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) from the catalog in .github/workflows/shared/graders/README.md.
Why distinct from existing built-in graders: gh-aw's built-in graders (tool-success-rate, retries, loops, trajectory-efficiency, execution-step-count, execution-duration, working-set-rebuild-factor, context-growth, artifact-production) measure aggregate efficiency/reliability signals, not why a run failed to meet its objectives. exploitation-error complements the already-shipped exploration-error: it isolates runs where the agent had sufficient evidence (observations ≥ distinct states visited) but still failed one or more declared objectives, and quantifies how much of that evidence was never consumed by a later action. This is mutually exclusive with exploration-error (which covers the insufficient-search case) on the same trace.
Required IR fields: objectives[].satisfiedAtEventIndex, events[] (kind state_change, ref), states[], observations[].consumedByActionIds.
Behavior:
All objectives satisfied → score 0.
Objectives unmet but exploration was insufficient → not-applicable (defers to exploration-error).
Objectives unmet and exploration sufficient → unusedObservations / observations, clamped [0,1]; lower is better.
Missing objectives/states/observations → not-applicable with explanatory message.
See .github/workflows/shared/graders/README.md for the full catalog — 9 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) from the catalog in.github/workflows/shared/graders/README.md.Why distinct from existing built-in graders: gh-aw's built-in graders (
tool-success-rate,retries,loops,trajectory-efficiency,execution-step-count,execution-duration,working-set-rebuild-factor,context-growth,artifact-production) measure aggregate efficiency/reliability signals, not why a run failed to meet its objectives.exploitation-errorcomplements the already-shippedexploration-error: it isolates runs where the agent had sufficient evidence (observations ≥ distinct states visited) but still failed one or more declared objectives, and quantifies how much of that evidence was never consumed by a later action. This is mutually exclusive withexploration-error(which covers the insufficient-search case) on the same trace.Required IR fields:
objectives[].satisfiedAtEventIndex,events[](kindstate_change,ref),states[],observations[].consumedByActionIds.Behavior:
exploration-error).unusedObservations / observations, clamped [0,1]; lower is better.See
.github/workflows/shared/graders/README.mdfor the full catalog — 9 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