Eight failed checks in ORBITAL's browser gate traced back to one real app bug

The rule that keeps this gate usable is in how a failed run gets read. Every raw `failed` result is checked against the verifier's structured verdict, `verdict.confirmed`, `family`, `category` and confidence, and never against the exit code ^[16]. T7 was the only real problem in the app across the twelve tasks ^[14]. "Eight failures, one bug. Read by exit code, all eight look the same," the author wrote ^[15].

The other seven were test or environment problems. T3's test asserted a hover effect on the allocation donut without ever hovering the segment, a step the script was missing ^[11]. T4 failed because a legitimate layout fix invalidated recorded replay baselines; re-recorded with `--author`, all eight test files passed ^[12]. On T9, T10 and T11 the feature steps passed and a redundant final re-check stalled afterwards, and on T11 the browser agent oscillated between two buttons until it detected a cycle in its own plan ^[13]. T2 failed three times at three different points, one attempt opening the wrong page entirely, with triage pointing at flaky headless hover targeting and an ambiguous unlock-state assertion ^[10].

Six overrides across twelve tasks is half the build ^[28]. T2 was recommended by a second Claude session and approved by the author explicitly, T3 needed triage evidence and then the author's approval, and T4, T9, T10 and T11 were cleared "under standing self-override authorization" ^[27]. Two of the six record a per-case human decision ^[30].

All of this is twelve tasks on one app, with one browser agent on one machine. For the seven-to-one split between automation failures and app bugs to carry to another suite, that suite would need assertions that depend on headless hover targeting and recorded baselines a layout change can invalidate ^[29][10][12].

The last layer is the write-up itself. Checking each ORBITAL claim in the repo's build log, `OBSERVATIONS-AND-REPORTINGS.md`, against the raw activity log ^[20], the author found notes crediting the sweep with catching "a real state-tracking defect" on the chart at T2, where the log has no sweep finding on T2 at all ^[22]. The notes call T9 "a genuine race"; the log shows all four export toasts confirmed in the same run, with a redundant step failing afterwards ^[24]. "My write-up had done to the verifier's failures exactly what the verifier had done to the app," the author wrote ^[21]. Kane's `bug_title` field made that easy to get wrong: "Replay misclassifies unlocked chart state" looks like a finding once it is lifted away from the `confirmed: false` beside it, and the first draft of Part 1 repeated the titles ^[25].