Repository navigation
docs: benchmarks on Opus 4.8 and GPT-6 Luna - #17
Merged
Merged
Conversation
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
There was a problem hiding this comment.
🟢 Approval recommended
The documentation-only changes have two minor clarity issues that do not block approval.
2 open issues
What changed in this PR
Extends the version 7 benchmark documentation with Claude Opus 4.8 and GPT-6 Luna results, including negative findings.
Changes:
- Adds 14-case suite results with two judges per model.
- Documents GPT-6 Luna’s synthetic-scenario losses and two prompt variants.
- Summarizes the new results in the README.
| File | Description |
|---|---|
| README.md | Summarizes additional model results and the negative outcome. |
| docs/RESEARCH.md | Adds benchmark tables, prompt experiments, and limitations. |
Review effort: Balanced
|
|
||
| ### GPT-6 Luna on the 40 synthetic scenarios | ||
|
|
||
| Here version 7 loses to no skill: 30 to 7 with MiMo V2.6 Pro (2 ties), 27 to 13 with Claude Haiku 5.5. Usefulness 4.63 without the skill, 4.48 with it (MiMo); expected behaviours met 69% against 58%. The skill halves the replies (median 834 to 400 characters), reading cost improves (3.97 to 4.69) but actionable (4.69 to 3.90) and safe (4.49 to 4.21) fall. The judges' notes on the lost cases point to the same gaps: the technical part deferred to a later turn, fewer checks and warnings, a missing detail that unblocks the request. |
|
|
||
| Here version 7 loses to no skill: 30 to 7 with MiMo V2.6 Pro (2 ties), 27 to 13 with Claude Haiku 5.5. Usefulness 4.63 without the skill, 4.48 with it (MiMo); expected behaviours met 69% against 58%. The skill halves the replies (median 834 to 400 characters), reading cost improves (3.97 to 4.69) but actionable (4.69 to 3.90) and safe (4.49 to 4.21) fall. The judges' notes on the lost cases point to the same gaps: the technical part deferred to a later turn, fewer checks and warnings, a missing detail that unblocks the request. | ||
|
|
||
| Two prompt variants were tried on the same replies: one sentence saying that "answer first" is an order and not a length, and a second that also narrows the three-sentence rule to purely factual answers. Replies grew by 17% and 6% and the expected behaviours rose to 60% and 64%, but the baseline still won 31 to 7 and 28 to 10. Neither variant is in `SKILL.md`. |
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
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.

Adds the 14-case suite results (no skill vs version 7) on Claude Opus 4.8 and GPT-6 Luna, two judges each, and the negative result on the 40 synthetic scenarios with GPT-6 Luna, with the two prompt variants tried. Stacked on #16.
🤖 Generated with Claude Code