---
id: t7-risk-expert-review-plan
title: Risk Expert Review Plan
module: GROW-S6
module_slug: grow-s6-govtech-compliance
cluster: Trust
type: playbook
version: v0.1.0
status: Gate-reviewed
tier: membership
contract_role: ""
canonical_url: "https://grow.goodcombinator.ai/library/registry/t7-risk-expert-review-plan"
download_url: "https://grow.goodcombinator.ai/library/registry/t7-risk-expert-review-plan.md"
license: CC-BY-4.0 (proposed — owner confirmation required)
source: GROW by Good Combinator
retrieved_at: 2026-05-29
---

# Risk & Expert Review Plan

> **[ATTORNEY REVIEW] — This entire playbook is legally operative.** The playbook defines where a compliance agent must stop and where a qualified human must take over. These boundaries have legal consequences in a public-sector context. Every section of this document must be reviewed by qualified counsel before it is treated as binding for any Florida special-district, county-agency, or public-body deployment. Nothing in this document constitutes legal advice.

The Risk and Expert Review Plan is the operational playbook for identifying compliance risk in agent-assisted govtech workflows, marking the points at which a determination must leave the agent and reach a qualified human, and enforcing the "not legal advice" boundary by mechanism rather than by disclaimer alone. The plan is built on `s1-hitl-review-policy` (which defines the HITL gate mechanics and the C2 override event payload) and feeds `t7-compliance-workflow` (which implements the gate triggers). A compliance agent operating without this plan in place is operating without a safety net: it may produce outputs that a recipient reasonably interprets as authoritative legal determinations, creating liability for the deploying organization and harm to the applicant. `[ATTORNEY REVIEW]`

## 1. Risk Taxonomy for Govtech Compliance Agents

Govtech compliance risk is distinct from general agent reliability risk because the harm pathway runs through public-sector authority. The taxonomy below classifies risk by severity (reusing the GROW glossary enum) and maps each class to its expert-review trigger.

| Risk class | severity | Description | Expert-review trigger |
|---|---|---|---|
| Statutory misapplication | `critical` | Agent applies the wrong statute, cites a repealed provision, or applies a threshold from a superseded rule version. | Any determination on a rule with `verify_flag: true` and `last_confirmed > 12 months`; any rule with `blocking: true` in `t7-rule-change-tracker`. |
| Legal interpretation required | `critical` | Determination requires interpreting ambiguous statutory language, resolving a conflict between rules, or applying a legal standard (e.g., "reasonable time," "de minimis") that is not quantitatively defined. | Any determination where the rule's `agent_trigger` requires a judgment call rather than a binary threshold check. |
| ERP/permitting authority boundary | `critical` | Agent output could be construed as an environmental resource permit determination, a denial, or a waiver of FDEP requirements. `[ATTORNEY REVIEW]` | Any determination touching `fdep-stormwater-general` or any rule row with `attorney_review: true`. |
| Public-records obligation | `high` | Agent output that becomes a public record under FS Chapter 119 [VERIFY] and contains an error that propagates into the official record. | Any agent output that will be filed, posted, or transmitted on behalf of a public body. |
| Multi-rule conflict | `high` | Two activated rules produce contradictory requirements; the agent cannot resolve the conflict with rule-based logic alone. | Any application where Phase 3 produces two rules with incompatible `approval_flow` requirements. |
| Stale-rule reliance | `high` | Agent cites a rule row where `last_confirmed` is beyond the freshness window, producing a determination that may be wrong under updated law. | Any rule with `last_confirmed > 12 months` (12-month flag) or `> 18 months` (hard block). |
| PII handling in determination | `high` | Applicant personal information appears in the determination output routed to a party not authorized for that data class. | Any output with `pii_flags` populated that is routed outside the authorized sink list. |
| Threshold ambiguity | `medium` | Measured value is close to a trigger threshold; minor measurement error would flip the determination. | Applied when the input value is within 10% (illustrative) of the threshold and the source `confidence_band` is `medium` or `low`. |
| Form or deadline discrepancy | `medium` | Agent cites a form ID or deadline that has been updated but the rule row has not yet been refreshed. | Applied when the `approval_flow` form ID does not resolve in the current forms library. |

## 2. Expert-Review Boundary Definition

The `expert-review boundary` (glossary term) is the line at which the compliance agent hands off to a qualified human. On this side of the boundary, the agent classifies. Beyond it, a human with appropriate credentials must make the determination. The boundary is not a suggestion; crossing it autonomously is a `critical`-severity failure mode (`unsafe-action-attempted`).

**The boundary is crossed when any of the following is true:**

1. The determination requires applying a legal standard that is not quantitatively defined in the rule row. `[ATTORNEY REVIEW]`
2. The rule row carries `attorney_review: true`.
3. The rule row has `verify_flag: true` and `last_confirmed > 12 months`.
4. The `determination` for any rule is `needs-review` after Phase 3.
5. The activated rule set contains a rule in the `critical` or `high` risk class above AND the source `confidence_band` for the triggering evidence is `medium` or `low`.
6. The application is a first-of-type fact pattern with no prior determination in the evidence store.
7. The applicant is a public body, government agency, or elected official — determinations involving these parties carry heightened accountability. `[ATTORNEY REVIEW]`

**What "crossing the boundary" means operationally:**
- `expert_review_required: true` is set on the C9 record.
- An S1 HITL event is raised with `decision_origin: escalation` and a rationale naming the specific boundary condition triggered.
- Phase 4 and Phase 5 of `t7-compliance-workflow` are blocked.
- The agent produces an internal expert-queue summary (not an applicant-facing output) describing the boundary condition and the information the expert needs to resolve it.

## 3. Expert Queue Protocol

The expert queue receives blocked determinations and manages the review SLA. The queue is implemented as a HITL gate per `s1-hitl-review-policy`; the specific mechanics use the C2 override event shape.

**Expert roles and their authority:**

| Role | Authority scope | SLA (business hours) |
|---|---|---|
| `licensed-environmental-engineer` | ERP threshold determinations, stormwater calculation review, FAC Chapter 62-330 threshold interpretation `[ATTORNEY REVIEW]` | 8 hours for `critical`; 24 hours for `high` |
| `district-counsel` | Statutory interpretation, public-records obligations, applicant-rights questions, any `attorney_review: true` rule `[ATTORNEY REVIEW]` | 4 hours for `critical`; 16 hours for `high` |
| `county-planning-staff` | Walton County LDC threshold and form questions | 8 hours for `critical`; 24 hours for `high` |
| `district-clerk` | Administrative completeness, form deficiency | 2 hours for all classes |
| `board-commissioner` | Determinations requiring board discretion or formal board vote | Per board meeting schedule; emergency powers apply for `critical` |

**Expert-queue packet contents.** The packet is the evidence bundle surfaced to the expert via the HITL gate. Per `s1-hitl-review-policy` §6, the reviewer must see exactly: the staged determination, the top three evidence pointers, the `failure_id` refs, the relevant rule row, and a single approve/modify/refuse/escalate control. For compliance contexts, add:

- The specific boundary condition triggered (from §2 above).
- The text of the activated rule (not a summary — the actual citation text). `[VERIFY]`
- The applicant's submitted evidence that triggered the rule.
- The prior determination history for this applicant and parcel, if any.
- The "not legal advice" boundary notice.

**Expert resolution actions:**

| Action | Effect on C9 record | Effect on workflow |
|---|---|---|
| Confirm agent determination | `determination` unchanged; `decision_origin` updated to `human-override`; HITL event closed | Blocking lifted; workflow advances to Phase 4 |
| Modify determination | `determination` updated; `decision_origin: human-override`; rationale required | Blocking lifted; modified determination logged |
| Refuse and instruct | Determination held; expert provides corrective instruction for re-evaluation | Re-evaluation run; new C9 record emitted |
| Escalate to higher tier | Routed to next role (e.g., district-counsel → outside-counsel) | New HITL event; SLA clock resets |
| Defer pending rule update | Determination held; `t7-rule-change-tracker` entry opened | Blocking maintained until tracker entry resolved |

## 4. "Not Legal Advice" Enforcement Mechanism

The "not legal advice" boundary is meaningless if it is enforced only by a disclaimer in an output document. This playbook enforces it structurally.

**Mechanism 1 — Block-before-output.** `expert_review_required: true` blocks the applicant-facing output at the workflow level, not at the presentation layer. The disclaimer cannot be stripped because the output does not exist until the HITL gate clears.

**Mechanism 2 — Decision origin tagging.** Any applicant-facing output carries a `decision_origin` trace. If any element of the trace is `compliance-determination` without expert validation for a rule above the boundary, the output is gate-failing.

