## llms.txt catches up with the insurability pitch · voice feedback interview prompt

- **The machine-readable surface was a release behind.** `llms.txt`, `llms-full.txt` and
  `.well-known/agent-content.json` still carried "Agents act. You own the risk" with zero mentions
  of insurability, carriers or the Index — so every agent reading us got the superseded pitch while
  the site itself led on insurability. Fixed at source in `src/data/site-content.json`, which
  generates all three.
- The concept set is rewritten and grown from six to eight: the insurability layer, the Insurability
  Index (five weighted dimensions, six levels), grant-is-not-mandate, no-deny, never-in-the-request-
  path, accepted-is-not-acceptable, the SG/Vault foundation, and positioning. The honest caveats
  travel with them — weights are underwriting judgement rather than actuarial fact; we define the
  mandate and map the gap but do not enforce it; the EU AI Act point is factual, not legal advice.
- **A smoke group now guards it**, so the machine surface cannot silently drift from the live
  positioning again.
- **Build ordering fixed.** The llm artifacts were written *after* the smoke suite ran, so a failing
  smoke left them stale — and the next run then validated yesterday's file. They are generated
  outputs that the smoke inspects, so they are now written first. This was a genuine
  chicken-and-egg: the brand assertion kept failing against a file the failure itself prevented from
  regenerating.
- Three stale two-word brand strings fixed in `assets/library/library.json`.
- **New: `team/prompts/voice-feedback-interview.md`** — a prompt for ChatGPT to run a voice-mode
  feedback interview with people who have seen a RiskMandate presentation, and produce a debrief.
  It reads the site first, treats the homepage as authoritative, keeps turns short for voice, leads
  the debrief with objections rather than praise, asks consent to quote explicitly, and instructs
  the interviewer that most capability is proposed rather than shipped.