# Modonome on this demo app: captured dry-run + maker/checker cycle

The dry-run output is verbatim tool output. The maker and checker cycle
described below is a real run whose events are recorded in
`runs/2026-06-26T11-46-00Z/metrics.jsonl`; the raw model output logs
were not committed (they are large and reproducible on demand; see
the Reproduce section in that directory's README).

## Step 1: the dry-run sweep

`node scripts/dry-run-sweep.mjs examples/demo-app` reports the posture and the bounded
work it would propose. The full verbatim output is in runs/2026-06-26T11-46-00Z/dry-run.txt and in
../dry-run-transcript.txt. The sweep changes nothing and refuses remote spend by default.

## Step 2: a maker proposes a bounded, test-only change

OrderService.refund has four branches (order not found, already refunded, wrong status,
and the success path), but the committed test suite covers only the order-not-found
branch. A maker running claude-haiku-4-5 was asked to add tests for the three uncovered
branches, adding assertions only and changing no source file. The maker_run event is
captured in runs/2026-06-26T11-46-00Z/metrics.jsonl.

## Step 3: an independent checker reviews it

A checker running claude-sonnet-4-6, a distinct model that did not author the proposal,
reviewed it. It approved and raised one real question: the success-path assertion is only
correct if the fake payment client returns exactly {status: "refunded"}, so that contract
should be confirmed before merge. The checker_review event is captured in
runs/2026-06-26T11-46-00Z/metrics.jsonl.

## Step 4: the change is recorded, not force-applied

The committed sample keeps its planted debt so the corpus and the CI negative control still
have something to detect. The cycle is recorded as evidence in runs/2026-06-26T11-46-00Z/.
The maker and checker identities and models differ, as the separation-of-duties rule requires.