← Lab Measured run2026-08-20

On real traces, does it agree with human experts?

It flags 87.2% of the traces human experts marked as broken, so it reliably tells you which transcripts to open. It describes them in its own vocabulary — 2 of the experts' 12 categories — and it does not pin down where inside a trace the failure sits.

Trace triage validated
87.2%broken traces flagged (34/39)
2/12expert categories reproduced
18.8%right span, vs 21.4% by chance
$2.14for all 40 traces

What we set out to measure, and what we measured

TRAIL taxonomy coverage (of 12 categories occurring >= 3x)>= 0.52/12under the bar
trace-level recall (traces with >= 1 annotated error)>= 0.70.872at or above the bar
trace-level Cohen's kappa>= 0.3-0.043under the bar

What we did

TRAIL pairs 148 real agent traces (GAIA, SWE-Bench) with 841 expert-annotated errors. We read 40 of them cold — the annotations were never visible to any stage — then matched our modes to theirs with a different model family.

What came back

Triage works: given a pile of traces it points you at the ones worth opening. Two limits came out of the same run. Its own modes line up with 2 of the 12 expert categories, and the turn it cites falls inside the annotated span 18.8% of the time — against 21.4% for a random pointer, so the turn index carries no signal.

What it changed

The product stopped implying it can name the turn. Evidence is quoted text you can search for, not a span index dressed up as a coordinate.

What it doesn't show

TRAIL is error-dense — 1 of 40 judged traces has no annotated error — so specificity is not estimable here at all, and 20 traces were clamped to fit the render budget.

MODEL google/gemini-3.7-flash · SAMPLE 40 traces, seed 42 · COST $2.14
REPRODUCE npm run validate:trail

Full metric tables and every caveat

Run it on your own traces.

Free, no account, parsed in your browser.

Upload a trace file