P4-TRANSFER Data Access Protocol

Status: DOOR PREPARED, NOT PASSED. This document is the acquisition protocol for the lab-B cohort that Door P4 — TRANSFER requires. It is named by the checklist itself (docs/EXTERNAL-DOORS-ACQUISITION-CHECKLIST.md:85-88, "the <TASK>-DATA-ACCESS-PROTOCOL.md").

What this document does NOT claim. It relabels no gate (X10 stays NOT_ESTABLISHED, H-AIF-G8 stays NOT_LOCATED, G08 stays BLOCKED_EXTERNAL). It registers no cohort. No lab-B data has been requested, received, or read as of 2026-08-30. The lab-A (Wadhwa 2022) holdout is SPENT (duration spent by B3; mark channel burned, D5) and nothing in this protocol touches it. Adverse first: the single hard external blocker on this door is that no independent second stator-dwell dataset exists in the repo (docs/EXTERNAL-DOORS-ACQUISITION-CHECKLIST.md:80-83), and this document cannot remove that blocker — only a real data release can.

Frozen counterpart: experiments/predictions/P4-TRANSFER-PREDICTION-RECORD.v1.json (the declarations this protocol feeds). Acceptance criterion, verbatim (experiments/cross-study-preregistration.v1.json:107-109):

"At least one mechanistic parameterization frozen on one laboratory/study predicts a second laboratory's commensurate raw observations with a predeclared advantage over baselines. Source-paper predictions or unit-incompatible mappings do not count."

1. What is requested from each candidate lab

The same five things from every lab, because the frozen scoring rule needs all five (docs/EXTERNAL-DOORS-ACQUISITION-CHECKLIST.md:41-56, 73-78):

1. Dwell state — integer stator occupancy (or the lab's recorded state variable, with its observation operator declared so the units axis can be assessed honestly).
2. Dwell duration — per event, in seconds, with the censoring flag (rightCensored).
3. Transition target — the next-state record. NOTE: the mark channel is under D5/D6 quarantine in-repo; transition targets are requested so the cohort is complete and future mark-process work is possible, but the P4 duration-only claim never reads them.
4. Motor identities — a stable per-motor identifier on every event. Without motor identities the cohort is unusable: the experimental unit is the MOTOR, and frames or events are never independent replicates (CLAUDE.md truth contract; score.py motor-equal aggregation).
5. Calibration release — the instrument calibration published with the raw data, so OBSERVED is earned (docs/EXTERNAL-DOORS-ACQUISITION-CHECKLIST.md:51).

Candidate labs and their standing caveats (from the checklist, :73-78):

2. Licence terms that are acceptable

• The licence must permit: (a) recording a committed SHA-256 of the received archive in this repository, and (b) publishing derived scoring artifacts (per-motor NLPD arrays, contrasts, verdicts) computed from the data.
• The licence need not permit redistribution of the raw data. The repo commits hashes and derived scores, not the bytes. A no-redistribution licence is acceptable; a no-derived-publication licence is not.
• A licence that forbids naming the source lab in the verdict is not acceptable: provenance is part of the claim.

3. Who signs

The operator signs. Agents draft, never sign. Data-access requests, licence acceptances, and the per-axis commensurability sign-off are operator actions (the transcription-not-judgement seam: agents may prepare every clerical artifact — the request letter, the axis-by-axis evidence table with locators, the hash manifest — but the judgement that a cohort is commensurate, and any outbound commitment to an external party, carries the operator's signature).

4. Intake steps (in order; the order is the protocol)

1. Verify the prerequisite commits exist. The prediction record (experiments/predictions/P4-TRANSFER-PREDICTION-RECORD.v1.json) and the committed real-data power-atlas derivation of N (motor-count target; the existing hierarchical-aif/results/motor_stack_aif/power_atlas.json is SYNTHETIC by its own truthLabel and does not qualify) must both be committed before any lab-B field is read. Prospectivity is decided by the commit graph (docs/EXTERNAL-DOORS-ACQUISITION-CHECKLIST.md:59-61).
2. Hash before reading. On receipt, compute the archive SHA-256 and register it in labBRegistration.cohortSha256 of the prediction record — before any field of the data is read. The registering commit must be a strict descendant of the commits in step 1.
3. Assess commensurability from released metadata only (papers, methods sections, calibration documents — not the event data). Produce the six-axis evidence table (agent draft, verbatim quotes with locators). Operator signs each axis or refuses.
4. Refusal is a first-class outcome. A cohort failing any axis is registered in the prediction record as REFUSED-INCOMMENSURATE with the failing axis named, and is never coerced — forcing a common coefficient across incompatible assays is forbidden, and is the exact axis on which the current corpus already fails (cross-study-parity-report.json:653; checklist :52-55).
5. Quarantine on ingest, per the existing tested machinery. Ingest validation reuses the built range-check and quarantine policy, not new code:
   • hierarchical-aif/src/motor_stack_aif/marks.py:63-92 — prepare_mark_dataset(policy="quarantine") keeps/quarantines with quarantineReason per row and never silently drops;
   • hierarchical-aif/src/motor_stack_aif/events.py:21-83 — the mark_quarantine mode enforces D5/D6 at the type level;
   • tested by hierarchical-aif/tests/motor_stack_aif/test_nextstate_range_check.py:60-66 (quarantine preserves, labels, and conserves counts: kept + quarantined == all). The duration channel of a defective row is quarantined with the row, never repaired by hand.
6. Score once. Only after registration is complete does experiments/predictions/p4_transfer_harness.py run — it refuses unregistered hashes and refuses a still-empty registration by construction. One scoring run per registered sha256; re-execution is for determinism checking only and must be byte-identical.
7. Report whatever comes out. All three falsifiers are legitimate reportable negatives; none may be tuned away. An underpowered cohort (below the derived N) is scored anyway and reported UNDERPOWERED at its actual interval width.

5. What would end this protocol

Receipt and registration of one commensurate cohort ends the acquisition phase; the scoring run ends the door attempt, in either direction. A refusal on every candidate is also an honest terminal state: it would leave X10 NOT_ESTABLISHED with the reason named per candidate, which is a better record than a coerced pass.