PROMETHEUS (PRM-001 v3) epic: Qwen 3.5 4B local lane — fast enough and good enough to break quorum ties #4354

owner: rex (lead aprender-84)

Epic for PRM-001 v3 "PROMETHEUS": make local Qwen 3.5 fast enough, and good enough, to break quorum ties. The Qwen 3.5 4B lane on released apr runs as a zero-weight shadow lane, then climbs a tripwire → tie-breaker → vote ladder. v3 supersedes REX-001, PRM-001 v1/v2 and PRA-001. Hosted-model outputs are never training targets.

One branch: rex/001; commits say Refs #4354. The row table is §7 of the spec. PRM-L lands rex/001 and the spec on main as its own PR. Budget: K̂ 2585 · K 2840 · andon 2275.

Phases (these replace APR-EPIC-001 §4d): P0 now = capture + shadow · P1 0.71 = speed instrument + ratchet · P2 0.71 = experiment + hardware ruling · P3 0.72 = tripwire · P4 0.73 = tie-breaker + prefill parity · P5 D-7 = local distill · P6 0.75 = vote.

Checklist (§7)

• PRM-00 v1 prereg (5f0ae10ac)
• PRM-00b lock v3 (rex-prereg-v2), 84 → on rex/001, lands with PRM-L (10-08)
• PRM-L land rex/001 + spec on main (own PR), 84 → 84/rex-001-wip @a685fd7dfa merged up to main ad8fafc, guard reds fixed; PR (PMAT-4524, PRM C6+C11: rex pool-admit backfill + FALSIFY-TAS-009 (land rex/001 @292c01d75c) #4524) waits on the open-PR cap (10-08)
• PRM-01 gx10 slice (split infra#1088), 84 drafts → gx10 slice done: infra#1132 landed 09-26 (infra#1142, feat(aprender-core): dataset-thestack-python-v1 INV-DATA-002 PARTIAL_ALGORITHM_LEVEL #1154, feat(ship-009): FALSIFY-SHIP-009 DISCHARGED via apr stamp local fixture-swap (MODEL-1 PARTIAL → DISCHARGED) #1054), closed 10-08. PRM-07 still needs a paiml-implement release carrying cbtop: BrickScore output hardcoded — 5 bugs corrupt profiling JSON #420 + feat: implement lazy transpose v2 #441 (fix: implement lazy transpose to resolve slow Tensor::transpose #436); newest tag v1.8.1 (09-24) predates both
• PRM-C1 capture at dispatch, transferred → paiml-implement#436
• PRM-C2 review-ledger-v2 (1c0529dca)
• PRM-C3–C5 almacen store, second copy, transcript retention (infra)
• PRM-07 shadow activation (gx10 slice + pi release)
• PRM-S1 replay benchmark review-replay-v1 vs llama.cpp, fb
• PRM-S2 workload profile, owner placed by 84
• PRM-10 perf ratchet wired to replay
• PRM-02 / 03 corpus v1, harness + scorer
• PRM-04 cell admission
• PRM-C6 secret gate GREEN (f9d93cbed); backfill waits on C1/C3
• PRM-C7–C9 contamination (C7 green 6ffeaeb26, cb), dedup, time split (a2)
• PRM-C10 outcome join + HRQ gold, a2
• PRM-C13 G-PROV gate + κ_err probe, cb
• PRM-C14 datacard + weekly receipt, 88
• PRM-05 pilot (parked)
• PRM-06 full run (parked)
• PRM-08 analysis + hardware ruling (parked)
• PRM-09 ladder + tie-breaker rung
• PRM-11 Q1–Q3 loop
• PRM-C11 sparse-logits-v1 (local only)
• PRM-12 Q4/Q5 local distill
• PRM-C12 terms tagging: removed in v3, G-PROV folds into C13

Parked

Closed over budget (APR-EPIC-001 rule 5). Reopen = pull (rule 7).

