feat(experiments/mcp): Write diagnosing-experiment-results skill and Evals#59029
Merged
mp-hog merged 5 commits intoMay 20, 2026
Merged
Conversation
mp-hog
added
the
skip-inkeep-docs
Contributor
|
🎭 Playwright didn't run on this PR — your changes touch code that could affect E2E behavior, but Playwright is opt-in via label now to keep CI cost down. Add the Most PRs don't need this. Real regressions still get caught on master and fix-forward. |
Contributor
|
Size Change: -7.37 kB (-0.01%) Total Size: 118 MB 📦 View Changed
ℹ️ View Unchanged
|
assign-reviewers-posthog
Bot
requested review from
a team,
TueHaulund,
arnohillen,
fasyy612,
ksvat and
nicowaltz
and removed request for
a team
![stamphog[bot]](https://avatars.githubusercontent.com/in/3000332?s=60&v=4){kind=small}
There was a problem hiding this comment.
Large PR (3176 lines, 16 files) with zero reviews that exceeded the size gate ceiling. The changes span new eval test infrastructure, extensive skill documentation, and auto-generated API type updates across two products — needs a human reviewer to validate correctness and scope.
Contributor
Prompt To Fix All With AIFix the following 1 code review issue. Work through them one at a time, proposing concise fixes.
---
### Issue 1 of 1
ee/hogai/eval/sandboxed/experiments/eval_midrun_changes.py:85-135
**Behavioral check bundled into `diagnosis_group` — violates the anti-bundling pattern established in `eval_interpretation_traps.py`**
The `ship_variant_under_uncertainty` case explicitly tests "should I ship?" and the case description states the correct answer is "not yet / not on this evidence." However, the behavioral check ("advises AGAINST") is embedded inside the `diagnosis_group` description text rather than enforced with a separate `AdvisesAgainstShipping` scorer + `"advises_against_shipping": True` in `expected`.
`eval_interpretation_traps.py` contains an explicit comment explaining exactly why this bundling is a regression vector:
> "Both are needed: the previous version of this case relied on `diagnosis_group` alone, where the LLM judge's 'must recommend waiting' criterion was bundled into the diagnosis description — splitting them separates content from behavior and makes regressions in either dimension visible independently."
The same failure mode applies here: the agent can identify the diagnostic correctly but still recommend shipping, and that behavioral regression would not be caught by `CitesDiagnosticGroup` alone. `AdvisesAgainstShipping` should be added to the scorers list, and `"advises_against_shipping": True` should be added to this case's `expected` dict — mirroring the pattern in `eval_interpretation_traps.py`.
Reviews (1): Last reviewed commit: "Fix CI" | Re-trigger Greptile |
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
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
2 participants
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.
Problem
Users frequently ask "why does my experiment look wrong / biased / empty?". There was no symptom-driven path from the user's complaint to the right diagnostic, and no eval coverage for this class of question.
Changes
diagnosing-experiment-resultswith a symptom-to-diagnostic-logic table and five reference files covering:How did you test this code?
Publish to changelog?
no
Docs update
No
🤖 Agent context
Authored with Claude Code across many different sessions