You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Implements exploitation-error — Tier 2, rank 4 in the catalog (the highest-priority Not started grader, walking Tier 1 → Tier 2 → Tier 3 top-to-bottom).
What it measures: For runs that left declared objectives unsatisfied, whether failure was due to misusing evidence already gathered rather than insufficient search. It computes unusedObservations / observations (an observation is "unused" if consumedByActionIds is empty/missing), but only when the run gathered enough evidence (observations >= distinctStatesVisited). When evidence was insufficient it reports not-applicable and defers to exploration-error, so the two graders partition failed-run causes without double counting. Runs with all objectives satisfied score 0.
Distinctness from existing built-ins:tool-success-rate and trajectory-efficiency measure call outcomes/redundancy, not whether gathered evidence was actually acted upon. This grader is the complement of exploration-error (already implemented) — together they attribute objective failure to insufficient search vs unused evidence.
Required IR fields:objectives[].satisfiedAtEventIndex, observations[].consumedByActionIds, events[] (kind state_change, ref) with fallback to states[].
See .github/workflows/shared/graders/README.md for the full catalog — 8 of 25 implemented after this PR.
Caution
Protected files were modified in this change.
This pull request is in
request_reviewmode and requires explicit human scrutiny before merge.Protected files:
README.md,.github/workflows/shared/graders/README.md,.github/workflows/shared/graders/exploitation-error.mdImplements
exploitation-error— Tier 2, rank 4 in the catalog (the highest-priorityNot startedgrader, walking Tier 1 → Tier 2 → Tier 3 top-to-bottom).What it measures: For runs that left declared objectives unsatisfied, whether failure was due to misusing evidence already gathered rather than insufficient search. It computes
unusedObservations / observations(an observation is "unused" ifconsumedByActionIdsis empty/missing), but only when the run gathered enough evidence (observations >= distinctStatesVisited). When evidence was insufficient it reports not-applicable and defers toexploration-error, so the two graders partition failed-run causes without double counting. Runs with all objectives satisfied score 0.Distinctness from existing built-ins:
tool-success-rateandtrajectory-efficiencymeasure call outcomes/redundancy, not whether gathered evidence was actually acted upon. This grader is the complement ofexploration-error(already implemented) — together they attribute objective failure to insufficient search vs unused evidence.Required IR fields:
objectives[].satisfiedAtEventIndex,observations[].consumedByActionIds,events[](kindstate_change,ref) with fallback tostates[].See
.github/workflows/shared/graders/README.mdfor the full catalog — 8 of 25 implemented after this PR.Note
This was originally intended as a pull request, but the git push operation failed.
Original error: The process '/usr/bin/git' failed with exit code 1
Workflow Run: View run details and download bundle artifact
The bundle file is available in the
agentartifact in the workflow run linked above.Create the pull request manually