id: 1f39fc6598af4ae9a80af05a333bebb4
parent_id: 9e086eb8c83749af939cda4f803929fd
item_type: 1
item_id: ec1518e71453490f8fe5ab75cff7180c
item_updated_time: 1786984702342
title_diff: "[]"
body_diff: "[{\"diffs\":[[0,\"-17 \"],[-1,\"night — likelihood fitting launched)\\\n\\\n> **Live: G51, server healthy.** Overnight: fitting corpus collecting + MLE fit chained (bgp_0106adf15, bgp_01063a9c4).\\\n\\\n## Overnight chain (autonomous)\\\n1. Collection → `/home/jan/gen5_data/equity_v7_fit` (~10M samples, mixed 9/6/3-max, g51 pool, corrector ON) — records now carry **pre-corrector ranges + per-street action codes + opp freqs + street sizes** (commit a087658)\\\n2. `fit_likelihood` runs on ≤600K records (chain: `/tmp/kilo/overnight_fit.sh`\"],[1,\"late — CRITICAL CORRECTION to range-bias diagnosis)\\\n\\\n> **Live: G51.** Yesterday's \\\"conditional structure, no scalar fix\\\" conclusion was WRONG — my sweep bin had a combo-enumeration bug (id-ordering check skipped ALL suited/offsuit combos; beat tables were pairs-only). Fixed (91a4fc2), fit chain restarted.\\\n\\\n## Corrected picture (HU, n=43K, v6_mc corpus)\\\n\\\n| Quantity | Value |\\\n|---|---|\\\n| Realized hero-best freq | 0.451 |\\\n| Prior-weighted prediction (t=0) | 0.349 |\\\n| Combo-uniform ATC (t=1) | 0.527 |\\\n| **Temperature match** | **t≈0.55-0.60 → 0.447-0.456\"],[0,\", \"],[1,\"s\"],[0,\"lo\"],[-1,\"g: `equity_v7_fit/fit_log.txt`)\\\n3. Output: fitted (center, steep, size_slope) × 5 action classes + realized-vs-predicted diagnostics\\\n\\\n## Fitting method (a087658)\\\nRealized holdings (revealed every hand) are draws from the TRUE posterior → MLE over c\"],[1,\"pe 0.86** |\\\n\\\n**Where the underestimation comes from (current best answer): the RangeNet prior is TOO STRONG / TOO CONCENTRATED.** Mass sits on premium types while real holdings include far more weak offsuit/junk. A single ATC-temperature explains most of the mean gap. NOT conditional struct\"],[0,\"ur\"],[-1,\"v\"],[0,\"e \"],[-1,\"params: maximize Σ log L(realized) − Σ log Σ prior·ΠL. Beat-probs enumerated once per (hero,board), Arc-shared. Smoke-verified (predicted tracks realized ±0.03).\\\n\\\n## Morning checklist (in order)\\\n1. `tail /home/jan/gen5_data/equity_v7_fit/fit_log.txt` — fitted params + diagnostics\\\n2. **Scale caveat**: fitted centers are on beat-prob scale; corrector uses Xpot percentile scale (both monotone in strength). Apply call/bet/check classes to a `cash_nl_g53.toml` (copy g51 + corr_* overrides); preflop classes fitted but NOT wired (corrector skips street 0) — later extension.\\\n3. Validate: KK replay (`replay_kk_overplay_variants`), **7_6 replay** (the #[ignore] test — the acceptance target), Harrington 51+1, sanity_check g53, A/B vs g51\\\n4. Calibration re-audit: rerun mc_calibration_analyze-style comparison if a short g53 collection is made — fitted curves should pull MC mean toward realized mea\"],[1,\"(that was the bug artifact); NOT the corrector (on/off identical — still solid); NOT MC math.\\\n\\\nWhy the prior overconcentrates (hypotheses): softmax 169-class classifier; training labels only from hands that reached observation; G48-guided collection distribution narrower than reality.\\\n\\\n## New cheapest experiment (supersedes yesterday's pessimism)\\\n**Range temperature as a production knob**: post-process NN ranges `q = (1-t)·p + t·combo_uniform` at inference (probs_to_hand_range or corrector input), t≈0.55 start. Measured target exists (realized mean/slope per bucket). Full gates required — G48 thresholds are co-tuned to the strong prior; expect decision shifts everywhere.\\\n\\\n## Still running overnight\\\n- Collection equity_v7_fit (~10M) → corrected fit_likelihood chained (bgp_01095181a)\\\n- Fitted curves remain valuable: they conditio\"],[0,\"n \"],[-1,\"(\"],[0,\"the \"],[-1,\"measured target: HU 0.31→0.45 direction)\\\n5. If 7_6 passes + calibration improves + sims neutral-to-positive → commit + live candidate\\\n\\\n## Context (from earlier today)\\\n- Range-bias deep-dive: corrector exonerated; bias = RangeNet conditional structure; no scalar fix possible (e3ebec4)\\\n- The fitted likelihood curves ARE the conditional-structure correction layer — this is the attack on both the measured bias AND 7_6 sequence-blindness (line-conditioned likelihoods)\\\n- EquityNet parked; A8-vs-AJ session hand = short-stack cooler (verified)\\\n\\\n## Key notes\\\nId\"],[1,\"temperature on action lines (per-class), potentially better than one global t\\\n\\\n## Morning: check fit_log.txt → g53 candidates: (a) fitted curves, (b) global temperature t=0.55, (c) both. Then KK + 7_6 replays, Harrington, sanity, A/B, calibration re-audit.\\\n\\\n## Lesson recorded\\\nReport sweep mechanics (pair-only tables should have shown as suspiciously narrow supports). The 'impossible' result (actual weaker than ATC) was the tell — an impossible result m\"],[0,\"ea\"],[1,\"n\"],[0,\"s \"],[-1,\"`f3c42f44` | Roadmap `47031623` | Sanity `d632cea3` | Archive `283e98fe`\"],[1,\"measurement bug, not exotic poker truth.\"]],\"start1\":31,\"start2\":31,\"length1\":2291,\"length2\":2206}]"
metadata_diff: {"new":{},"deleted":[]}
encryption_cipher_text: 
encryption_applied: 0
updated_time: 2026-08-17T16:45:47.190Z
created_time: 2026-08-17T16:45:47.190Z
is_locked: 0
type_: 13