README.md

Backbone Observability Gate Ratio Sweep

This experiment compares multiple Layer 2 metric families on the hard branch:

• legacy_random_5seed
• one_shot_random
• persistent_mode_informed

It uses the Layer 2 trial table from:

• backbone_observability_gate_informed_bank_trials.csv

For each method:

1. choose threshold and direction on calibration
2. freeze that rule
3. score holdout
4. score confirmation
5. rank metrics by out-of-sample mean balanced accuracy

The compared metric families include:

• raw Layer 2 signals
• spread-normalized ratios
• support-normalized ratios
• mixed spread-and-support ratios

This is a comparative test harness, not a claim that every ratio here is a valid control law.

Main Result

For the informed-bank Layer 2 regime, the strongest tested candidate in this sweep is:

• ratio_candidate_times_anchored_span_over_std

That is:

• mean_candidate_count * mean_anchored_alpha_log_span / mean_anchored_alpha_log_std

With calibration-frozen thresholding on persistent_mode_informed, it reached:

• calibration balanced accuracy: 0.6818
• holdout balanced accuracy: 0.7500
• confirmation balanced accuracy: 0.6667
• out-of-sample mean balanced accuracy: 0.7083

That makes it the strongest current Layer 2 implementation candidate for the informed-bank regime in this repo.

Outputs

• backbone_observability_gate_ratio_sweep_leaderboard.csv
• backbone_observability_gate_ratio_sweep_summary.json
• backbone_observability_gate_ratio_sweep_top_metrics.png