round5-sv-cells — SELF-VERIFIER feedback cells (sv7 / sv7d)

2026-08-16. Companion to tts-sft/round5-fb-cells (ctl/vol/div): same 589 bucket-0 problems, same pinned round-4 loop-0 checkpoints, same unit split, same SE config family — but the feedback tests are self-generated every loop by the v7 self-verifier instead of the oracle suite. The oracle cells are the controls; together they measure, at scale, how much of feedback-SE's bucket-0 reach and densification survives when the oracle is replaced by a verifier the model builds for itself. Verifier evidence base: TTS/SELFVERIFY_DIAGNOSTIC_2026-08-15.md (+ §8b/8c) and TTS/V7_AB_RESULTS_2026-08-16.md.

cell	feedback source	env delta	runner
sv7	v7 self-verifier (per-loop)	round-4 contract (MAX_SHOWN=3)	run_shard_r5_sv.sh
sv7d	v7 self-verifier (per-loop)	+ LCB_FB_CASE_SELECT=shuffle (case diversity over the self bank)	run_shard_r5_sv.sh
sv8	v8 = v7 + loop policy	--probe-calls 2 --eff-cache + FREEZE_AFTER=2 (all-green freeze)	run_shard_r5_sv8.sh
pw7	v7 + PAIR-conditioned white-box (--wb-pair)	each wb call sees a cyclic PAIR of dissimilar all-green candidates, asks for inputs where the two DISAGREE (+1 shared-misread breaker); call count = sv7	run_shard_r5_sv.sh
pop7	v7 + POPULATION-conditioned white-box (--wb-pop)	2 wb calls, each sees ALL (≤8) picks: one hunts behavior DIVERGENCE among candidates, one hunts the SHARED bug/misread	run_shard_r5_sv.sh
sel7	v7 + test SELECTION (--emit-order failcount)	the emitted suite is ordered by how many current candidates each test fails (certified labels win ties), so the renderer's first-failures window shows the fail-many tests	run_shard_r5_sv.sh
sum7	v7 + failure-mode SUMMARY (--fb-summary)	one extra call summarizes the top fail-many tests' common failure modes; injected into next-loop feedback prompts via LCB_FB_SUMMARY (emission order unchanged)	run_shard_r5_sv.sh
hm7	Harman's full O(N) pipeline	--wb-pair --emit-order failcount --fb-summary composed	run_shard_r5_sv.sh

Harman mechanism cells (added 2026-08-16 late, after the meeting)

One cell per verifier-information mechanism from Harman's 2026-08-16 meeting (pair-conditioning, population-conditioning, test selection, failure-mode summarization) plus the full composed pipeline (hm7: 16 solutions → covering pairs → fail-many test selection → summary → next loop, at O(N) LLM calls). Design rules:

• sv7 is the shared control for every mechanism cell (same units, same pinned loop-0, byte-identical seed payloads; GENSEEDs 52007-52011 vs 52004).
• Label provenance is held constant: all cells keep the full v7 labeling stack (sample anchors + label-free probes + wb-certify). Mechanisms change WHERE discriminating inputs come from / WHICH failures are shown — never how expected outputs are decided (the exp3 lesson: candidate-derived labels poison repair).
• Call budget: pw7 = sv7 exactly; pop7 spends 2 wb calls instead of ≤8 (cheaper); sum7/hm7 add 1 summary call + one CPU fail-count pass per problem-loop; sel7 adds only the CPU pass.
• Only u00+u01 per mechanism cell (~295 problems) — small-shard fast-turnaround for the one-window handoff; extendable to u02/u03 later since the build is deterministic from the sv7 units.
• Analysis (ours, on returns): paired per-problem vs sv7 on the same units; primary = next-loop Δ official pass-fraction (partial credit) + reach; secondary = feedback coverage, poison band, conversion, within-arm silent-wrong drift; selffb/ audits carry wb_mode, fail_counts (emitted order), summary_chars; fb_audit logs summary_injected.
• Mechanism validation: 21-check CPU integration test (mocked LLM, real exec harness + real pipeline) ran all five flag sets end-to-end incl. operator injection round-trip; plus the 198-check cut verify. No live-GPU smoke — the changed paths are phase-A-only and env-gated in phase B.

