P-SOL-2: corrected two-rung G0 fidelity ladder (fast-d3s7 proxy + native D3 N7M6 guardrail), fast-engine corpus continuations, and the G1/G2 offline gates

Successor to EX-20260823-sol-corpus-and-offline-gate-4d3d86e4, whose G0 budget arithmetic was wrong by 20-77x for every native D3 N7M6 stage (S1 halted before any leased seed was read; erratum and corrected design: runs/RUN-20260823T191900Z-b9f8f80d/kimi-k3-psol-v2-correction.md, Kimi K3).

completedtier: CHECKalgorithmicno-gameplaypublic-policyEX-20260823-sol-corpus-and-offline-gate-v2-f95f56e4

On this page

Linked approachThe data factory, and the switch that was left off

Theories

TH-20260823-sibling-outcome-labels-fix-leaf-0c3d480d

Results

RS-20260823T225753Z-0fbd48c3

Dates

Created: 2026-08-23T21:52:46Z
Updated: 2026-08-23T22:57:53Z

Record idEX-20260823-sol-corpus-and-offline-gate-v2-f95f56e4

No explanation has been written for this record yet.

Technical recordThe registered protocolEX-20260823-sol-corpus-and-offline-gate-v2-f95f56e4

Hypothesis: Successor to EX-20260823-sol-corpus-and-offline-gate-4d3d86e4, whose G0 budget arithmetic was wrong by 20-77x for every native D3 N7M6 stage (S1 halted before any leased seed was read; erratum and corrected design: runs/RUN-20260823T191900Z-b9f8f80d/kimi-k3-psol-v2-correction.md, Kimi K3). All engineering from v1 is retained: the panel2 generator, byte-identity/CRN/mirror/legality gates (all passed at HEAD 49b4631), the training recipe, the G1/G2 offline gate and the incumbent comparison. Changes, per the v2 correction: (G0 rung 1, powered) D1 and D2 continuation engines are certified against a fast-engine depth-3 seven-stratum M=1 reference (fast-d3s7, ~0.004 CPU-s/move) on 1,200 seed-free synthetic roots, 7 siblings x K=8 x H=50, CRN-paired tapes; proceed per candidate iff mean per-root Kendall-tau LB95 >= 0.75 AND top-1 agreement LB95 >= 0.80. (G0 rung 2, guardrail) fast-d3s7 is spot-checked against native D3 N7M6 on 6 roots harvested from retained C0 traces (already-read data), 7 x K=4 x H=40 (equivalent-cost alternative K=6 H=26), CRN-matched; kill iff top-1 agreement <= 4/6 OR mean tau < 0.5 OR any root tau < 0; this is a guardrail, not an equivalence claim, and every downstream conclusion is worded 'M1-continuation evaluation; M6 transfer assumed and guardrailed'. (Corpus) main-corpus continuations switch from D3 N7M6 to fast-d3s7; a 500-root fast-d3s7 pilot (2.5 CPU-h) precedes the full corpus; behaviour mix, seed arithmetic (0xa5217000-0xa52177ff), training and G1/G2 are unchanged from v1. Budget to the offline gate: rung 1 5.2 CPU-h + rung 2 12.8 (priced at 6.83 s/move) + regates 0.5 + pilot 2.5 + preregistered reserve 14.0 = 35.0 CPU-h cap; reserve uses only (i) rung-2 extension to 12 roots if realised rate <= 2.5 s/move, (ii) rung-1 top-up if a CI straddles a threshold, (iii) returned.
Arms: Arm Name Entry point Manifest
Candidate panel2-km-label-nnue-leaf-fastd3s7 approaches/lifetime-objective/sibling-corpus/generate.cpp –
Comparator incumbent played-action LeafNet (finding-08) and exact D1/D2/D4 orderings approaches/lifetime-objective/learned-leaf/train_leaf.py –
Classification: algorithmic
Information boundary: public-policy
Benchmark tier: CHECK
Lifecycle: completed
Theories tested: Complete-sibling long-horizon outcome labels fix the learned leaf's within-root discrimination (P-SOL)
Primary metric: G1: student top-1 vs exact-D4 argmax on the 4,096-root gate set (clustered by origin), against 0.55 absolute and incumbent+0.03 (paired McNemar 95% LB > 0)
Secondary metrics: G0: per-engine within-root Kendall-tau vs D3 N7M6 with cluster-bootstrap CIs
G2 target quality: label-argmax top-1 vs exact D4 (>= 0.60 required to call the target right)
G2 fit quality: student-vs-label within-root tau against 0.7x the split-half label ceiling
G2 coverage slice: tau by behaviour source (d3 vs d2/eps roots)
incumbent LeafNet row for every metric; exact D1 (0.486) and D2 (0.568) anchors
init-seed spread over 5 seeds
corpus counts, censoring rate at H=48, per-stage CPU-hours
