fix(bench): reliable aksops → valid iso-accuracy ops result (~neutral)#17
Merged
Conversation
The aksops JIT arm was flaky (2/6 success) because the mock az/kubectl were one shared, obviously-fake stub that the agent sometimes refused to run. Fix: distinct az and kubectl test doubles (different bodies, headers stating they're the intended local tools) and a prompt that frames the task as a SIMULATION where running the runbook IS the deliverable — removing the "these are fake, why bother" escape hatch. Result — now both arms succeed 6/6 (100%), so the comparison is finally valid at iso-accuracy: baseline 100% mean 293,418 (med 332,396) vs jit 100% mean 308,486 (med 308,858) -> saving -5.1% by mean (JIT slightly WORSE); ~neutral by median. High baseline variance (one 141k rollout) => genuinely noisy, no clear signal either way, but definitely no JIT win. So even on the ops/tool-use shape, at true iso-accuracy, a compiled skill does not clearly pay off. The earlier +14% was a reliability artifact. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
After hardening the fixture (both arms 6/6), the valid iso-accuracy number is ~neutral (−5% mean / +7% median, noisy), not the retracted +14%. Updates the TL;DR and aksops section: across all four shapes the study finds NO clear JIT win — code-edit shapes neutral-to-negative, ops shape ~neutral once measured reliably. A win, if any, likely needs workflows with much larger per-episode work a compiled action can wholesale replace. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Makes
aksopsreliable so the ops question can be answered at iso-accuracy — and answers it.Fix: the JIT arm was flaky (2/6) because
az/kubectlwere one shared, obviously-fake stub the agent sometimes refused to run. Now they're distinct test doubles and the prompt frames the task as an explicit simulation where running the runbook is the deliverable.Result (rollouts=6, both arms now 100%):
Verdict: even on the ops/tool-use shape, at true iso-accuracy, a compiled skill does not clearly pay off. The earlier +14% (retracted in #16) was a reliability artifact. Across all four shapes the study now finds no clear JIT win.
Includes the findings-doc update (TL;DR + aksops section).
golangci-lintclean; aksops unit tests green.🤖 Generated with Claude Code