Instruction scenarios

Four small scenarios exercise the engineering instructions. They are examples to run when relevant, not an automated qualification suite. Each has seed files, a normal request, and a few observable expectations. Seed fixtures contain deliberate defects and stay outside the root source/format gates.

Scenario	What to check
Scope fix	A small correct fix without adjacent cleanup; expectations.
Design only	A useful, compatible timeout design without implementation; expectations.
Review a change	An evidence-backed authorization finding with no edits; expectations.
Near miss	A concise explanation without engineering ceremony; expectations.

Run and inspect

Use the disposable scenario recipe. Copy only the chosen fixture/ contents into a fresh linked worktree, keep the expected answer outside it, and launch Pi there with candidate instructions and a separate session directory. Obtain approval before paid model runs; start with the primary engineering model.

Inspect the native session, changed and untracked files, and actual check output against the expectations. Record the instruction revision and working diff, model/runtime, observed outcome, and limits beside the saved evidence. A correct review can demonstrate a failing test; the scope fixture deliberately has an unrelated failure. Do not score exact headings, tool-read patterns, model names, or worker counts.

Report what happened on that model and configuration, not blanket qualification. These four scenarios do not cover every specialist method, delegation pattern, durable plan, or worker boundary. Historical results remain historical; the retired thirteen-case runner, activation scoring, automatic judge, and qualification voting are not maintained.

Shared and runtime checks

instruction-root.ts projects candidate instructions without changing the live agent directory; process.ts is the shared process helper. Their consumers and useful structural checks remain tested offline.

Tool-surface evals and Goal-reader evals are separate runtime checks. Their runners, package-owned graders, and comparison rules are unchanged by this simplification. bun run test:eval runs the surviving helper, tool-surface, and package grader tests; it makes no model calls.