source_line	date	decision	by	source
75	2026-10-11T15:51Z	All seven changes the B1e interview asked for (blind reads, logged width, gate check, jq/awk and tokenizer, rule 0 also-present, concept-level placements, provisional gates with member recheck), audited before they are given to the analyst.	J	J chat 2026-10-11 17:50 Berlin: "Do all changes and audit them before giving them"
96	2026-10-12T21:40 Berlin	B4 selection by three placements (two readers + B1 grouping), classes 1-5 per orchestrator note 2026-10-12T19-00Z; class 2 keeps the B1 grouping (readers' list first), tagged b1_grouping_kept; blind sample 500 = 250 uniform + 250 adjudicated, two fresh readers each, strata apart	J 12 Oct (relayed by orchestrator note; J pasted it 21:01).	tools/assemble.py 91aa6410 (new), stage_B3 b6fbd8e7, stage_B4 2323fc7d, STAGE-B4 9d21b505. Dry run on the run: classes 2549/1977/2432/828/58; 56 missing-replicate items to class 3; 18 unopened-family flags.
155	2026-10-15 23:33 Berlin	Codex D2 labelers: lead checkpoints from the external drive ten rooms at a time, by size; resume each lead with its recorded GPT 5.6 medium, then switch the same session to GPT 6.1 medium for the return task (replaces HANDLER.md's recorded-model rule for the return only); run.json records model (6.1) and resumed_with (5.6); returns to 00_GIVE-THE-AGENT/DROPS/D/<sub>/. Brief: handler/memory/analysis/2026-10-15-to-codex-D2-labelers.md.	J	J 23:33
248	2026-10-21 17:18 Berlin	Re-check labeler model corrected to GPT-6.1 Sol medium (OpenAI note relayed by J); handler prompt had 5.6 in error; D2 run.json resumed_with field flagged as unreliable.