Assert the marker-eval per-slot storage contract in eval code, not just prose
kind: infra#agent-ok
track:Highlight text on any card (body, plan, or an event) to anchor a comment, or leave a whole-task comment here. Mention @claude to summon a reply.
No comments yet.
Marker evals stored only post-softmax log-probs, forcing paid GPU re-runs (#530 needed a re-eval pod, #531 an 80-run re-score). The storage contract (4 floats per slot) landed in CLAUDE.md and rules same day, but nothing enforces it at runtime. Action: add a code-level assertion in the eval path so a violation of the per-slot storage contract fails fast instead of being discovered after the pod is gone. source: logs/daily/2026-06-09.md, approved by Thomas 2026-06-10 ('Apply these')
Activity