Who catches the signals?

Seven decision models tested on five buying signals from Gong Call Playbook. Each model made 1,000 decisions on test examples.

Best accuracy

82.1%

Model comparison

Macro average across five tasks · 0.5 decision threshold

Accuracy by model

Each bar is the unweighted average of five test-set accuracy scores. The y-axis is truncated to make close scores easier to compare.

Full results +

All model metrics and median latency, sorted by the selected chart metric
Model	Accuracy	Precision	F1	Recall	AUROC	ECE ↓	p50 ms
Mo	0.821	0.830	0.721	0.708	0.918	0.113	117
Luna decisions	0.815	0.761	0.757	0.815	0.884	0.129	245
Perplexity decider	0.814	0.853	0.703	0.664	0.917	0.115	376
GPT-6.1 Sol	0.787	0.703	0.743	0.875	0.920	0.183	3,140
Mo Fast	0.774	0.717	0.718	0.824	0.904	0.160	120
Jev latest	0.773	0.706	0.728	0.840	0.899	0.153	210
Clef	0.769	0.775	0.625	0.590	0.882	0.160	394

GPT-6.1 Sol ran with eight workers; other models ran with one. Latency reflects different concurrency loads and excludes retries.

Methodology +

Five binary tasks from Gong Call Playbook contribute 200 test rows each. No training rows, call excerpts, or individual predictions are published here. Dataset revision d83dd6557965.

Models return p(signal present). A fixed 0.5 threshold determines the positive class. Reported classification metrics are unweighted means of five task scores. ECE uses ten probability bins; lower ECE is better. No threshold was tuned on test data.

Latency percentiles include each successful HTTP attempt only, excluding retries and backoff. This measures signal detection in excerpts, not full-call coverage.