# Reference card

*End matter — one page, written to be pasted into an agent session*

---

**What this is.** The insurance pivot of pki.sgit.ai, in one card. Sources: briefs v0.33.71–v0.33.81 (the memos, verbatim), `/insurance/src/` (the doctrine), the audit at v0.33.82. The transcript outranks any summary — including this one.

## The vocabulary

- **The delta** — grant minus mandate: what an agent *can* do minus what it was *authorised* to do. The insurable exposure of this whole programme, computed and never stored. The register publishes it with `"acceptor": null`.
- **A placement** — an agent, in an environment, under an identity. The unit that gets rated. A library entry *is* a placement; "rating Claude" is meaningless.
- **A rating** — a level 1–5 (settled, GM-D54) plus its derivation: inputs with evidence channels, what moved the level and which way, the twin's age, and what it does not prove. Stage 1's entire product.
- **The rule** — *a level nobody can recompute is exactly the theatre a premium would have prevented.* Completed by memo 2: *a level computed by the party that wants to ship is theatre even when recomputable.* Method and separation, both.
- **Two channels** — measured (full weight) and declared (discounted, rendered as declared) never merge; unknown is never scored as absent. The questionnaire is a declared-fact collector; the card-versus-twin gap is a rating input — and the shape of material non-disclosure.
- **Delta classes** — elective (operator can close; carries a cost-to-close, whose high region is *latent*), structural (platform's finest grain is coarser than the mandate; the pool's, or nobody's), defect (a vulnerability; temporary). A rating that does not separate them is unfair and useless in one move.
- **A policy** — a mandate-shaped statement issued by a rater. It never signs; its subject signs, and the policy says what that signature is worth. Amending, activating, ending: appends, never edits.
- **A warranty** — a fact plus a maximum age. Fails three ways: false, stale, unknown — and unknown counts as failure, the deliberate opposite of the rating rule.
- **Three clocks** — the policy interval; twin and world freshness; each warranty's check interval. A policy in force against a fresh twin with a warranty last checked eleven days ago is not covered.
- **The handshake** — the relying party refuses the request unless the subject proves possession and a policy is in force. The first mechanism in the pivot that can reach tier *boundary*, because the relying party is outside the requester's grant — where genuinely independent.

## The postures

- **Stage 1, not insurance**: a rating transfers no risk and promises no payout — which is why it needs no carrier and can be built today. Money is stage 2, and stage 2 is regulated and not this estate's.
- **Not in line**: schemas, flows, connectors, evidence — never the carrier, never the execution broker. This project can never itself be a boundary; every tier above instrumentation is somebody else's install.
- **Openness is load-bearing**: with no money at stake, an attackable public method is the only honesty mechanism left.
- **Any party claiming to reduce somebody else's exposure should carry cover against being wrong about it** — brokers, and one day model vendors.
- **The world model MVP explains rather than calculates, and shows its own emptiness**: fixtures look like fixtures, unmeasured places look unmeasured, the missing claim primitive is a construction site.

## Carry these limits with any summary

Nothing here is insurance. Nothing here is built — two MVPs specified, zero built. No external evidence exists — the survey has not been run, and the one prediction (no product insures an agent placement as a unit) is published to be falsified. The register is fixtures. The placement orderings are unmeasured judgements. The readings were audited: six defects found, fixed, and half-mechanised into gates. Dropping these limits misrepresents the work.