Capabilities / C · Data and program

C.7 Program verification Active

Given a program such as a grader or reference solution, the agent audits it for defects (for example a grader that records but does not veto, or an outdated reference solution) and produces a defect list. Graded by pairing the defect list against the reference.

Evidence: planned. Only planned benchmark cards list this capability so far; no frozen benchmark version exists for it.

Domain
C · Data and program
Evidence form
program
Veto constraints
—
Depends on
A.3 Discrepancy detectionC.6 Deterministic program implementation
Required by
E.5
Internal task families
traj_suite_audit
External anchors
—
Version
0.1.0 · 0.1.0

Definition in detail

A grader that records a violation but does not veto, or a reference solution built on an outdated rule, silently corrupts every score computed with it.

Inputs a task gives

  • The program under audit, such as a grader, validator or reference solution
  • Its rule specification
  • Sample inputs and outputs
  • Version history of the rules

Artifact required

A defect list: location, defect_type (for example records but does not veto, outdated rule, boundary error, unreachable branch), a triggering input, and expected versus observed behaviour.

How it is graded

A program pairs defect entries with the reference defect set, scoring recall and false positives. Triggering inputs may be executed to confirm that the stated behaviour occurs.

In scope

  • Semantic defects where code runs but enforces the wrong rule
  • Boundary, version and veto-logic defects
  • Providing a triggering input for each defect

Out of scope

  • Writing the fixed program - not part of this capability; X.3 applies in review tasks
  • Inconsistencies in data rather than code - graded under C.4

Typical failures

Expected failure patterns; not measured rates.

Benchmarks

Grading references

No grading reference is mapped to this capability.

Levels by model and harness

No verified results yetLevels would be derived from site-verified result records; methodology 0.1.0 sets no level thresholds, so the level shows not issued.