Skip to content

Reword the plain text and comment rules - #84

Merged
kaveraa merged 1 commit into
mainfrom
fix/text-rules
Nov 13, 2026
Merged

kaveraa merged 1 commit into
mainfrom
fix/text-rules

Conversation

@kaveraa

@kaveraa kaveraa commented Nov 13, 2026

Copy link
Copy Markdown
Owner

explain-retry showed two rules giving way with the plugin on. Both are reworded, from three measured trials whose traces were read.

  • Plain text: "long answers too" and "(write -, 0-100 too)". Sonnet's dash was an en dash in a range of numbers, not in its prose.
  • Comments: "only a reason the code cannot show", "never a summary of a function or a line", "in doubt, none". Haiku opened each function with a comment that sums it up.
On explain-retry, plugin arm Before After
Sonnet 5.5, plain answer 1 / 3 5 / 5
Sonnet 5.5, comments give a reason 1 / 3 5 / 5
Haiku 5.5, comments give a reason 0 / 3 4 / 5

The ruleset ceiling goes from 600 to 630 estimated tokens. The two lines take it to 628 at its heaviest level, and nothing could be cut elsewhere without touching a wording already measured. This is the one decision in this change that is not mine to make: say so and I look for a rule to give up instead.

The nine cases with the reworded rules, three runs, Opus as judge:

Model Score with, now / before Score without
Fable 5.2 0.99 / 0.98 0.88
Opus 5.5 1.00 / 0.99 0.85
Sonnet 5.5 1.00 / 0.98 0.86
Haiku 5.5 0.98 / 0.97 0.84

Ruleset, AGENTS.md, skill, tests, README and eval card in both languages, gain card, notebook, CHANGELOG. 97 tests passing.

@kaveraa
kaveraa merged commit b333a42 into main Nov 13, 2026
5 checks passed
@kaveraa
kaveraa deleted the fix/text-rules branch November 13, 2026 17:55
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant