What was reviewed
[[2026-09-26-cohort-feedback-synthesis|reviews/2026-09-26]]:11-34; founder-grades.md:1-27; rubric-v0.2.html:24-56; [[2026-10-01-accepted-spec-v0|accepted-spec-v0.md]]:40-69; feedback/*.md; ledger Q1-Q10; rounds.md:638-697; escalations.md:436-488; queue.md:1-31,610-655; round-common.md:4-11; gh.
Findings
- Critic ACCEPTs don't yet predict his grade. Same iteration: critic ACCEPT, REVISE logged from two iMessages, 6 of 6, all on art, even the v0.2 pair, critic A8 Y (Q3). No accepted iteration has a human grade (Q10); Ray's proposed trust bar: 9 of 10 (accepted-spec-v0.md:42).
- Tonight's ACCEPTs vs his complaints: grown-up text and the 5's strokes fixed (rounds.md:689-692); A1 (clipping) Y; none is cut-and-paste (Q8). Open: giraffe gaps may print solid, stroke spacing at threshold, anchor nameability untested (queue.md:649-651; escalations.md:485).
- A3 is the widest gap. He failed it on 4 of 5 sheets (Q2), the critic on 1 of 12 (Q1). Inference (different sheets): the critic reads pictures generously. Iteration burn: A9, then A4, A8, B6 (Q9); A9 and A4 have brief fixes queued (queue.md:653-654), A3 none.
- Cohort loop. Last recorded kid play: batch 1, returned 10-01 (Q4), first since 09-08 (vaughn:15). The niece's journal line reached draft #382 (rounds.md:659).
- The gate is outside calibration. The gold set scores Ray against humans (accepted-spec-v0.md:41). The critic has no row on the human-graded PDFs (Q7), 7 of its 19 v0.2 rows lack per-check answers (Q5; queue.md:22), and shares the builders' model (round-common.md:4; spec:42).
- Throughput. 41 open PRs (21 draft) vs 17 on 09-26 (2026-09-26:12); none merged since 09-23 (gh). Inference: building outruns merges and grading.
Proposed queue.md changes
--- a/queue.md
+++ b/queue.md
@@ -25,0 +26,2 @@
+- [process, S, next non-game slot, before :16] CALIBRATION prep (2026-10-03 cohort review): backfill per-check answers on its Q5 rows, else mark notes-only; add eval-ledger/calibration.sql (agent vs human per iteration, v0.1 map per accepted-spec-v0.md:69), tested on non-held rows only (:25). When the homework returns, critic-grade the same PDFs: matching rubric, non-builder model (spec:42).
+- [process, XS] A3 name-it list beside the 10-03 brief-wording fix: builder lists picture names; critic names each from the image first; a miss = A3 N.
Decisions needed
- Throughput (escalations.md:444, since 09-27). Ray's reversible proposal: 2 new games per round until the homework scores the critic; freed slots to calibration. Bear: below the 09-01 majority-new-games rule (queue.md:8); no tap gain (:444).
- Your time: add one accepted Quiver-art sheet, kid-played, to the homework (finding 1)?
- The 1: you noted "multi-stroke" (founder-grades.md:20); built: one flagged stroke; HWT: no flag (escalations.md:488). Which?