Statistical unit: root
Uncertainty method: cluster bootstrap over origin games (10,000 resamples, seed 0xb0071eaf) for tau and gaps; paired McNemar for the incumbent comparison
Data role: no-gameplay
Seed leases: SL-20260823T215000Z-a5216000
Whole-origin split: yes
Reuse disclosure: G0 consumes zero leased seeds: rung-1 and pilot roots come from a seed-free synthetic pool, rung-2 roots from retained already-read C0 artifacts (runs/RUN-A525-reveal, data status unchanged). The training lease SL-20260823T215000Z-a5216000 (still reserved, never opened under v1) is opened only for the main corpus and gate set after G0 passes, with the same sub-ranges as v1.
Pass criteria: Re-gates for the fast-d3s7 configuration: byte-exact agreement with the native single-knob d3 s7 M1 search on probe decisions, CRN tape parity, mirror invariance, seed-free tapes only.
G0 rung 1: at least one of D1/D2 with mean per-root Kendall-tau LB95 >= 0.75 AND top-1 agreement LB95 >= 0.80 vs fast-d3s7 (cluster bootstrap, 10,000 resamples, seed 0xb0071eaf).
G0 rung 2 guardrail not triggered: top-1 agreement > 4/6 AND mean tau >= 0.5 AND no root tau < 0 (fast-d3s7 vs native D3 N7M6, 6 roots).
G1 (unchanged from v1): student top-1 vs exact-D4 argmax >= 0.55 AND >= incumbent LeafNet + 0.03 (paired McNemar LB > 0); marginal band [0.52,0.55) proceeds to a d4s5-only screen.
On pass: As v1: register the successor gameplay experiment (P-SOL-v1 section 8, including the D3 N7M6 arm) before any gameplay seed is opened. If the fast-M6 port passes its equivalence gates before rung 2 executes, the preregistered contingency upgrades rung 2 to fast-M6 with R2=48 inside the same envelope.
On fail: Rung-1 total failure or guardrail kill: record valid + fail; the cheap-continuation label mechanism is rejected for this configuration; the fast-M6 engineering port (E-FAST-M6) is the named reopening route. G1 fail: as v1, routed by the G2 decomposition.
Gate fixed before controlled data: yes
Resources: Wall seconds 18000
CPU threads 26
Max host bytes 34359738368
Max GPU bytes 17179869184
GPU devices gfx1151
Stop conditions: S1 (re-armed on the corrected cost model): any stage projecting over its envelope halts before consuming roots or seeds; partial recorded.
Rung-1 kill: both candidates fail -> stop before the corpus.
Rung-2 guardrail kill -> stop before the corpus.
Corpus overrun fallback as v1.
Reserve spends only on its three preregistered uses.
Expected artifacts: runs/RUN-20260823T191900Z-b9f8f80d/kimi-k3-psol-v2-correction.md (corrected design, Kimi K3)
runs/RUN-20260823T215500Z-sol/g0v2/ (regates, rung1.json, rung2.json, decision)
runs/RUN-20260823T215500Z-sol/corpus/ (pilot + main)
runs/RUN-20260823T215500Z-sol/train/ and gate/
Amendments: Timestamp Before controlled data Reason
2026-08-23T22:57:53Z no Closed out after G0: rung 1 S1-halted on measured fast-d3s7 cost (54.6 ms/move vs the 4 ms planning figure), rung 2 guardrail killed the M1 proxy (mean tau 0.370, top-1 4/6). failureAction applies; no leased seed was opened; lifecycle completed and hash recomputed per repo convention. The run record retains the preregistration hash cdf85bc4b3ca7925d0805524aa609e7717906c5d21e68d47aa4dba2e6c4813f6.

Arm	Name	Entry point	Manifest
Candidate	panel2-km-label-nnue-leaf-fastd3s7	`approaches/lifetime-objective/sibling-corpus/generate.cpp`	–
Comparator	incumbent played-action LeafNet (finding-08) and exact D1/D2/D4 orderings	`approaches/lifetime-objective/learned-leaf/train_leaf.py`	–

Wall seconds	18000
CPU threads	26
Max host bytes	34359738368
Max GPU bytes	17179869184
GPU devices	gfx1151

Timestamp	Before controlled data	Reason
2026-08-23T22:57:53Z	no	Closed out after G0: rung 1 S1-halted on measured fast-d3s7 cost (54.6 ms/move vs the 4 ms planning figure), rung 2 guardrail killed the M1 proxy (mean tau 0.370, top-1 4/6). failureAction applies; no leased seed was opened; lifecycle completed and hash recomputed per repo convention. The run record retains the preregistration hash cdf85bc4b3ca7925d0805524aa609e7717906c5d21e68d47aa4dba2e6c4813f6.

