# LEDGER EXTRACT — openhands-index survey # # Append-only custodian ledger. Every stamped line is reproduced verbatim EXCEPT the # path column, which is reduced to a basename: the full paths were on the operator's # machine and say nothing a reader can use. The sha256 values are untouched, so each # line still verifies against the corresponding file in this directory. # # Lines beginning '#' in the original are custodian notes and rulings, kept as written # except that the custodian is named by role rather than by name. # including the ones that record our own process failures. # --------------------------------------------------------------------------- # PIN A2 (Claude arms) model=claude-fable-5 CLI=2.1.193 recorded=2026-07-25T19:54:16Z (before any arm launched) # PIN checker (post-lock judgment) model=claude-fable-5 — SAME as executor A2; shared-blind-spot risk disclosed in ANALYSIS_PLAN_iter2.md §5.1 # PIN B2 (Gemini arms) model=PENDING — set at RUNBOOK §2A after /model, with billing smoke test # NOTE A2-id: mid-arm prereg stamp WAS NOT TAKEN (custodian process deviation, 2026-07-25). # Lock-time stamps below fix content from lock forward only. The L1 witness for this # arm therefore rests on filesystem ordering + continuous human approval of every tool # call, NOT on a declaration-time hash. Recorded, not worked around (L9). 2026-07-25T21:03:31Z A2-id lock 8527843c2e6e3c90cc435a9c4a61d987d14c9b1c3ad4a1e59f739557b2577f86 16722 PREREG_A2-id_2026-07-25.md 2026-07-25T21:03:31Z A2-id lock 5bd3cead90c35a7cc67e95b510c88dc208832c5955aade09a84cc9abe8eaf040 2757 PREDRAFT_A2-id_both-directions.md 2026-07-25T21:03:31Z A2-id lock fa2f18db77a2db94c7014ce346196838cb171b1c140701d966427f8265486602 2400668 RESULTS_A2-id_2026-07-25.txt 2026-07-25T21:03:31Z A2-id lock a22e7f2c991b9aa26b44ad919a82033f9b6caf848715cba65f57d81b48311db3 39417 REPORT_A2-id_2026-07-25.md 2026-07-25T21:03:31Z A2-id lock 7a1688bb054c3fa0de029fe6d0d4162dcc37cad0b11a83b1432e697700c7af1e 7400 RECON_A2-id_2026-07-25.md 2026-07-25T21:03:31Z A2-id lock 3d27cf765604dd96e0b6833a7d35014720290c2e37d22adf9ab3d2b2727209dd 2107 LOCK_A2-id.md # PIN iteration-2 RESTART (run 1 archived): all arms model=claude-fable-5 CLI=2.1.220 bypassPermissions recorded=2026-07-25T22:35:31Z (before any restart arm ran) 2026-07-25T23:12:48Z A2-id lock 8b5818428faf71d1a6c0256a1e517adf3f6d95f92c17e6dcd066331e367c0eb8 16352 PREREG_A2-id_2026-07-25.md 2026-07-25T23:12:48Z A2-id lock c4b5609581e88f1b37af62d5b3d030c095214df092af6d3954c30d58e5f8c8fe 2770 PREDRAFT_A2-id_both-directions.md 2026-07-25T23:12:48Z A2-id lock 38ce432250dec5dccf16d00e0290c9e59bfd6653c5b6066091947f11f7a2b354 65951 RESULTS_A2-id_2026-07-25.txt 2026-07-25T23:12:48Z A2-id lock 9e44b2b187ca2a9fbace66ef20e5defb02aa26e969cd8882c450a4d726891285 36288 REPORT_A2-id_2026-07-25.md 2026-07-25T23:12:48Z A2-id lock ef1b39276c13eae32eca29b4ac118664e582aec22b1dc07f6e17278f13db902f 6937 RECON_A2-id_2026-07-25.md 2026-07-26T00:15:52Z CALIB prereg 153d4b9aa8df5d6b930e3b9172b510aff10138a9f28ff292dc574c9a79f456f3 277 CALIBRATION_PREREG.json 2026-07-26T00:18:44Z CALIB lock 6f6ebf6aa853bc5ebd21587c26f2770776adf210e4528224d2939cbb5000f7bc 314675 base_A.json 2026-07-26T00:18:44Z CALIB lock 4840a810430e9b3375c9e87ed81215e207f5c5ce76d3acf9aed1fe4c6b0dd243 314542 base_B.json 2026-07-26T00:18:44Z CALIB lock 360473422333335d33d40033184b817d26bdf001d7480e7ce0f6c611895cada5 314676 pert_A.json 2026-07-26T00:18:44Z CALIB lock 8c051a7189db6554e2254e48e693b79f24dc6f6226ec2c7d10e4c53482244c4c 314542 pert_B.json # CALIB addendum 2026-07-26T00:24:59Z: sweep-flagged cell (issue_resolution_score, OpenHands/Gemini-3.1-Pro, # pub 76.8 vs recomp 77.0) CHECKED AND DISMISSED as divergent — both frozen implementations # classify it reconciled_T2 (|0.2| <= preregistered T2=0.5); it is 1 of 14 T2 score cells, # not an outlier. The custodian sweep was a scouting pass; the arms preregistered tolerances # are the authority. No publication-rule flag attaches to this cell. # RULING 2026-07-26T09:22:23Z (custodian): R2 = CLOSE NOW. B2-id and blinded pair NOT EXECUTED, stated in full # per L9 in every published surface. Replication seat left deliberately open — the O9 repro # path invites OpenHands (or any reader) to be the second executor on independent hardware; # community-setting conversation starter, not a gap. B2-id retained as optional named # follow-up; blinded pair deferred to the parked kit-hardening WO where its value lives. # RULING 2026-07-26T09:25:25Z (custodian): R1 = A2-id ACCEPTED as the study identified-arm result (O14 # acceptance is the custodian's; the arm's own O14 line stands as written in its REPORT). 2026-07-26T20:59:18Z CALIB-COST prereg 1dc1994490143c6652ac8712321be3a8c9bc097eb285fbc13ce9000b34bec142 428 CALIBRATION_COST_PREREG.json 2026-07-27T01:13:06Z ADDENDUM lock 884f32bad3f63361db3941b414dd9b64c82f090836e783f590a9752614e43bf2 5185 ADDENDUM_2026-07-26_two-surfaces.md 2026-07-27T01:13:06Z ADDENDUM lock 4779a4c7f0c4063655c37fdd65b1f15b8e8f20f544a64bdfb98753c1da09821e 23911 extract_instance_results_3015ac6.py # NOTE 2026-07-27T01:13:06Z: two surfaces examined AFTER the arm locked and after publication — # the HF Space (missed by R3) and extract_instance_results.py (an unexecuted leg). # Recorded in ADDENDUM_2026-07-26_two-surfaces.md. The arm census is unchanged; the # cost columns are disclosed as non-independent (per-instance costs are normalised # toward the published cost_per_instance by the extractor). 2026-07-27T10:39:48Z CALIB-COST lock 6f6ebf6aa853bc5ebd21587c26f2770776adf210e4528224d2939cbb5000f7bc 314675 base_A.json 2026-07-27T10:39:48Z CALIB-COST lock 4840a810430e9b3375c9e87ed81215e207f5c5ce76d3acf9aed1fe4c6b0dd243 314542 base_B.json 2026-07-27T10:39:48Z CALIB-COST lock e58464ef3ca7e54fa76d1438cfba4f0be0c0dee30c063097a9f3245f59feddf4 314659 pert_A.json 2026-07-27T10:39:48Z CALIB-COST lock 6667a74f8a7450a901497a2c9ce02014fb2f5971dbdf64ac1f6b27b781adc660 314525 pert_B.json # NOTE 2026-07-28T01:00:37Z: the publisher's extract_instance_results.py was hashed above and briefly # mirrored into the published annex. The mirror is WITHDRAWN: their repository declared no # licence in its API metadata while its README named MIT (recorded in RECON), so # redistributing it asserted a permission we cannot point to (L7). It is now cited by URL # and by the sha256 already stamped above. The stamp line stands — it records what we read.