eval-builder
Installation
SKILL.md
Eval Builder (legacy Pentagon entrypoint)
First output is a harness plan. Do not start with unit tests or a raw proof matrix. Pentagon work asks: “what proof would survive a hostile optimizer?”
Current eval canon is core/eval, docs/specs/spec-eval.md, and
dna/core/eval.dna.yaml. core/testing, lev pentagon, and
core/daemon-pentagon are legacy migration inputs, not target authority.
Naming and migration frame
Use eval and eval-builder language in new plans, tasks, docs, APIs, and
operator output. Treat Pentagon as a historical compatibility name only.
This skill has three modes: