inc-0006
data-integrity
citable URL: https://halobench.com/records/inc-0006/ — this address never moves; the anchor /records/#inc-0006 keeps resolving
HG-002 smoke parser rejected a valid task with no required tool call
detected
2026-08-19 · window 2026-08-19T12:12:22Z → 2026-08-19T12:18:31Z · by summary-review-against-copied-tau2-results · lag immediate at arm close
signature: Five rewards and terminations, six valid nonempty tool calls, no empty turns, no infrastructure errors and mean 0.80, but rc 1 solely because task 1 had zero tool calls.
should have caught it
The parser fixture should have included a successful task whose scenario legitimately terminates without a tool call.
diagnosis
initial: invalid smoke because one task lacked a nonempty tool call
actual: The non-vacuity rule applies to the smoke as a whole, not every task. Task 1 completed successfully without needing a tool; the harness parser produced a false negative after the scientific result was already complete.
false-path cost: n1 was not run in the first paired root; a corrected paired retry followed
falsified by: copied results JSON, summary and corrected positive fixture
actual: The non-vacuity rule applies to the smoke as a whole, not every task. Task 1 completed successfully without needing a tool; the harness parser produced a false negative after the scientific result was already complete.
false-path cost: n1 was not run in the first paired root; a corrected paired retry followed
falsified by: copied results JSON, summary and corrected positive fixture
blast radius
configs affected: cfg-0122
lesson
Capability validity gates must test whole-smoke non-vacuity unless the task contract itself requires a call. Preserve a scientifically complete result separately from the harness exit status and do not claim the arm formally passed until the corrected parser closes rc 0.
state
resolved · class code
Cited by — computed at build time, never stored
claims clm-0088