sample-loop

sample-loop — the v0.6 inner loop, on a real artifact

Generated by:

node bin/ed-agent.mjs --refine examples/sample-loop/input-artifact.md --mission code --target 6

• input-artifact.md — a deliberately messy payment-retry handler: a hardcoded secret, an empty catch, AI-tone filler, and an unquantified claim.
• loop-report.md — the loop's output. Severity falls 47 → 18 over four iterations as it fixes the worst finding each pass (secret → empty-catch → AI-tone filler), then plateaus and stops. The residual — an unquantified claim and the marketing "tell" words a deterministic pass should not rewrite — is surfaced and escalated to a human, never silently "resolved."

What this demonstrates:

• The loop is real and deterministic (~0 token): the verifier is the same assessor suite the rest of Ed Agent uses; the producer here is the in-harness remediator.
• It converges to the floor of what deterministic fixes can prove, then stops — the four safety mechanisms (severity gate, overshoot rollback, iron-law hard-halt, budget fuse) mean it can neither run forever nor ship a worse artifact than it started with.
• exit 2 because the residual is above the target — useful as a CI gate.

Wire a non-deterministic host LLM as the producer and the overshoot-rollback guard is what keeps it honest; the verifier, rollback, iron-laws and budget stay deterministic.