Harness Lab

the harness

step budget 20 turns

Too tight truncates real work; too loose funds doom loops.

verifier

The only thing standing between a plausible patch and a wrong fix in production. Costs tokens; strict also false-rejects.

shape observations
Return the error region, not the 2MB log. compaction
Caps the window — and sometimes drops a load-bearing fact. typed tool errors
A failure that carries information breaks retry loops. no-progress detector
Stop after 3 dead turns and hand back findings.

—pass rate

—wrong fix shipped

—gave up / burned out

—tok / solved ticket

—avg tok / ticket

—p95 turns

— — —

run a batch to see where you stand

Every config you run leaves a dot. Cyan = clears the gate, amber = latest. The shaded corner is shippable. You are hand-exploring a Pareto frontier.

trajectory · one ticket, turn by turn