sv8 (v8 loop policy), motivation all measured 2026-08-16: (a) all-green freeze — a problem with zero feedback blocks for 2 consecutive loops stops evolving, its population carried forward in every checkpoint (cc-03250 went 15/16 passing → 0/16 one loop after its bank went all-green; freeze also recovers the wasted post-success generation); (b) doubled probe call — two independent validator+input sets unioned (single-call probe sets measured high-variance across arms); (c) eff-ref caching — certification refs are synthesized once and re-validated by execution on later loops (saves 2 calls/problem/loop from loop 2 on).

sv8 validation status (2026-08-16, be aware): mechanisms are unit-tested (freeze mark/stage/merge cycle, label-free contract, 150-check cut verify) and the loop-1 cycle ran live end-to-end (stage → synthesis with probe-calls 2 → evolution on staged ck → merge → mark; eff-cache write confirmed). The freeze-FIRING path and eff-cache HIT are unit-tested but were not exercised in a live run (GPU window ended), and the 25-problem quality pre-flight vs v7 is pending. Run sv7/sv7d freely; start sv8 only after we confirm its pre-flight (or accept the reduced validation).

The v7 verifier per loop (~5-6 LLM calls/problem on the pod's own vLLM):

1. official-sample anchoring (extract + validate statement samples — carries most of the crack power on its own);
2. constraint-aware label-free probes: a generated VALIDATOR program admits weird-but-valid inputs; kept probes ship output: null and fail candidates ONLY on crash/timeout — zero label-poison by construction (measured 0/182);
3. basis-aimed white-box tests on all-green candidates, with wb-certify: an expected output survives only if two independently-prompted efficient references (sample-hard-gated) reproduce it; otherwise the input is kept label-free. (Measured: first labeled-wb arm under the 9% poison band at L1; the residual is problems where every ref family shares one misread.)

What is in a unit (vs a round5-fb-cells unit)

units_cells/r5sv_120b_{sv7,sv7d}_u{00..03}/
  seeds.jsonl            byte-identical to the ctl unit
  ck_nonsat/<unit>_loop0.json   byte-identical pinned round-4 loop-0
  selffb_seeds.jsonl     seeds minus SPJ-checker + functional problems (545 of
                         589 total) — the v7 machinery is stdin-only; excluded
                         problems run with NO feedback blocks (call-type has
                         none in ANY r5 cell; the 26 SPJ rows lose feedback vs
                         the oracle cells — a documented parity cost)
  budgets.jsonl          {id, L: 8} for the selffb seeds
  unit.env               RUN_NAME/LOOPS/CELL/GENSEED/FB_*/V2_FLAGS
  (NO fb_tests.jsonl — the oracle suite is never fed)

GENSEED 52004 (sv7) / 52005 (sv7d) — fresh per cell like r5's 52001-3 (replay-trap rule); routing.seed stays 1234 (same loop-1 parent grouping as r4 and every r5 cell).

Running

Node prep (idempotent, self-testing — includes the case-select validation AND the new label-free-probe contract test 自检 4):

bash minipatch/install.sh /path/to/tts-sft

Per unit (any order, cells independent, resume = re-run):

UNIT=<dataset>/units_cells/r5sv_120b_sv7_u00 RUN_DIR=<out>/r5sv_120b_sv7_u00 \
  bash run_shard_r5_sv.sh

The runner alternates per evolution loop: phase A (self-verifier test synthesis, writes RUN_DIR/selffb/loop<t>_tests.jsonl + persistent test_bank.jsonl + audits) then phase B (one SE loop with LCB_FB_PUBLIC pointed at phase A's file). Both phases resume. Runner FATALs unless the SV minipatch is installed (sanity 6 checks the label-free harness support and the selffb scripts directly).

Cost (estimate)

Generation ≈ an r5 oracle cell (1.1B billed/cell). Phase A adds ~5-6 short calls/problem/loop ≈ +5-10% tokens but real wall time: expect **1.5-2× the oracle-cell unit wall** (15-20 h/unit vs 9-10). 8 units ≈ **2.5B billed total**. Returns include selffb/ (banks + audits) — please send those back too; the verifier-side analysis needs them.

Analysis when returns land (ours)

Grade with the round-4 checker-aware fleet. Primary readouts, all paired against the oracle cells on the same problems: (a) stalled-500 crack count: sv7 vs ctl (how much reach survives self- verification) and vs div; (b) guard-set retention; (c) per-candidate partial credit (distribution improvement on uncracked problems — the goal-2a metric); (d) poison/coverage audits from the returned banks + fb_audit.jsonl (shown_cases works in all cells); (e) sv7d vs sv7 = case-diversity's marginal on self banks.