A 170-goal agent field test costs $0.49. Proving it actually passed costs more.

Zero failures is a claim about the harness before it is a claim about the agent, and the v0.1.0 run of PlannerCritic ^[6] had four separate ways to return a number that meant nothing. Eighty-eight percent of the assertion files were in a format the runner did not recognise, and it returned 0/0 rather than an error ^[23][1]. The budget and replan dimensions also reported 0/0, because dispatch passed five arguments to a function that takes four ^[9]. The in-memory SQLite store reset between runs, losing cross-dimension state ^[11]. The results parser read `trace.get("status")` instead of `trace["result"]["status"]`, converting five passes into reported failures ^[10]. On the author's own breakdown, four of the ten findings were in the measuring apparatus rather than the planner ^[25]; exactly one was the sort of defect a conventional test suite would have called a failure ^[8].

That is the reason a green sweep needs something other than a pass rate to be readable. Three of those four bugs produced zeros or inversions that no summary line would flag. What distinguishes a clean pass from a silent one is the denominator: assertions evaluated per dimension, published next to the result, plus at least one goal that is supposed to fail and whose failure the rig has to report. Neither release note describes such a canary ^[3][17].

The economics are more interesting once you divide. v0.1.0 spent about $0.0019 per goal ^[19]; v0.2.1 spends about $0.0029 ^[20]. Per-goal cost rose roughly half while the suite grew about eight percent ^[21], and the source does not say why, though v0.2.1 added a live-critic boundary evaluator and an operational benchmark ^[17]. More usefully: the first run cost about three cents per issue found ^[22], and that metric goes undefined the moment the gate starts returning zero. What replaces it in the source is the assertion that a sweep is cheaper than debugging one production incident caused by a bad plan ^[27], which is not a number you can check.

Two findings put a floor under the price and a ceiling over the fix. Local Qwen3.5-4B and 9B models could not produce structured JSON ^[14], so the run cannot be moved off a paid API to get to zero marginal cost. And GPT-4o reproduced the same defect patterns as gpt-4o-mini ^[16], so the planner gap is prompt and schema work, not a billing decision. Meanwhile the critic is non-deterministic: strict goals re-run at `revision_cap=4` all escalated, with different blockers on each revision ^[15].

The accounting deserves one more look. v0.2.0 is described as 170 goals across 40 domains against v0.1.0's 157 across 35 ^[3][7], with 14 new goals in five new enterprise domains and three new adversarial-policy goals ^[18]. Seventeen additions on a base of 157 gives 174, not 170, leaving four goals unaccounted for ^[26]. The whole 10-to-0 arc ^[5] assumes the two suites are comparable, and a four-goal discrepancy in the published totals is the same class of problem as the assertion files: the number is stated, and nothing in the write-up demonstrates it was counted.