Handback, stage B2

Written at stage B2, leg B2d, 2026-10-12.

What the stage delivers

• The frozen scheme. scheme_final.jsonl (197 themes, tier "main" 183 and "rare" 14, J at QB2-014) and scheme_final.md, with families.jsonl (21 families) beside them. All 496 gate pairs are confirmed against the frozen sentences in gate_checks.jsonl.
• The record of the pilot.
  • scheme_edits.jsonl (306 lines) with its tier_history.jsonl lines.
  • rechecks-B2.jsonl (1,594 records). 526 of them are the withdrawn split's and enter no figure.
  • pilot_rounds.jsonl (6 lines: rounds 1 to 3, gold, stab-1, stab-2).
  • exemplar_candidates.jsonl (68 lines, hard-set items only).
• The reports. stability-B2.md (the stability figure, 0.807) and pilot_report.md (the rounds, the gold number, the edits by theme, the author-error verdicts, the prediction scorecard, the cost).
• The return specification. RETURN-B2.md carries fleet ids, the three-concept floor for also_present, the rule that a bag with no other theme at the floor is not flagged, and "twenty themes at most".

What I did not do as the stage file says

• The packets. Rounds 2 and 3, gold and the stability rounds were not built by packets.py B2. My own builders (out/_analysis/b2_packets_r2b.py, b2_packets_fleet.py, b2_packets_fleet2.py) built them. They apply J's rulings: self-contained packets under the caps, fleet ids only, and codes_by_item in each MANIFEST entry. The shared pages still come from packets.py's scheme_pages.
• The blind window. The ten hard items of the blind re-placement had been visible to me in round tables earlier in this session. The window guarded the reads, not my memory, so the self-agreement of 0.883 is an upper bound.
• The bag split, made and withdrawn. At B2b-r2-p2 I split six bags into children. J withdrew the split. No one cuts the bags (J, 12:24 Berlin), and the records stay in out/recut/ and out/_analysis/.

What is wrong with the instrument, or unclear

Four gold items show sentences a stranger reads otherwise. The scheme is frozen, so they go to C1:

• T3-001 against T3-048 on an unattributed fake (uofmn t01, nine of nine readers against me).
• T3-004 against T3-033 on a new plan for a mark already standing (cuboulder t04).
• T3-084 against T3-002 and T3-005 when the asker is from another room (mariokart t04).
• T3-160 against T3-002 on calls to join a memorial (ireland t12).

Pairs that stayed divided. T3-004/T3-070 and T3-009/T3-073 are the only predicted confusions that divided readers more than once.

The cluster's count rule. The rule in T3-002, T3-033 and T3-038 asks a reader to sort concepts by act before counting. It is the sentence I am least sure a stranger reads as I meant.

Second placements. Readers drop most second placements. The deals with neighbours (G14) show it most: readers place T3-010 alone.

What I ask for

• The ruling on C1's handling of the four gold-item wording defects above, since the scheme is frozen.
• The bag count. J counts 841 B1 bags, and my head has said 846. It is not reconciled.
• Fleet ids in the validator. validate.py reads room ids in _fleet_raw through the fleet table. B3's returns will be in fleet ids from the start, and the check should keep reading them so.

What stage B3 should know

• Gold placements. The gold set's own placements carry the round-1 author-error: on ten items I placed a second theme no reader keeps. Gold fidelity in B3 should be read against the rechecked gold placements as well as against gold_items.jsonl. Even so, seven of the ten are not yet rechecked to one theme.
• Floor and reference figures. The stability figure, 0.807, is the floor for agreement. On gold, primary match is 0.800, fidelity 0.624 and pairwise 0.818.
• Return specification. Carry RETURN-B2.md's bag rule and the phrase "twenty themes at most" word for word, with fleet ids and codes_by_item.
• Batch cap. The B3 cap stays at 40 items with a 10-item warm-up. The packets of this stage were too short to show a first-ten against last-ten curve.
• No gold exemplars. No exemplar may come from the gold set. exemplar_candidates.jsonl holds hard-set items only.

This handback reports agreement with my own placements. It bounds whether the instrument decides, not whether it is right.