id: 10723d70e38a409ab42e7c9c00a88388
parent_id: 69e18c9e84a44742adaec36208c87f2e
item_type: 1
item_id: ec1518e71453490f8fe5ab75cff7180c
item_updated_time: 1787322484586
title_diff: "[{\"diffs\":[[0,\"08-2\"],[-1,\"0 — g55 gate fail\"],[1,\"1 — call model fix\"],[0,\"ed)\"]],\"start1\":27,\"start2\":27,\"length1\":24,\"length2\":25}]"
body_diff: "[{\"diffs\":[[0,\"08-2\"],[-1,\"0 — v19 trained, g55 FAILED gate\"],[1,\"1 — call model fixed, pool v2 stats\"],[0,\")\\\n\\\n>\"]],\"start1\":29,\"start2\":29,\"length1\":40,\"length2\":43},{\"diffs\":[[0,\"used\"],[-1,\" (user decision), G51 unchanged.** v19 models trained; board-exact arm wins offline; g55 candidate fails Gate 3 vs Gen2 — regression isolated and real\"],[1,\", G51 binary unchanged on server (stopped anyway).** Preflop call model structurally fixed; pool personalities recalibrated to human stat ranges. Sanity gates running on g51-with-fix\"],[0,\".\\\n\\\n#\"]],\"start1\":83,\"start2\":83,\"length1\":158,\"length2\":190},{\"diffs\":[[0,\"nkt \"],[-1,\"3: RangeNet v19 A/B — DONE (winner: board-exact)\\\n- 10.2M v19 samples, 30 epochs, GPU, both arms same recipe (eps \"],[1,\"5: Preflop Call Model Fix — DONE (commit a5b6264)\\\n- **Root cause of call-fest**: facing-raise path double-counted matched money as phantom implied odds (all behind at min(vpip,\"],[0,\"0.\"],[-1,\"0\"],[0,\"5)\"],[-1,\"\\\n- **Board-exact 449-dim: NLL 4.470, top1 4.92%, top10 25.1%, calErr 0.0491** vs 85-dim: 4.551/4.07%/22.9%/0.0495\\\n- Inference wired: shared `append_board_exact_features` (train/infer identical, layout test), predictor auto-detects 449-dim (commit 95bdc4d)\\\n- Models: `range_v19_board.safetensors`, `range_v19_85d.safetensors`\\\n\\\n## Punkt 4: g55 gates — FAILED (not promoted)\\\n- Gate 1 Harrington (NN path, 449-dim model confirmed loaded): PASS 52/52 (1 known ignore)\\\n- Gate 2 self-play: clean. Gate 3: Gen3 +0.6, G48 +1.9 (g51: +1.7), TAG +1.8 (g51: +1.5), Nit +0.1, **Gen2 −0.27 BB/100 → FAIL** (g51 sanity was +0.6). Gate 4 Mixed: 0.0 (g51: +0.4)\\\n- A/B g55 vs g51: −0.11 BB/100 dead heat over 200K cand-hands\\\n- **Control run (identical seeds, 4×g51+5×Gen2×5000h×5 seeds): g51 +2.80 BB/100 vs g55 −0.27.** Per-seed swing −490…−1140 BB, ALL seeds negative → real regression vs Gen2, not noise.\\\n- **Diagnosis**: sharper board-exact prior wins in-distribution (G48/TAG = pool family) but is exploited by Gen2'\"],[1,\"); opponent ranges capped at 0.50 (matrix supports 0-1) → ~0.30 equity floor under trash → VPIP 0.85-0.94 pool-wide, 99.5% flops\\\n- **Fix**: call range = raiser's PFR (same estimator as reraise path, epf_raise_range_factor removed); implied money only from NOT-yet-matched players at P(call)=max(vpip−pfr,0)/sizing (measured when observer warm, config priors when cold); all arbitrary 0.50 caps → [0.02, 1.0]; call_thresh cap removed\\\n- **4 regression tests** in call_model_tests (trash HU fold, no-phantom-money, 55 set-mine, K8o vs shove); 538 lib tests green, 52/52 Harrington replays pass\\\n- **Live check (241 hands)**: hero VPIP 32.4%, NOT table-scaled — leak never ignited live (vpip-gated implied odds held). Problem was sim-pool + training-data realism.