Curve Parameter Sweep Results (2026-06-27)

# Curve Parameter Sweep Results — 2026-06-27

## Overview
Three rounds of parameter sweeps on the Gen 4 curve config (cash_nl_g4_live).
Each round: one parameter varied at a time, tested vs both flock (tight) and chump (loose) fields.

## Round 1 (2 seeds × 5K hands)
Baseline: normal-tuning thresholds + curve + narrowing + ts=1.0

Top improvements (both fields positive):
| Parameter | Change | Combined |
|-----------|--------|----------|
| bet_turn_base | 0.50→0.55 | **+459** |
| raise_base | 0.75→0.80 | **+71** |
| raise_w_protection | 0.25→0.35 | **+61** |
| scurve_floor | 0.03→0.05 | **+59** |
| bet_flop_base | 0.35→0.30 | **+54** |
| bluff_hs_max | 0.10→0.05 | **+53** |
| scurve_steepness_mult | 0.5→0.3 | **+45** |

## Round 2 (3 seeds × 5K hands)
Baseline: round 1 winners applied

Top improvements:
| Parameter | Change | Combined |
|-----------|--------|----------|
| **range_tightness_scale** | **1.0→0.7** | **+112** |
| range_tightness_scale | 1.0→0.9 | **+61** |
| bet_turn_base | 0.55→0.58 | **+58** |
| narrow_center_offset | 0.1→0.2 | **+46** |
| range_looseness | 0.10→0.05 | **+21** |

## Round 3 (3 seeds × 5K hands)
Baseline: round 2 winners applied. Targeted at live hand analysis findings (too loose calling draws).

Top improvements:
| Parameter | Change | Combined |
|-----------|--------|----------|
| range_tightness_scale | 0.7→0.8 | **+31** |
| strong_hands_fraction | 0.25→0.30 | **+18** |
| semi_bluff_max_bet_ratio | 0.35→0.30 | **+16** |
| call_w_draw | 0.80→0.75 | **+16** |
| call_ppot_base | 0.32→0.25 | **+8** |

Notable: call_w_draw 0.70 was WORSE (-10), 0.60 was neutral. Sweet spot is 0.75.

## Final Live Config (cash_nl_g4_live.toml)
```
strong_hands_fraction = 0.30
call_base = 0.45
call_ppot_base = 0.25
call_w_draw = 0.75
call_w_pot_odds = 0.70
bet_flop_base = 0.30
bet_turn_base = 0.58
raise_base = 0.80
raise_w_protection = 0.35
bluff_hs_max = 0.05
semi_bluff_max_bet_ratio = 0.30
range_looseness = 0.05
range_use_scurve = true
scurve_steepness_mult = 0.3
scurve_floor = 0.05
narrow_center_offset = 0.2
range_tightness_scale = 0.8
```

## Live Hand Analysis Findings
Stack went from 7.4M → 5.0M (240 BB loss). Key leaks:
1. Called pot-sized turn shove with Q-high + gutshot (ThQh): -280K
2. Called 150K turn bet with open-ender at 2:1 (65s): -200K
3. Called 240K river shove with K8 top pair bad kicker (Kc8d): -120K
4. Overbet flop with A-high then shut down (AhQd): -90K
5. JJ top set vs straight cooler: -361K (unavoidable)

Root cause: too loose calling big bets with draws. Round 3 addressed this with call_w_draw 0.75.

## Linear vs Curve
- Linear consistently wins by ~+20 combined (small, <1%)
- User prefers curves for live play, kept


id: d666f66c3d364aadbc17a28265ea8481
parent_id: 2c8da247905946c3aa19eb4936e16323
created_time: 2026-06-27T08:07:52.108Z
updated_time: 2026-06-28T05:18:11.598Z
is_conflict: 0
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: 1782547672108
user_created_time: 2026-06-27T08:07:52.108Z
user_updated_time: 2026-06-27T13:48:10.766Z
encryption_cipher_text: 
encryption_applied: 0
markup_language: 1
is_shared: 0
share_id: 
conflict_original_id: 
master_key_id: 
user_data: 
deleted_time: 1782623891598
type_: 1