Turn findings into the next clear step.
When a required check cannot run, Tbench Certified withholds the mark and hands back the reason: which checks blocked, what each one measured, and the fix that would resolve it. The next submission is measured on its own evidence.
A refusal is a finding you can act on.
The outcome is withheld, and the evidence for withholding it is handed over in full. Three parts of the record guide the next submission.
- 01The checks that blocked, keyed and namedThe outcome names each blocking required check, its key and the evidence it measured.
- 02The fix for each one, where a fix is knownFindings include the recorded remedy. An unrecorded remedy stays identified for reviewer follow-up.
- 03The evidence the rest of the audit did gatherCompleted checks remain available with their verdicts, durations and details, alongside the required coverage blockers.
Two ways a required check fails to run.
A skipped check and an errored one block for the same reason -- neither measured anything -- and the record keeps them apart because the fix differs. This is the shape of the blocking-check record.
- 01Did not run
Tests fail on a partially correct solution
tests.fail_on_partialA partial-solution fixture was unavailable, so the check could not measure partial-solution rejection.
Add the missing partial-solution fixture, then re-submit for an independent audit.
- 02Errored
Reward is deterministic across repeats
reward.determinismA verifier dependency was absent from the audit image. The check errored before producing reward evidence.
Install the verifier dependency at build time, then repeat the determinism check.
The report distinguishes recorded remedies from findings that still need a reviewer to define the next step.
Remediation is its own piece of work.
A refused set leaves findings open, and the console holds them in a queue worked by remediation authors. The repair scope is agreed separately, then the revised source enters a new audit.
Tbench AI keeps the submitted source read-only, so remediation is explicit. The work is to change the set -- or the verifier -- that produced the finding, in the submitter’s own repository, and to bring it back as a new submission when it is ready to be measured again.
The billing terms are the same on both sides of a refusal: an audit is charged for the measurement it performed. Remediation, where an organisation asks for it, is quoted and engaged separately.
Re-submit the repaired source.
Repair the blockers, submit a new revision and follow the independent audit result. Issuance depends on the new evidence and complete required coverage.
- 01
A revised submission receives a new audit under its recorded methodology revision, with a verdict based on its own evidence.
- 02
The original outcome remains in the register when the task is repaired and re-submitted.
- 03
Audit charges fund measurement. Passed and refused audits use the same published task rate.
Holding a refusal?
Check the identifier against the public register to see the recorded outcome and status, or write to the team about remediation scope.
support@tbench.in