id: ac46e4e60481424abf9c9f8b0f59d17d
parent_id: 4c9ad101739c441c912da740cdd7a357
item_type: 1
item_id: ec1518e71453490f8fe5ab75cff7180c
item_updated_time: 1787143203772
title_diff: "[{\"diffs\":[[0,\"26-0\"],[-1,\"7-31\"],[1,\"8-19 — v19 collection\"],[0,\")\"]],\"start1\":24,\"start2\":24,\"length1\":9,\"length2\":26}]"
body_diff: "[{\"diffs\":[[0,\"— v1\"],[-1,\"8 training + oracle evaluation\"],[1,\"9 collection RUNNING\"],[0,\")\\\n\\\n>\"]],\"start1\":35,\"start2\":35,\"length1\":38,\"length2\":28},{\"diffs\":[[0,\" v18\"],[-1,\" calibration-aware training sweep running; oracle-in-corrector evaluated and parked\"],[1,\"-e05 trained (parked as range model for v19 collector). **v19 dual-purpose collection running** (8×150K hands, ETA ~8-9h)\"],[0,\".\\\n\\\n#\"]],\"start1\":90,\"start2\":90,\"length1\":91,\"length2\":129},{\"diffs\":[[0,\"nkt \"],[-1,\"1:\"],[1,\"3: v19\"],[0,\" Range\"],[-1,\"Net 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 audit with 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 — parked with finding)\\\n- Implemented fully (corr_use_oracle knob, multi-path table load verified, personality class from stats, per-combo 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)\"],[1,\" Collection (RUNNING — started 2026-08-19 14:33)\\\n- **Ziel**: RangeNet v19 mit board-exakten Features (A: 5×52+2×52 One-Hots) + Verteilungs-Targets gegen Überkonzentration; Oracle-Tabelle regeneriert aus denselben Daten (Sizing-Änderung invalidiert alte Tabelle)\\\n- **Recorder dual logging** (commit 1e3144f): `RangeSample.raw` = `{hero_cards[2], opp_cards[2], board[n], actions[(street,code,chips)], faced_chips[4], pot, bb}` — Karten-IDs `rank*4+suit`, Chips roh (Xpot-Re-Bucketing offline). Alte Corpora bleiben deserialisierbar (serde default). `RANGE_RECORDER_MAX_SAMPLES` env cap neu.\\\n- **Sizing-Diversität** (füllt leere medium/large Oracle-Zellen): Lag 0.55/0.35, Maniac 0.75/0.50, neu **g51_tagbig_nn** 0.45/0.33\\\n- **Run**: `scripts/v19_collect.sh` → `/home/jan/gen5_data/equity_v19`, g51-NN-Pool + TagBig, 9/6/3-Max gemischt, Caps range 1.6M / equity 400K pro Tisch, Collector `gen5_equity_collect_v19` (range_v18_e05 + corr + eq_cc)\\\n- **KRITISCHER BUG gefixt** (dc92fc0): `min_raise = 20` (vom alten gen5_equity_collect.sh geerbt) degeneriert das Spiel bei Blinds 500/1000 → 1.1% Flop-Rate, 100x Hand-Rate. **Immer min_raise = BB = 1000.** Validierung nach Fix: Flop 15%, Street-Verteilung {3:45%,4:31%,5:24%} = v6-Referenz, ~5 Hände/s/Tisch = v7_fit-Referenz.\\\n- Call-Realismus des Pools: pro Klasse steil & differenziert (Tag/Reg/Nit 180-380x, Fish/Station 6-7x weak→strong) — KEINE Persönlichkeits-Chirurgie nötig; frühere \\\"2.7:1 flach\\\"-Messung war Aggregations-Artefakt.\\\n- **Nach Abschluss**: (1) merge + train v19 (A/B: board-exakt vs 85-dim, Log-Score-Metrik), (2) Oracle-Tabelle aus v19-equity_data regenerieren, (3) Verteilungs-Targets via Kontext-Clustering auf raw-Block.\\\n\\\n## Punkt 1: RangeNet v18 (DONE — e05 geparkt)\\\n- NLL+calErr-Checkpoint-Selection, ε-Sweep auf 15M: **e05 gewinnt** (NLL 4.555). v18-e05 verbessert Produktions-Kalibrierung, ist jetzt das Range-Modell des v19-Collectors.\\\n- Überkonzentration (v17-Krankheit) nur teilweise behoben → v19-Ansätze (board-exakt + Verteilungs-Targets) sind die Fortsetzung.\\\n\\\n## Punkt 2: Oracle-in-Corrector (DONE — parked with finding)\\\n- corr_use_oracle regrediert KK-Gate (Pool-P(call|strength) zu flach\"],[0,\" vs \"]],\"start1\":223,\"start2\":223,\"length1\":1065,\"length2\":2190},{\"diffs\":[[0,\"and-\"],[1,\"ge\"],[0,\"fit\"],[-1,\" curves' 8:1. Our bots are a poor behavior\"],[1,\"te 8:1 — Bots sind schlechte Verh\"],[0,\"al\"],[-1,\" \"],[0,\"te\"],[-1,\"mplate for HUMAN opponents (the corrector's target domain).\\\n- Keepers: the tab\"],[1,\"ns-Vorlage für Menschen). Tabel\"],[0,\"le + \"],[-1,\"l\"],[1,\"L\"],[0,\"ookup\"],[-1,\" ma\"],[1,\"-Mas\"],[0,\"chiner\"],[-1,\"y remain fo\"],[1,\"ie bleiben fü\"],[0,\"r Fu\"]],\"start1\":2414,\"start2\":2414,\"length1\":167,\"length2\":115},{\"diffs\":[[0,\"nNet\"],[-1,\" (absolute P(action) on sim side) — that use doesn't need human-transfer.\\\n- g54 config exists, Harrington-passes, NOT a c\"],[1,\".\\\n- g54 existiert, kein K\"],[0,\"andidat\"],[-1,\"e\"],[0,\".\\\n\\\n#\"]],\"start1\":2538,\"start2\":2538,\"length1\":137,\"length2\":40},{\"diffs\":[[0,\"t\\\n- \"],[-1,\"Live variance: day sessions net noise; A8-vs-AJ cooler verified standard.\\\n- Corpus disk: ~200G across gen5_data; retention pass due (equity_v6_mc/nocorr/g53audit can be culled after v18 work — keep v7_fit, v8_oracle, v6 main)\"],[1,\"Corpus-Aufräumen durchgeführt 2026-08-19: equity_v6_mc/nocorr/g53audit gelöscht (~43G frei). Behalten: v6, v7_fit, v8_oracle, v19.\\\n- CUDA OOM bei 8 Paralleltischen → CPU-Fallback ist normal/erwartet\"],[0,\".\\\n\\\n#\"]],\"start1\":2594,\"start2\":2594,\"length1\":233,\"length2\":206}]"
metadata_diff: {"new":{},"deleted":[]}
encryption_cipher_text: 
encryption_applied: 0
updated_time: 2026-08-19T12:46:04.876Z
created_time: 2026-08-19T12:46:04.876Z
is_locked: 0
type_: 13