\\\n\\\n## Punkt 6: Pool stat recalibration — DONE (probe-verified)\\\n- Fish-tier knobs loosened (safety 0.55-0.65, domination 0.03-0.05, npot 0.03-0.04): **VPIP spread now Tag 0.23 / Lag 0.38 / Maniac 0.46 / Whale 0.66 / Fish 0.71 / Station 0.78 / Gen2 0.13** (was 0.85-0.94 all), flop 86.5%\\\n- Per-class targets now TRAINABLE (classes separable in feature space). Machinery ready: opp_name in RawRangeContext, per-class oracle rows in build_range_targets (17d4d21)\\\n- **Next major cycle (v22)**: collect on recalibrated pool → per-clas\"],[0,\"s \"],[1,\"s\"],[0,\"of\"],[-1,\"f-distribution rollout lines — overconfident wrong reads. Suspects: (1) corrector corr_* knobs tuned for v17-width priors over-concentrate with v19, (2) v19 training pool lacks Gen2-style lines.\\\n- Next: g55-nc (corr off) + g55-85d isolations; consider corrector re-fit on v19 priors.\\\n\\\n## Punkt 1-2 (done earlier): v19 collection + oracle regen — see commits 1e3144f/3d97fd9/c23dc19/cab5173\"],[1,\"t targets → train → g58 → gates\\\n\\\n## Earlier points (2026-08-20, status unchanged)\\\n- v19 corpus (10.2M) + v21 Gen2-coverage (3.2M): g57 = best new-gen candidate (Gen2 isolation +1.25 vs g55 −0.27, G48 +1.7/TAG +1.5 retained) but FAILED Gate 1 (KQs top-pair passive, 4_5_99 turn barrel) — pool-average labels too wide for described opponents; per-class fix is the answer\\\n- g57nc (corr off): worse everywhere → corrector stays\"],[0,\"\\\n- O\"]],\"start1\":277,\"start2\":277,\"length1\":1522,\"length2\":1901},{\"diffs\":[[0,\"cle \"],[-1,\"v19 \"],[0,\"table\"],[-1,\": all size contexts populated after street_sizes index fix; superseded old table\\\n\\\n## Gen 6 (renamed): mirrors + rollouts — details in Ideas note f3c42f44. NN-for-speed = distilled EV net (primary) / rollout-policy net (secondary).\\\n\\\n## Session context / lessons\\\n- Live range collection: needs showdown→range-sample a\"],[1,\" v19 regenerated (street_sizes index bug fixed, cab5173)\\\n- Training RAM ceiling: 13.4M samples ≈ 80.5GB; RANGE_MAX_SAMPLES cap; u8 one-hot storage would 7x it\\\n\\\n## Gen 6 (renamed): mirrors + rollouts — details in Ideas note f3c42f44.\\\n\\\n## Key lessons (2026-08-19..21)\\\n- Tightness death spiral / call-fest equilibrium: pool equilibria are structural, not configurable — vali\"],[0,\"da\"],[-1,\"p\"],[0,\"te\"],[-1,\"r (live only reveals\"],[1,\" flop rate +\"],[0,\" per-\"],[-1,\"player at showdown, no all-reveal event); wire with next live rebuild\\\n- Pool realism ladder agreed: (b) dynamic curves w/ shrinkage first → (a) pool v2 from live stats → (c) live oracles; all three draw from live hand histories → resume live reasonably soon as data engine\"],[1,\"class stats with probes before collecting\\\n- Symmetric-pool leaks invisible in A/B (Gen2 = only dissimilar opponent, always thinnest margin)\\\n- Live hand histories at /tmp/super_marvin_hand_history.log (hero = seat 0/Bolsa); 241 hands parsed OK\"],[0,\"\\\n\\\n##\"]],\"start1\":2180,\"start2\":2180,\"length1\":634,\"length2\":647}]"
metadata_diff: {"new":{},"deleted":[]}
encryption_cipher_text: 
encryption_applied: 0
updated_time: 2026-08-21T14:36:19.697Z
created_time: 2026-08-21T14:36:19.697Z
is_locked: 0
type_: 13