• REX-05: pilot: 10 dev items per admitted cell, cold-start runs, H6 reference workload alone x5 #4360 REX-05: pilot: 10 dev items per admitted cell, cold-start runs, H6 reference wor — REX-05..12 blocked on infra#1088 (spec §4a c10)
• REX-06: full run: test split on admitted cells, 9B control, Haiku+agy baselines, 10% determinism re-run, H6 co-located runs #4361 REX-06: full run: test split on admitted cells, 9B control, Haiku+agy baselines, — REX-05..12 blocked on infra#1088 (spec §4a c10)
• REX-07: day-0 shadow lane: 4B lane as non-voting 4th quorum lane on gx10, writes review-ledger-v1 #4362 REX-07: day-0 shadow lane: 4B lane as non-voting 4th quorum lane on gx10, writes — REX-05..12 blocked on infra#1088 (spec §4a c10)
• REX-08: analysis + hardware ruling: §3 tests, §4 rule, rex-001-hardware-ruling.md, per-divergence/per-gap issues #4363 REX-08: analysis + hardware ruling: §3 tests, §4 rule, rex-001-hardware-ruling.m — REX-05..12 blocked on infra#1088 (spec §4a c10)
• REX-09: promotion ladder: H4/H5 shadow->tripwire->vote gates in paiml-implement and arbiter decide (filed issues) #4364 REX-09: promotion ladder: H4/H5 shadow->tripwire->vote gates in paiml-implement — REX-05..12 blocked on infra#1088 (spec §4a c10)
• REX-10: perf ratchet: per-tag §5.1 job on primary cell, ratchet file, knob sweep per train #4365 REX-10: perf ratchet: per-tag §5.1 job on primary cell, ratchet file, knob sweep — REX-05..12 blocked on infra#1088 (spec §4a c10)
• REX-11: B1 loop: prompt versions + retrieval few-shot evaluated via §5.4 champion/challenger #4366 REX-11: B1 loop: prompt versions + retrieval few-shot evaluated via §5.4 champio — REX-05..12 blocked on infra#1088 (spec §4a c10)
• REX-12: B2 loop: 27B teacher logits (gx10, top-k, m=1) over train pool; B2a/b/c acceptance tests wired to qwen35 finetune/distill/merge #4367 REX-12: B2 loop: 27B teacher logits (gx10, top-k, m=1) over train pool; B2a/b/c — REX-05..12 blocked on infra#1088 (spec §4a c10)

Rows (folded as checklist lines, ISSUE-TREE-001, aprender-wtix-even)

• REX-00: pre-register: commit spec, analysis plan (§3 tests as code signatures), empty test-hash manifest, prompt v1 + sha #4355 REX-00: pre-register: commit spec, analysis plan (§3 tests as code signatures), empty test-hash manifest, prompt v1 + sha
• REX-02: corpus: P/R/G review corpus, strata, dev/test splits, sealed hash manifest, contamination contract with fixtures #4357 REX-02: corpus: P/R/G review corpus, strata, dev/test splits, sealed hash manifest, contamination contract with fixtures
• REX-03: harness: Rust bench binary driving resident apr serve + deterministic verdict parser + receipt writer + scorer (Wilson, McNemar exact, seeded bootstrap, Holm) #4358 REX-03: harness: Rust bench binary driving resident apr serve + deterministic verdict parser + receipt writer + scorer (Wilson, McNemar exact, seeded bootstrap, Holm)
• REX-04: parity admission: apr parity per cell vs llama.cpp d1d3c3396 on exact weights #4359 REX-04: parity admission: apr parity per cell vs llama.cpp d1d3c3396 on exact weights
• REX-05: pilot: 10 dev items per admitted cell, cold-start runs, H6 reference workload alone x5 #4360 REX-05: pilot: 10 dev items per admitted cell, cold-start runs, H6 reference workload alone x5
• REX-06: full run: test split on admitted cells, 9B control, Haiku+agy baselines, 10% determinism re-run, H6 co-located runs #4361 REX-06: full run: test split on admitted cells, 9B control, Haiku+agy baselines, 10% determinism re-run, H6 co-located runs
• REX-07: day-0 shadow lane: 4B lane as non-voting 4th quorum lane on gx10, writes review-ledger-v1 #4362 REX-07: day-0 shadow lane: 4B lane as non-voting 4th quorum lane on gx10, writes review-ledger-v1
• REX-08: analysis + hardware ruling: §3 tests, §4 rule, rex-001-hardware-ruling.md, per-divergence/per-gap issues #4363 REX-08: analysis + hardware ruling: §3 tests, §4 rule, rex-001-hardware-ruling.md, per-divergence/per-gap issues
• REX-09: promotion ladder: H4/H5 shadow->tripwire->vote gates in paiml-implement and arbiter decide (filed issues) #4364 REX-09: promotion ladder: H4/H5 shadow->tripwire->vote gates in paiml-implement and arbiter decide (filed issues)
• REX-10: perf ratchet: per-tag §5.1 job on primary cell, ratchet file, knob sweep per train #4365 REX-10: perf ratchet: per-tag §5.1 job on primary cell, ratchet file, knob sweep per train
• REX-11: B1 loop: prompt versions + retrieval few-shot evaluated via §5.4 champion/challenger #4366 REX-11: B1 loop: prompt versions + retrieval few-shot evaluated via §5.4 champion/challenger
• REX-12: B2 loop: 27B teacher logits (gx10, top-k, m=1) over train pool; B2a/b/c acceptance tests wired to qwen35 finetune/distill/merge #4367 REX-12: B2 loop: 27B teacher logits (gx10, top-k, m=1) over train pool; B2a/b/c acceptance tests wired to qwen35 finetune/distill/merge