id: 8ffcfa5259c94e308e01fb85112ba84f
parent_id: 
item_type: 1
item_id: f3c42f4435644e19b186265dd3fbf672
item_updated_time: 1786734671497
title_diff: "[{\"diffs\":[[1,\"Improvement Ideas — Top 10 (2026-08-14)\"]],\"start1\":0,\"start2\":0,\"length1\":0,\"length2\":39}]"
body_diff: "[{\"diffs\":[[1,\"# Idea #7 (revised): Action-Likelihood Range Re-weighting (2026-08-14)\\\n\\\n## Problem\\\nKK-type overplay: overpair assessed 0.58-0.65 equity on A-high boards vs call-call-call lines (true ~0.10-0.25). Production G50 path (alpha=1.0) skips ALL board-interaction narrowing — range comes purely from RangeNet `probs_to_hand_range`, which does card exclusion only. `narrow_board_hit_boost` exists (range_builder.rs:514) but only runs in the heuristic fallback; percentile S-curve narrowing ignores draws/overcards.\\\n\\\n## Revised design (user proposal): Bayesian action-likelihood re-weighting\\\nP(hand | actions, board) ∝ P(actions | hand, board) × P(hand)\\\n- P(hand) = RangeNet prediction (prior)\\\n- P(actions | hand) = likelihood from per-combo strength profile: hs (static LUT eval percentile vs actual board, 169 evals, cheap), ppot/npot (draw classification tier or small shared-board MC via HandPotential), nutpot/rpot (GPU MC infra exists, commit f2312ab; matters for big-bet lines)\\\n- Likelihood shapes per StreetAction × bet-size bucket: CalledBet(big) → high hs or ppot+odds or nutpot; Bet/RaisedBet → hs or ppot semi-bluff + bluff floor; Checked → mid/low hs. Compounded multiplicatively per street (naive Bayes), renormalized.\\\n\\\n## Properties\\\n- Subsumes hand-crafted board-interaction classes: overcards, paired boards, straights, flushes, river overtake all appear as hs changes automatically\\\n- Board-exact — corrects RangeNet's structural blindness (board seen only via 4 texture + 10 card features)\\\n- Bet-size conditioning per-hand, not global\\\n\\\n## Integration / pitfalls\\\n1. REPLACE percentile S-curve + narrow_board_hit_boost (supersedes both), same TOML knobs\\\n2. Keep bluff floors (narrow_floor) — avoid over-narrowing\\\n3. Calibrate likelihoods from equity_v5_2m4 data (19.2M samples with showdown hands + action lines; logistic fits per action × size bucket) instead of hand-tuning 6-8 knobs\\\n4. Cost: negligible vs multiway equity MC downstream\\\n\\\n## Verification\\\nreplay_kk_overplay (#[ignore], real profiles) → Harrington 18/18 → sanity_check.sh cash_nl_g50 → sweep vs NN pool\\\n\\\n## Effort/Gain\\\n1-1.5d (+0.5d if fitting likelihoods from data). +1-3 BB/100 live, largest vs loose-passive fields. Interim until EquityNet (#1); improves #1 training targets. Generic action-likelihood refinement, not just the KK patch.\"]],\"start1\":0,\"start2\":0,\"length1\":0,\"length2\":2311}]"
metadata_diff: {"new":{"id":"f3c42f4435644e19b186265dd3fbf672","parent_id":"2c8da247905946c3aa19eb4936e16323","latitude":"0.00000000","longitude":"0.00000000","altitude":"0.0000","author":"","source_url":"","is_todo":0,"todo_due":0,"todo_completed":0,"source":"joplin-desktop","source_application":"net.cozic.joplin-desktop","application_data":"","order":1786733906375,"markup_language":1,"is_shared":0,"share_id":"","conflict_original_id":"","master_key_id":"","user_data":"","deleted_time":0},"deleted":[]}
encryption_cipher_text: 
encryption_applied: 0
updated_time: 2026-08-14T19:15:24.342Z
created_time: 2026-08-14T19:15:24.342Z
is_locked: 0
type_: 13