id: 4c9ad101739c441c912da740cdd7a357
parent_id: 6014d2d75c21482aa3617929af79f3a0
item_type: 1
item_id: ec1518e71453490f8fe5ab75cff7180c
item_updated_time: 1787115968681
title_diff: "[]"
body_diff: "[{\"diffs\":[[0,\"-19 \"],[-1,\"morning — g53 temperature verdict: PARK)\\\n\\\n> **Live: G51 (unchanged, healthy).** Range-temperature candidate g53 fully evaluated overnight: **parked** — the offline calibration gain does not survive the composite production pipeline.\\\n\\\n## g53 evaluation (complete)\\\n\\\n| Gate | Result |\\\n|---|---|\\\n| Harrington | 52/52+1ign pass |\\\n| KK replay | REGRESSED — river hs 0.29→0.46, overplay Raise returns (temp partially undoes corrector) |\\\n| Calibration audit (917K, temp VERIFIED flowing: Gen5T HU mc 0.243 vs baseline 0.314) | **Mean bias NOT fixed**: overall 0.059 vs realized 0.152 (unchanged); HU slope 0.79→0.83 only; **nutpot destroyed 0.018→0.000** |\\\n| A/B vs g51 (200K hands) | −0.21 BB/100 dead heat; fish field +2.29 vs +2.19 control — play-neutral |\\\n\\\n**Why the offline gain didn't transfer**: the corrector runs AFTER the temperature and re-concentrates on call lines (its job) — netting out the softening exactly where it matters. The measured t≈0.55 was fitted on prior-only E[beat]; production hs passes through temp→corrector→MC. Composite systems resist single-knob fixes.\\\n\\\n## Incidents found & fixed this session (all committed)\\\n1. **Comma-spliced TOML** in collector configs (`key = 1.0, key2 = true` on one line = invalid TOML) — silently skipped (warning buried), stale same-named config satisfied bot_type → **the first \\\"g53\"],[1,\"— v18 training + oracle evaluation)\\\n\\\n> **Live: G51 (unchanged).** v18 calibration-aware training sweep running; oracle-in-corrector evaluated and parked.\\\n\\\n## Punkt 1: RangeNet v18 (IN PROGRESS)\\\n- Trainer upgraded (d955a70): RANGE_SMOOTHING label smoothing, **NLL + marginal-calErr checkpoint selection** (top-k selection rewarded overconcentration — the v17 disease), multi-file round-robin input\\\n- Data: 15M from equity_v6 (19.2M, 85-dim, v17's corpus) + g53full mixed sizes\\\n- ε-sweep chain running: **0.10 done — NLL 4.555, top1 3.94%, top10 22.9%** (v17: 4.35/24.2 — v18-e10 close on discrimination, calibrated by construction); 0.05 and 0.20 queued\\\n- Next after sweep: pick ε by NLL → **calibration\"],[0,\" audit\"],[-1,\"\\\"\"],[0,\" w\"],[-1,\"as actually a g51 re-run**. Fixed + unique collector names + glob-exclusion.\\\n2. **Registration failures now FATAL** in holdem_bots main (d2f9287) — immediately caught two more broken configs (sim.toml\"],[1,\"ith v18** (does MC-vs-realized mean gap close?) → if yes: KK/Harrington/sanity/A/B gates as g55 → live candidate\\\n\\\n## Punkt 2: Oracle-in-Corrector (DONE —\"],[0,\" par\"],[-1,\"sed as bot config → skip; v6/_nc collector splices → repaired).\\\n3. Relabel nohup killed by shell timeout → restarted via background_process.\\\n\\\n## Where the range bias stands (cumulative knowledge)\\\n- Bias = RangeNet prior ov\"],[1,\"ked with finding)\\\n- Implemented fully (corr_use_oracle knob, multi-path table load verified, personality class from stats, p\"],[0,\"er\"],[1,\"-\"],[0,\"co\"],[-1,\"ncentration (measured 3 ways: audit, temperature bracketing, overlap score −0.60)\\\n- Corrector exonerated (on/off identical); NOT fixable post-hoc by temperature (g53 proved this)\\\n- **The fix is v18 retrain**: calibration-aware training on the new corpora (multi-hand-per-context distributions from equity_v7_fit/v8_oracle), model selection by log-score/calibration (NOT top-k), plus the empirical oracle table as behavioral grounding\\\n\\\n## Assets built this session\\\n- range_temp knob (probs_to_hand_range_temp) — kept, off by default, for future experiments\\\n- g53 audit corpus (917K, temp-flowing, relabeled) — docum\"],[1,\"mbo table likelihoods; smoke-tested)\\\n- **Result: REGRESSES the KK gate** (turn 0.535→0.707 AllIn, river Raise returns). Mechanism: pool bots' P(call|strength) is nearly FLAT (2.7:1 strong:weak ratio — G48 pot-odds discipline) vs hand-fit curves' 8:1. Our bots are a poor behavioral template for HUMAN oppon\"],[0,\"ents \"],[1,\"(\"],[0,\"the co\"],[-1,\"mposite-pipeline effect\\\n- Empirical oracle table (models/oracle/oracle_table_v1.json, 5M decisions) + build_oracle_table bin\\\n- fit_likelihood --score / --score-oracle overlap scoring modes\\\n- Fatal-registration sim hardening\\\n\\\n## Recommended next (decision for user)\\\n1. **RangeNet v18 retrain** (the real fix): train on equity_v7_fit corpus (context-matched range data exists there via pre_ranges? — no: use the range_recorder data from the same runs) with (a) label smoothing/temperature-scaled targets, (b) checkpoint selection by calibration slope/log-score, (c) mixed-table data. ~2-3 days.\\\n2. Oracle-in-decisions (Tier 1): swap corrector curves for oracle-table lookups (per-config likelihoods) — improves eq_cc and corrector grounding without retraining.\\\n3. 7_6 still open; FutureActionNet \"],[1,\"rrector's target domain).\\\n- Keepers: the table + lookup machinery remain for FutureActionNet (absolute P(action) on sim side) — that use doesn't need human-transfer.\\\n- g54 config exists, Harrington-passes, NOT a candidate.\\\n\\\n## Session context\\\n- Live variance: day sessions net noise; A8-vs-AJ cooler verified standard.\\\n- Corpus disk: ~200G across gen5_\"],[0,\"data\"],[1,\";\"],[0,\" re\"],[-1,\"ady (oracle corpus).\\\n\\\n## Live\\\nG51 on persistent server; session results variance-dominated (day1 +122BB, day2 negative, A8-vs-AJ cooler verified standard\"],[1,\"tention pass due (equity_v6_mc/nocorr/g53audit can be culled after v18 work — keep v7_fit, v8_oracle, v6 main\"],[0,\").\\\n\\\n\"]],\"start1\":31,\"start2\":31,\"length1\":3362,\"length2\":1791}]"
metadata_diff: {"new":{},"deleted":[]}
encryption_cipher_text: 
encryption_applied: 0
updated_time: 2026-08-19T05:16:01.365Z
created_time: 2026-08-19T05:16:01.365Z
is_locked: 0
type_: 13