Drop7 Research

Type to search every page, record and vocabulary term.

GitHubPlay

P-SOL-1: panel2 sibling corpus with CRN Kaplan-Meier outcome labels, NNUE retraining, and the G0/G1/G2 offline gates

Full protocol: runs/RUN-20260823T191900Z-b9f8f80d/kimi-k3-main-design.md sections 3-7 and 10 (P-SOL-v1), fixed as written; this record freezes its decision structure.

supersededtier: CHECKalgorithmicno-gameplaypublic-policyEX-20260823-sol-corpus-and-offline-gate-4d3d86e4
On this page
Linked approachThe data factory, and the switch that was left off
Theories
Dates
Created
Updated
Record idEX-20260823-sol-corpus-and-offline-gate-4d3d86e4

No explanation has been written for this record yet.

Technical recordThe registered protocolEX-20260823-sol-corpus-and-offline-gate-4d3d86e4
Hypothesis
Full protocol: runs/RUN-20260823T191900Z-b9f8f80d/kimi-k3-main-design.md sections 3-7 and 10 (P-SOL-v1), fixed as written; this record freezes its decision structure. Stages: (G0) extend the sibling-corpus generator with a panel2 mode (PanelRecordV2 as specified; v1 output byte-identical, gated) and run the continuation-engine fidelity ladder on 256 D3-N7M6 behaviour games (seeds 0xa5216000-0xa52160ff), K=6, H=48, all three engines on identical roots with identical CRN tapes; mini-ladder of 64 roots (D1 vs D3 arms) runs first. Proceed with the cheapest engine whose within-root Kendall-tau vs D3 N7M6 has 95% LB >= 0.85 (cluster bootstrap by origin, 10,000 resamples, seed 0xb0071eaf); if none, stop (F1). (C) Main corpus 2,048 games on 0xa5217000-0xa52177ff (70% D3 N7M6, 15% D3 N7M6 eps=0.10, 15% D2 behaviour), ~48,000 roots at panel stride, K=16, H=48 continuations with the G0-selected engine; D2 fallback branch preregistered (6,000 roots, K=8, H=40) if the D1 branch is unaffordable. Whole-origin split origin mod 10 -> train 0-7 / val 8 / test 9. (T) Train the unchanged 572k NNUE with L = 1.0 pairwise-logistic ranking on delta KM expected lifetime (pair weight min(1,|delta|/(2*SE_CRN))) + 0.5 censor-aware log-lifetime + 0.25 masked-BCE hazard + 0.25 flow; AdamW 3e-3, wd 1e-4, cosine, batch 512 roots, 30 epochs, early stop and model selection on val within-root tau; 5 init seeds. (G1/G2) Gate set: 512 fresh D3-N7M6 games on 0xa5219000-0xa52191ff, 4,096 roots with exact D4 sibling values (d4-rank.cpp adaptation), exact D1/D2, incumbent LeafNet and student on identical roots.
Arms
ArmNameEntry pointManifest
Candidatepanel2-km-label-nnue-leafapproaches/lifetime-objective/sibling-corpus/generate.cpp–
Comparatorincumbent played-action LeafNet (finding-08) and exact D1/D2/D4 orderingsapproaches/lifetime-objective/learned-leaf/train_leaf.py–
Classification
algorithmic
Information boundary
public-policy
Benchmark tier
CHECK
Lifecycle
superseded
Theories tested
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
Training-lease block only (0xa5216000-0xa52191ff inside the SEEDLEASE-A52 reserve): G0 ladder 256 origins, main corpus 2,048, gate set 512 (gate origins become development-read on use and never train). No development, protected or final cohort is read. The offline gate compares against the incumbent LeafNet checkpoint (runs/RUN-A52-LEAF) without re-reading its cohorts. Gameplay (blend tuning and 256-game screens, including the D3 N7M6 arm) is a successor experiment registered only if G1 passes.
Pass criteria
  1. Panel2 byte-identity gate: v1 panel and default state output byte-identical with panel2 off; featurizer parity on 1,000 stored RUN-A51D rows.
  2. G0: cheapest engine tau 95% LB >= 0.85 (else stop, F1).
  3. G1: student top-1 >= 0.55 AND >= incumbent + 0.03 (McNemar LB > 0). Marginal band [0.52,0.55): proceed to a d4s5-only screen. Below 0.52 or below incumbent+0.01: fail (F2/F3 routed by G2).
On pass
Register the successor gameplay experiment exactly as P-SOL-v1 section 8 (blend tuning on reusable-development cohorts; 256-game paired screens at d4s5/d4s7 with rarity and futility stops; the D3 N7M6 deployment arm) before any gameplay seed is opened. No tier above CHECK is claimed by this experiment.
On fail
Record valid + fail with the G2 decomposition stating whether the target or the fit failed; the label family (F3) or the coverage claim (F2) is closed for this configuration; no gameplay cohort is opened; the D2-fallback and scale-out routes are noted, not run.
Gate fixed before controlled data
yes
Resources
Wall seconds18000
CPU threads26
Max host bytes34359738368
Max GPU bytes17179869184
GPU devicesgfx1151
Stop conditions
  1. S1: any stage exceeding 1.3x its budget line (32.45 CPU-h total to the gate, cap 35; GPU cap 4 h) halts for coordinator decision; partial recorded.
  2. S2: G0 kill (F1) stops at ~14 CPU-h.
  3. S3: G1 fail stops before any gameplay.
  4. Fallback diagnostic on corpus overrun: halt at the 50% milestone if projection exceeds 1.3x, keep whole completed games, run the fit-quality half of G2 on the partial corpus; a scale-up request routes to the drop7-scale-out skill.
  5. Mini-ladder first: if D1 tau point estimate < 0.75 on 64 roots, skip full D1 assessment and evaluate D2.
Expected artifacts
  • runs/RUN-20260823T191900Z-b9f8f80d/kimi-k3-main-design.md (design, Kimi K3, P-SOL-v1)
  • runs/<sol-run>/g0/ladder.json and gates.log
  • runs/<sol-run>/corpus/ (panel2 files + summary)
  • runs/<sol-run>/train/ (5 seeds, metrics)
  • runs/<sol-run>/gate/g1-g2.json
Amendments
TimestampBefore controlled dataReason
2026-08-23T21:53:26ZyesSuperseded by EX-20260823-sol-corpus-and-offline-gate-v2-f95f56e4 before any leased seed was opened: the G0 budget lines priced D3 N7M6 at 292 moves/s (the D3 N5M1 figure); measured cost is 1.754-6.83 CPU-s/move (erratum in runs/RUN-20260823T191900Z-b9f8f80d/kimi-k3-psol-v2-correction.md). Stop rule S1 fired on projection; the panel2 implementation and all engineering gates from this protocol carry forward unchanged. This record is retained immutably; the hash is recomputed over the superseded record per repo convention, and the run record RUN-20260823T215500Z-fc74e0b4 retains the original preregistration hash 017a6bcd6a44ce65ef961259e10618470213e6494236aeac3a6917d704797a37.
Agent contextHow to extend this record

To add a reader-facing explanation, write web/content/research/EX-20260823-sol-corpus-and-offline-gate-4d3d86e4.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-4d3d86e4.json, validated against research/schemas/experiment-v1.schema.json. Protocol hash: 770db739af6802ebc489dc112a4cef53703a2d2a4af9210aef8e54d330e02d23.

Registered by Claude Code / claude-fable-5.