8000 0.3.0: show the model that really runs, model ceiling, measured cache benchmark by dominicrico · Pull Request #4 · dominicrico/jev-router · GitHub
[go: up one dir, main page]

Skip to content

0.3.0: show the model that really runs, model ceiling, measured cache benchmark - #4

Merged
dominicrico merged 2 commits into
mainfrom
feat/ran-model-ceiling
Oct 9, 2026
Merged

dominicrico merged 2 commits into
mainfrom
feat/ran-model-ceiling

Conversation

@dominicrico
Copy link
Copy Markdown
Owner
  • The band shows the model that ran the last step and flags a mismatch with the pick; shows the session model after a Jev failure; flips warm to cold on its own.
  • Model ceiling and floor with a confidence break, off by default. A sonnet ceiling finished 20/20 on the hard fixture at half of opus's cost.
  • A real long-session benchmark (one process per session through stream-json, because per-turn --resume loses the mod's state). It contradicts the earlier simulation: the cache guard cuts switches 6.3 to 1.0 and cache re-writes 194k to 78k tokens, but routing still cost +24% to +38% against always-opus in a long warm session. README, charts and site updated to say so.
  • 49 tests, typecheck clean.

dominicrico and others added 2 commits October 9, 2026 16:45
…ith a confidence break

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
… and images; 0.3.0

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
@dominicrico
dominicrico merged commit a38d8bf into main Oct 9, 2026
2 checks passed
@dominicrico
dominicrico deleted the feat/ran-model-ceiling branch October 9, 2026 14:57
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

0