# Authority Geometry for AI-Assisted Delegated Insurance

**Admissible underwriting paths, delegated-authority boundaries and Authority Receipts.**

MAND8 research note — the insurance-specific application layer.

> **Dependency.** This note assumes the **Holonom** engine as a completed
> dependency. Holonom already implements the underlying computation:
> institutional action paths, minimum-distortion reference paths,
> deviation, boundary proximity, evidence-defect taxonomies, reconstruction
> uncertainty and normalized exposure. This note describes only MAND8's
> insurance-specific application layer on top of that engine.

---

## 1. The primitive is the institutional action path

Most AI-governance systems take the model, vendor, prompt or log file as the
primary object. This note takes a different primitive.

The primary object is the **institutional action path**: the sequence through
which information becomes an AI output, the output influences a workflow,
authority rules are applied, a human reviews or overrides the result, and an
institutional action follows.

For delegated insurance, that path may include:

    Risk submission
      → data extraction
      → AI-supported assessment
      → binder and endorsement lookup
      → authority-rule evaluation
      → referral check
      → human review
      → quote or bind decision

The central question is not whether AI participated. It is whether the
AI-influenced portion of the path remained inside the authority, policy and
review boundaries that applied at the time.

## 2. Authority as an admissible region

Every delegated underwriting action operates within constraints: permitted
classes, territories, limits and aggregates, premium thresholds, exclusions,
referral conditions, approval roles, endorsement dates, required review
states.

Together, these constraints define an **admissible region** for the action.

- An action is **admissible** when each required authority, policy, review
  and evidence condition is satisfied.
- A **boundary** is reached when a condition becomes tight or changes regime
  (e.g. a proposed limit approaches a referral threshold).
- A **boundary crossing** occurs when the action proceeds outside the
  authority that applies.

## 3. Authority is not merely a log field

Authority is often stored as a reference, role name or approval status.
This research treats it more fundamentally. Authority constrains which
actor may take which action, under which binder or mandate, within which
limits, during which effective period, subject to which referral or review
requirement.

The same proposed action may be admissible for one actor and inadmissible
for another. The same action may change status when a binder is amended,
an endorsement becomes effective, an aggregate is exhausted, the risk
moves into another territory, a referral condition is triggered, or a
human approval expires.

MAND8's product thesis: these authority conditions should be attached to
the action itself — not reconstructed later from disconnected documents
and systems.

## 4. The Authority Receipt as a coordinate chart

An Authority Receipt is not a full re-recording of every event. It records
the **coordinates** needed to reconstruct the authority question:

- action and workflow
- actor or system influencing it
- binder and endorsement version
- effective authority period
- applicable class, territory and limit
- rules and thresholds evaluated
- required human-review state
- referral, exception or override
- resulting authority classification
- evidence hash and timestamp

The receipt is a **partial coordinate chart** over the institutional
action path. It does not prove that the underwriting judgment was correct;
it is designed to show whether the authority path can be reconstructed.

## 5. Evidence defects

Evidence failure is more than an incomplete form. Defects include:

- **Missing coordinate** — applicable binder version, reviewer, authority
  limit or policy reference was not recorded.
- **Delayed coordinate** — a human approval occurred after the bind action
  but appears as though it happened beforehand.
- **Contradictory coordinate** — the receipt records no referral while
  another system records an exception.
- **Unobserved segment** — the decision contains AI-generated reasoning
  but the model interaction or source evidence was not preserved.
- **Boundary crossing** — the action proceeded after a delegated-authority
  or referral threshold had been crossed.
- **Unverifiable transition** — the workflow moved from recommendation to
  approval but the evidence does not show who or what authorised it.

MAND8 does not automatically label a missing coordinate a breach. The
defensible conclusion is often narrower: *the available evidence is
insufficient to certify whether the action remained inside delegated
authority*.

## 6. Reconstruction uncertainty

A receipt is a partial observation. When important coordinates are
missing, several action paths may be compatible with the same record. One
may have been correctly referred and approved; another may have proceeded
without required approval. If the receipt cannot distinguish them,
**reconstruction uncertainty** is high.

Better-structured Authority Receipts should reduce the set of plausible
action histories, and therefore reduce uncertainty about what occurred.
This is *not* the same as claiming better evidence reduces the underlying
underwriting risk.

MAND8 separates three quantities:

- the risk of the insurance decision
- the authority status of the action
- the quality of the evidence available to reconstruct it

## 7. Research questions for Lloyd's Lab

1. **Representation.** Can a representative delegated-authority structure
   be expressed as typed constraints without erasing contractual
   ambiguity?
2. **Boundary detection.** Can the system distinguish actions that are
   within authority, approaching a referral boundary or outside authority?
3. **Receipt sufficiency.** Which coordinates must an Authority Receipt
   contain for a practitioner to reconstruct the action?
4. **Uncertainty.** Which missing coordinates make the authority status
   undecidable from the receipt alone?
5. **Institutional usefulness.** Would managing agents, MGAs, coverholders,
   compliance teams or auditors use this evidence in an existing
   oversight process?

A negative result improves the research. It would show which parts of
delegated authority resist formalisation, where human interpretation
remains essential, and whether the Authority Receipt is the wrong
primitive.

---

*This note is the Lloyd's-facing slice of a broader working paper on
computational actuarial geometry for institutional AI. The broader paper
covers banking, insurance, legal and other high-assurance institutional
workflows. Neither paper constitutes an approved actuarial, capital or
regulatory framework.*