Technical recordResults recorded against this protocol1 record

valid runoutcome: failnot-supported-as-testedtier: pilotRS-20260823T225753Z-0fbd48c3

P-SOL-2 stage G0 fails per the frozen failureAction, with no leased seed opened. Re-gates 7/7 pass, including fast-vs-native d3s7 parity (240 decisions, 0 action/work mismatches). Rung 1 S1-halted on measured cost: fast-d3s7 continuations run at 54.6 ms/move, ~13x the design's 4 ms estimate, projecting 47.9 CPU-h against the 6.76 ceiling; no tau was computed and the 1,200-root seed-free pool is retained. Rung 2 executed within envelope and the guardrail KILLED the proxy: on 6 CRN-matched C0 roots (exact replay verified), fast-d3s7 vs native D3 N7M6 within-root KM-lifetime orderings agree at mean tau 0.370 (LB95 0.283, min 0.053), top-1 4/6 - both kill conditions trigger, worst at late-game roots. Scientific consequence: the M=6 reveal quadrature genuinely changes within-root orderings, so an M=1 continuation corpus cannot carry D3 N7M6 label semantics; the fast-engine M=6 port (E-FAST-M6) is the required reopening route for any cheap-continuation label programme. Total ~7.5 CPU-h of the 35 cap; training lease still reserved and unopened.

What it had to pass

✓Re-gates: v1 byte-identity, fast-vs-native d3s7 decision parity, CRN tape parity, mirror invariance, thread/run determinism, legality — observed: 7/7 pass; parity 240 decisions, 0 action/work mismatches (work 38,179,907 both sides); CRN 2,016 comparisons 0 mismatches
✕Rung 1: at least one of D1/D2 with tau LB95 >= 0.75 and top-1 LB95 >= 0.80 vs fast-d3s7 on 1,200 roots — observed: not run: S1 halt - measured fast-d3s7 54.6 ms/move (design assumed 4), projection 47.9 CPU-h vs 6.76 ceiling; root pool retained
✕Rung 2 guardrail: top-1 > 4/6 AND mean tau >= 0.5 AND no root tau < 0 (fast-d3s7 vs native D3 N7M6, 6 CRN-matched C0 roots) — observed: mean tau 0.370 (LB95 0.283, median 0.304, min 0.053), top-1 4/6 - kill conditions 'mean tau < 0.5' and 'top-1 <= 4/6' both trigger; native 3.40 s/move realised

Technical recordRecorded metricsRS-20260823T225753Z-0fbd48c3

regates

7/7 pass; fast-vs-native parity 240 decisions 0 mismatches, work 38,179,907 both sides

rung1

halted

S1 on projection

ratesMsPerMove

d1: 0.0740
d2: 1.9900
fastD3s7: 54.6000

projectionCpuH

47.9000

ceilingCpuH

6.7600

rung2

roots: 6
K: 4
H: 40
meanTau: 0.3700
tauLB95: 0.2830
medianTau: 0.3040
minTau: 0.0530
top1Agreement: 4/6
kill: mean tau < 0.5
top-1 <= 4/6
nativeSecondsPerMove: 3.4000
cpuHours: 4.6000

totalCpuHours

7.5000

Limitations

Rung 2 is a 6-root guardrail: it detects gross misalignment (kill power ~0.89 at true agreement 0.5) and its kill here is decisive for the proxy, but the tau point estimates carry wide intervals.
Rung 1's powered comparison never ran, so D1/D2 fidelity to fast-d3s7 is unmeasured.
The 20-minute wall overage over the subagent's 2.5 h cap was spent in the single-threaded determinism gate and is disclosed.

Open the result record

Agent contextHow to extend this record

To add a reader-facing explanation, write web/content/research/EX-20260823-sol-corpus-and-offline-gate-v2-f95f56e4.mdx; it renders above this record on the next request. The registered protocol itself is in the technical record above.

Record file: research/experiments/EX-20260823-sol-corpus-and-offline-gate-v2-f95f56e4.json, validated against research/schemas/experiment-v1.schema.json. Protocol hash: 2635b0477b90cb559455576e7744417d872283529909202f1a69805e15b28509.

Registered by Claude Code / claude-fable-5.