id: 247e7bdc40c9490fbe87166ed92fb7be
parent_id: 00d3a064c9684d1280da2297523a8cfd
item_type: 1
item_id: ec1518e71453490f8fe5ab75cff7180c
item_updated_time: 1786980695659
title_diff: "[]"
body_diff: "[{\"diffs\":[[0,\"-17 \"],[-1,\"— MC calibration audit: range-model bias measured)\\\n\\\n> **Live: G51, server healthy.** EquityNet parked. **NEW: production equity audit over 1.02M labeled decisions — the range model underweights hero's winning spots 2.4x.**\\\n\\\n## The calibration audit (\"],[1,\"evening — range-bias deep-dive complete)\\\n\\\n> **Live: G51.** Range-model bias fully localized: NOT the corrector, NOT concentration — the conditional structure of RangeNet's predictions.\\\n\\\n## Measurement chain (all \"],[0,\"commit\"],[1,\"ted:\"],[0,\" 551\"]],\"start1\":31,\"start2\":31,\"length1\":264,\"length2\":230},{\"diffs\":[[0,\"cc, \"],[-1,\"corpus /home/jan/gen5_data/equity_v6_mc)\\\n\\\nMethod: \"],[1,\"e3ebec4)\\\n\\\n1. **Audit** (1.04M samples, corr ON vs OFF): bias identical → cor\"],[0,\"rec\"],[1,\"t\"],[0,\"or\"],[-1,\"der logs production MC's own equity at decision time + exact labels from actual holdings (all hands revealed in sims — no showdown selection bias, verified hand_runner.rs:1162). Mixed 9/6/3-max, neutral g51 pool, 1.02M sampl\"],[1,\" exonerated.\\\n2. **Label correction found+fixed**: fresh audit corpora had showdown-equity labels vs current-strength MC (river rows were definitionally clean; HU/flop inflated). Both corpora relabeled (676K/681K records).\\\n3. **Labels verified** by hand on spot records (board-trips, monotone-flop cas\"],[0,\"es\"],[1,\")\"],[0,\".\\\n\"],[-1,\"\\\n| Predictor | label-MAE | mean pred | mean label | slope |\\\n|---|---|---|---|---|\\\n| **Production MC** | **0.114** | 0.063 | 0.148 | 0.796 |\\\n| EquityNet v14 (parked) | 0.136 | 0.049 | 0.145 | ~0.6 |\\\n| EquityNet v10 (parked) | 0.138 | 0.016 | 0.145 | ~0.5 |\\\n\\\n**Findings:**\\\n1. **Hero is best-now 14.8% of the time at decision points; MC believes 6.3%.** The\"],[1,\"4. **Corrected audit** (definitions aligned): HU mean_pred 0.314 vs actual 0.451, slope 0.79; multiway slopes 1.15-1.24 with means 2.5x low; river 1.28.\\\n5. **Temperature sweep** (`range_calibration` bin, HU n=43K): softening toward uniform moves mean 0.216→0.352 (combo-uniform ATC\"],[0,\" b\"],[-1,\"i\"],[0,\"as\"],[-1,\" is in the RANGE MODEL (RangeNet+corrector jointly), not MC math — MC exactly computes E[beat|range]; the gap means ranges overweight opponent strength.\\\n2. Bias concentrated in hero-STRONG spots: label=1 → MC 0.28; label=0 → MC 0.007 (dead-spot detection near-perfect). Worst HU (slope 0.595) and flop (0.649); river overcorrects (1.272).\\\n3. nutpot mean 0.018 vs actual 0.136 (7.6x under).\\\n4. **Behavioral signature: the bot underestimates its own winning spots → thin value missed → passive lines.** Consistent with 7_6 (KK no river value bet vs capped line). Possibly explains part of day-2 live losses.\\\n\\\n**Caveat**: G48 thresholds were tuned WITH this bias — a global recalibration shifts every decision; must go through full sweep + sanity. Do not \\\"just fix\\\" the range mean.\\\n\\\n**IMMEDIATE FOLLOW-UP (flagged)**: A/B audit with `corr_enabled=false` (30-min collection) to localize the bias: RangeNet prior vs corrector over-concentration. The corrector's call-curves deliberately concentrate on stre\"],[1,\"eline at t=1) — **cannot reach 0.451**. Also: type-uniform ≠ combo-uniform (type weighting overweights pairs/suited).\\\n\\\n## Interpretation\\\n- The predicted ranges overweight opponent strength GIVEN THE ACTION SEQUENCE — a conditional-structure error: which hands take which lines. Retraining v17 with better labels/features is the fix path; no post-hoc scalar (temperature/floor) can correct it.\\\n- The residual 0.451 vs ATC 0.352 gap suggests actual opp holdings are anti-correlated with hero strength beyond card removal — likely a mix of selection effects in HU pot formation; not fully explained.\\\n- **Behavioral impact confirmed directionally**: bot underprices its strong spots → thin value missed. A8 vs AJ session hand consistent with the inverse case (overpricing opp aces less likely than underpricing own).\\\n\\\n## Options forward (decision needed)\\\n1. **7_6 likelihood fitting** — now has clean targets + verified labels + a calibration metric (slope per bucket). The corrector's action-likelihood curves ARE the conditional structure — fitti\"],[0,\"ng\"],[1,\" \"],[0,\"th\"],[-1,\" (\\\"callers have it\\\") — prime suspect on call-heavy fields.\\\n\\\n## Next steps (revised order)\\\n1\"],[1,\"em to realized showdown frequencies attacks exactly the diagnosed defect class. (~1-2d)\\\n2\"],[0,\". **\"],[-1,\"Bias localization** (corrector on/off audit) — hours, decides everything below\\\n2. **7_6 likelihood fitting** now has a measured target: fit curves so predicted P(beat) matches realized frequencies (proper scoring on the labeled corpus) — fixes bias + sequence-awareness together\\\n3. FutureActionNet / live fine-tune unchanged\\\n4.\"],[1,\"RangeNet v18** — retrain with the new corpora (they double as range data), richer action-history features. Bigger (2-3d + training).\\\n3. Both, in that order (1 is fast and de-risks 2's data).\\\n\\\n## Note for next session\\\n-\"],[0,\" Equ\"]],\"start1\":263,\"start2\":263,\"length1\":2077,\"length2\":2040},{\"diffs\":[[0,\"Net \"],[-1,\"stay\"],[1,\"remain\"],[0,\"s parked\"],[-1,\" (audit confirms MC closer to reality — but note both share the range bias; fixing ranges improves MC directly)\\\n\\\n## Session history 2026-08-17\\\n- v15-v20 EquityNet iterations → parked (2625149); live day-2 losses = variance, no defects found in spot checks\\\n- Commits today: 678080e, 1deaf4b, 1a1bcd1, 548ee48, 2625149, 55192cc\"],[1,\"; all infra reusable.\\\n- equity_v6_mc + equity_v6_nocorr corpora: relabeled, MC-attached — the calibration dataset for any range work.\\\n- range_calibration bin: t-sweep, support/combo-uniform variants.\"],[0,\"\\\n\\\n##\"]],\"start1\":2306,\"start2\":2306,\"length1\":345,\"length2\":221}]"
metadata_diff: {"new":{},"deleted":[]}
encryption_cipher_text: 
encryption_applied: 0
updated_time: 2026-08-17T15:35:46.719Z
created_time: 2026-08-17T15:35:46.719Z
is_locked: 0
type_: 13