**Mechanism 3 — Mandatory disclaimer language.** `[ATTORNEY REVIEW]` The following notice must appear on all applicant-facing determination outputs, in accessible language, without modification: "This screening summary is produced by an automated compliance tool operated by [District Name]. It is not a final agency determination, is not legal advice, and does not constitute a permit, approval, denial, or legal opinion. You have the right to seek independent legal counsel and to request review of any determination by district staff. For questions about your rights under Florida law, contact [District Contact Information]." Counsel must review and approve the specific language for each jurisdiction before it is used.

**Mechanism 4 — Audit trail integrity.** The `decision_trace` in `s3-provenance-metadata-schema` records the `decision_origin` for every element of every determination. A determination that reached an applicant without a `human-override` or `escalation` event for every `expert_review_required: true` rule is a provenance integrity failure and triggers an immediate correction workflow per `t7-accountability-layer`.

## 5. Risk Review Cadence

Govtech risk does not diminish over time; it accumulates as rules change and as the agent handles more fact patterns. The following cadence maintains the safety level.

**Weekly (operational).** Review the expert-queue backlog: any HITL events older than 2x the role SLA are escalated to the next tier automatically.

**Monthly (trend).** Review the distribution of `determination: needs-review` outcomes. If the `needs-review` rate for any rule exceeds 40% of activations (illustrative threshold), the rule's `agent_trigger` may be too broad, or the rule may need a threshold recalibration. This is a signal for a hardening task per `s1-hitl-review-policy` §5.

**Quarterly (structural).** Full review of the expert-review boundary conditions in §2 against current rule-map state. Any new `attorney_review: true` row in `t7-jurisdiction-rule-map` must be reflected in the boundary conditions before the next deployment cycle.

**Annual (policy).** Full `[ATTORNEY REVIEW]` of this playbook against current Florida statutes, FDEP rules, and applicable federal authority. All SLA values and role definitions reviewed with district counsel.

**Triggered.** Any instance where a determination crossed the boundary without raising an HITL event is a `critical`-severity incident. Immediate incident review, `s1-monitoring-rollout-postmortem` postmortem triggered, corrective action required before the next determination is issued.

## 6. Worked Example — ERP Boundary Crossing, South Walton County

**Scenario.** Application `app-2026-0515` triggers `fdep-stormwater-general`. The agent evaluates: impervious increase is 0.4 acres. The rule row has `verify_flag: true`, `last_confirmed: 2024-11-01` (18 months prior — auto-block threshold), `attorney_review: true`. The rule is also in the active tracking item `rc-2026-0001` in `t7-rule-change-tracker` with `blocking: true`.

**Boundary conditions triggered:** (1) `verify_flag: true` + `last_confirmed > 18 months`, (2) `attorney_review: true`, (3) rule in `blocking: true` tracker item.

**Agent actions:**
1. Sets `expert_review_required: true`, `determination: needs-review`, `decision_origin: escalation` on the C9 record for `fdep-stormwater-general`.
2. Raises S1 HITL event with rationale: "Rule fdep-stormwater-general has not been confirmed in 18 months, carries attorney_review flag, and is blocked in t7-rule-change-tracker pending threshold verification. ERP applicability determination requires licensed engineer + district counsel review."
3. Blocks Phase 4 and Phase 5. No applicant-facing output produced.
4. Routes expert-queue packet to `licensed-environmental-engineer` (8-hour SLA) and `district-counsel` (4-hour SLA, given `critical` risk class).
5. Engineer reviews FDEP FAC Chapter 62-330 and confirms the threshold has NOT changed; `[VERIFY]` cleared by counsel. Tracker item `rc-2026-0001` is updated; `blocking` lifted.
6. Determination re-run: ERP pre-application meeting is required. `determination: needs-review` updated to `non-compliant` (application cannot proceed without ERP pre-app). `decision_origin: human-override`. `expert_review_required` remains recorded as `true` (the boundary was crossed; that fact is permanent in the record).
7. Phase 4 aggregates; Phase 5 produces clerk-reviewed applicant notice: "Your application requires an FDEP ERP pre-application meeting before the district can process the permit. See attachment for FDEP contact information." Disclaimer language appended. `[ATTORNEY REVIEW]`

**Outcome.** The applicant was not told anything by the agent. The applicant received a clerk-transmitted notice grounded in expert confirmation. The provenance record shows the full trace including the HITL escalation, the expert review, and the override event.