AUTORESEARCH · CAUSAL RULE OPTIMIZER

How the graph learns its own thresholds

A Claude-driven loop proposes one change to train.py per iteration, runs the experiment, and either keeps or reverts the change based on F1 delta. The kept diffs are the threshold config the ZWM causal engine actually ships.

BASELINE F1

0.6994

BEST F1

0.8524

DELTA

+0.1530

+21.9%

ITERATIONS

5 kept

F1 TRAJECTORY

Each point is one experiment. Filled teal = kept, faded grey = reverted.

KEPT DIFFS (5)

iter 01

lower compliance_violation_score 40→35

F1 0.7069 · P 0.601 · R 0.859 · Δ -0.0075

iter 02

lower fitiq_risk_threshold 50→42

F1 0.7196 · P 0.639 · R 0.823 · Δ -0.0127

iter 04

lower compliance_warning_score 60→55

F1 0.7535 · P 0.698 · R 0.818 · Δ -0.0339

iter 08

raise violation contribution 0.60→0.70, lower warning 0.30→0.25

F1 0.8476 · P 0.939 · R 0.773 · Δ -0.0941

iter 13

continuous proximity feature in compliance warning zone

F1 0.8524 · P 0.950 · R 0.773 · Δ -0.0048

FULL HISTORY

#	DESCRIPTION	F1	Δ	STATUS
1	lower compliance_violation_score 40→35	0.7069	-0.0075	KEPT
2	lower fitiq_risk_threshold 50→42	0.7196	-0.0127	KEPT
3	lower violation_prediction_threshold 0.45→0.38	0.6372	+0.0824	REVERTED
4	lower compliance_warning_score 60→55	0.7535	-0.0339	KEPT
5	increase HIGH severity weight 0.25→0.40	0.7500	+0.0035	REVERTED
6	biological×compute joint interaction term (0.35 boost)	0.7535	+0.0000	REVERTED
7	amplified joint compliance+fitiq penalty (1.5x)	0.7535	+0.0000	REVERTED
8	raise violation contribution 0.60→0.70, lower warning 0.30→0.25	0.8476	-0.0941	KEPT
9	raise violation contribution 0.70→0.80	0.8476	+0.0000	REVERTED
10	raise standalone fitiq penalty 0.20→0.30	0.8453	+0.0023	REVERTED
11	lower violation_prediction_threshold 0.45→0.42	0.8453	+0.0023	REVERTED
12	explicit joint bio HIGH + compute < 0.92 crisis signal (0.55)	0.8476	+0.0000	REVERTED
13	continuous proximity feature in compliance warning zone	0.8524	-0.0048	KEPT
14	continuous proximity for FitIQ contribution	0.8524	+0.0000	REVERTED
15	lower fitiq_risk_threshold 42→40	0.8524	+0.0000	REVERTED

Loop source: zwm-autoresearch/run_loop.py · model claude-sonnet-4-6 · experiment log updates after every iteration.