T16 — Before and after, run on our own deployment

From: docs/briefs/direction__one-rule-a-stranger-can-paste-is-the-way-in.md §3.2, §4.2 · Size: a day
Status: done, v1.37.0 (8 October 2026)

Why

The memo's liking it stage is the moment a person sees the line change what their agent did. A rule page with the line and no evidence asks to be trusted. The before and after is the evidence, and it is the only part of a rule page that is measured.

What to build

• A method page, site/rules/method.html: what a run is, on which deployment, with which model,

  what one run can and cannot show, and why a line is an expectation however well it behaves.

• One run per rule, for RM-R0001 to RM-R0004: the same request to a Claude Code session we are

  entitled to run, once without the line and once with it in the project instructions. The request asks the agent to do the thing the rule is about, and stops before anything irreversible: the schedule rule asks for a schedule and the run ends at the agent's answer or its approval prompt, never at a schedule that fires.

• Recorded in the rule's JSON under before_after: date, model as the session reports it, the

  deployment (claude-code-web), the request, the two answers excerpted, and one sentence of what changed. The page renders it; the status moves to run.

Rules

• Only on our own deployment. Never on somebody else's account, tenant or system.
• Nothing irreversible is exercised to prove it exists; the run stops at the agent's reply.
• One run each, said as one run. No percentages, no works every time.

Done when

The four rules show a dated before and after, the method page explains the limits in plain words, and the status on each is run.