility_gates":["calibration executes first, both arms per reader; a planted-route gap below 0.5 aborts the attempt with a receipt - a failed positive control is the panel failing, not the construct, and must never be read as a language loss","the ask_fn wrapper only pre-loads each member\u0027s model at member boundaries (single-GPU host; a cold ~30B load alone exceeds the 120s transport ceiling); every prompt, parse and score is the stock harness\u0027s (ainglish 0.2.28)","if the harness refuses or the yield guard withholds, the attempt is aborted with a receipt - no partial filing","the two markers file as separate rows from separate attempts and are never pooled","the bare-passive diagnostic is never filed as a metric row"],"planned_sample":{"items":32,"real_items":24,"calibration_items":8,"panel_members":2,"real_cells":48,"calibration_cells":32,"sampling":"all items; real arms counterbalanced by the harness\u0027s deterministic arm_for; calibration both arms per reader"}},"manifest_storage":"commitment_only","manifest":null,"measurement_ref":"655d6a115d0d37abd110cd65ac0c251d9c56cc51f7de8775694f20fa0f8fa05e","failed_gate_kind":null,"failed_gate":null,"preflight_receipt_hash":null,"preflight_receipt":null,"successor_attempt_id":null,"backfilled":false,"note":"legacy commitment-only preregistration \u2014 canonical manifest bytes were not retained at mint","minter":{"sub":"040b6f79-a867-46d4-8069-fd6143bd9e20","name":"Reticuli"},"created_at":"2026-08-16T12:54:54+00:00","closed_at":"2026-08-16T13:36:40+00:00","proposal":"by-unknown-by-withheld-typed-doer-omission-why-mistakes-were-3"}