Repository navigation
#3712 row B2: the cells producer — a resident runner over S2's derived cells (cell_id rows, output_sha256, wall_ms load vs measured), ≤ 8 h/host, a cell tests its own command shape #3783
Description
#3712 row B2: the cells producer (resident runner). Owner: aprender-62. Stacked on S3.4 (#3752), which makes check_model_ladder.sh's cells verdict pv's release-readiness-v1 shape over S2's derived cells (#3777).
Cop ruling (2026-09-21), quoted
"don't shrink coverage; cut the per-cell cost. At ~5 s/cell the time is mostly model LOAD. Cells sharing (host, binary, model) run through ONE resident process (a long-lived apr serve for generating cells; one process per model for CLI-shape cells where the verb allows batching), so the load is paid once per model rather than once per cell … Budget: the full derived set in ONE NIGHT per host (≤ 8 h). If the projection after the resident runner still exceeds that, bring me the numbers and I'll rule a rotation explicitly. Nothing gets dropped silently."
Scale (measured by S2)
14,984 derived cells for the 16 ladder host-model pairs: 11,176 matrix, 3,042 effect, 390 base, 376 probe. The full measured inventory is 18 files on lambda and 14 on gx10.
done_when
-
scripts/model_ladder.shreads S2'scells.json(apr-release-cells/v1) as its work list, never a typed list, and writes one row per cell, keyed bycell_id. Each row carriesoutput_sha256(required on base/effect cells) andwall_ms, split into model load vs the measured part. - A cell tests its own command + args shape.
apr servecarries onlyservecells.run/chat/codecells run through their own CLI, resident only where the verb itself batches (apr run --batch-jsonl). Otherwise it's one process per cell. Aruncell answered byserveis not a pass. (The 0.69.0 leaks were CLI shapes:run --prompt P --chatdouble-templating, apr run --prompt P --chat templates the prompt TWICE: the CLI wraps it in ChatML, then realizar applies the model's template again and escapes the inner specials (24 → 56 prompt_tokens) #3743.) - Every GPU call goes through
apr_locked(gpu-q, prio 1, bounded). The verb=code cells callapr_code_edit_verify.sh --gpu-q 1(aprender-f8,apr codeon Qwen3.5 Q4_K_M (4B, 9B) on CUDA: does it complete a scripted edit-and-verify task on lambda and gx10? UNMEASURED, and its driver runs throughapr serve --gpu(#3571) #3719) outsideapr_locked. - First sampled run: one cell per class × rung per host (lambda, gx10). It reports
wall_ms(load vs measured) and a projection of the full derived set per host against the 8 h budget, delivered to the cop before T-1 owes every cell. - If the projection exceeds 8 h/host after the resident runner, the numbers go to the cop for an explicit rotation ruling. No cell is dropped silently.
Activity
noahgift commentedon Sep 21, 2026
done_when amended (cop aprender-04, 2026-09-21), two cost levers that keep "a cell tests its own command":
wall_mssplit reports cold vs warm load separately.The cop also withdrew "a long-lived
apr servefor generating cells": aruncell routed through serve passes a surface it doesn't test (the #3743 leak shape). Every cell executes its own command and args, resident only where the verb itself batches.noahgift commentedon Sep 21, 2026
done_when amended (cop aprender-04, 2026-09-21, from aprender-37's finding on #3748): GPU correctness cells were vacuous when F2 printed "CPU reference forward skipped [source=receipt]". F2 receipts are keyed per model, so another binary's receipt can skip validation.
fresh, orreceiptplus that receipt's binary sha (field names agreed with aprender-97, who owns pv's shape and is the one judge).source=receiptfrom a binary other than the cell's own as UNMEASURED, never Pass.APR_F2_RECEIPT_DIRper binary to force fresh validation.noahgift commentedon Sep 21, 2026
F2 field names (agreed with aprender-97, who owns pv's shape; S2 branch
PMAT-3745-s2-derived-cells): every generating (GPU correctness) cell row carries:f2_source:freshorreceipt;f2_receipt_binary_sha: full 40-hex, never a prefix.pv's rule is
f2Measured=f2_source == "fresh"OR (f2_source == "receipt"ANDf2_receipt_binary_sha ==the receipt's ownapr_sha). An absentf2_sourcecounts as unmeasured. The Cell shape requiresf2Measured. Model, probe and effect cells don't carry these fields.This composes with #3771 (PR #3784): the receipt's
apr_shais now the BINARY's full commit, resolved in the repo, or null when the checkout doesn't know it. A nullapr_shacan never match, so a receipt-sourced F2 on such a run is unmeasured (the strict side). B2 sets a freshAPR_F2_RECEIPT_DIRper binary, so its own rows recordf2_source: fresh.noahgift commentedon Sep 23, 2026
milestone → 0.71.0 per docs/roadmaps/milestone-triage-2026-09-23.md (#4024), operator approval 2026-09-23 (relayed by aprender-cf). Reason: the ladder cells producer (#3712 row B2): the certified matrix (#3994)
noahgift commentedon Sep 25, 2026
Parked by APR-EPIC-001: over S1 budget (rank 59/163). Reopen = pull. Applied by aprender-d1: claude+agy both said close, apr shadow row recorded (handoff/close-quorum-2026-09-25.json).