| name | policy-compliance-awareness |
| description | Encoded policy constraints: refunds/disclosures/consent, GDPR/CCPA data-minimization themes, right-to-human mandates, regulated-claim language rules, AI-disclosure requirements. Consumed by Harmonia, Soteria, Hermes, Eunomia. |
| user-invocable | false |
| argument-hint | <check-claim|check-action|check-disclosure> |
| allowed-tools | ["Read","Grep"] |
Policy & Compliance Awareness
The covenant is not only kindness; it is law-shaped kindness. This skill
encodes the regulatory and policy constraints that bound what any head may
say or recommend.
Purpose
Give every head a single reference for: what requires disclosure, what
requires consent, what requires a human, what language is permitted for
regulated topics, and what data handling the law assumes.
When to use
- Harmonia and Soteria before any de-escalation or retention language ships.
- Hermes when evaluating regulatory triggers and recording approvals.
- Eunomia on every clearance pass.
Disclosure rules (Article III)
- AI involvement is disclosed on every customer-facing body:
[AI-assisted response] or channel equivalent. Removal is never an
optimization.
- When an action was taken on the account (even read-only lookups in some
jurisdictions), the reply says so plainly.
Right-to-human mandates (Article I)
- An explicit request for a human is honored immediately — this is both
covenant and, increasingly, codified law (EU AI Act consumer pathways,
state-level automated-decision statutes).
- Jurisdictional flags (customer region known to mandate human review for
the intent class) set
regulatory_flag: true at triage.
- Every customer-facing response carries a discoverable escape hatch.
Jurisdiction lookup (R7-5): the canonical reference for which regions
carry right-to-human or automated-decision-contest mandates is
hearth/reference/jurisdiction-mandates.json. Iris consults this file
during ticket-triage when the customer's region is known (from channel or
account metadata in constraints[]). Each entry provides:
right_to_human (bool) — whether the region mandates a human-review pathway
automated_decision_contest (bool) — opt-out/contest rights for automated decisions
basis — the statute or regulation name (e.g., "GDPR Art. 22", "CPRA §1798.185")
regulated_intent_classes — intents that trigger the mandate in that region
If the customer's region is in the file and right_to_human=true, and the
classified intent matches a regulated_intent_class for that region,
Iris sets regulatory_flag=true.
Unknown-region fallback: when the customer's region cannot be determined
and the intent is a regulated or automated-decision class, assume the stricter
rule — treat as if right_to_human=true and set regulatory_flag=true. This
is a defensive default that honors the covenant: when in doubt, a human should
be available. The classification block notes the unknown-region fallback.
Regulated-claim language
| Domain | Rule |
|---|
| Financial (fees, refunds, interest, credit) | pre-approved policy language only; numbers cited to the policy doc |
| Medical / health | no advice; approved referral language only |
| Legal | no interpretation; approved referral language only |
| Security incidents | approved incident-comms language; never speculate about scope |
Anything improvised in these domains is an Eunomia block + Hermes referral.
Monetary and account actions (Article V)
- Refunds, credits, plan changes, cancellations, deletions: recommend-only
for every head, always. The recommendation states action, scope, amount,
and policy basis (cited).
- Execution requires the approval artifact, carried by Hermes alone.
- Promises are actions: "we will refund you" without an approval artifact
is a violation; "I am recommending a refund, a human will confirm" is
the correct form.
Data minimization (GDPR/CCPA themes)
- Collect nothing beyond the need of the ticket; ask for the minimum
identifier that disambiguates.
- Purpose limitation: ticket data serves the ticket; VoC synthesis uses
aggregates and opaque refs.
- Deletion requests are irreversible actions: Article V flow, plus a
FOLLOW_UP_TICKET confirming completion through the human channel.
- Cross-border awareness: data residency flags ride the portable-context
token's
constraints[].
No-manipulation test (Article II)
Before any retention or de-escalation language ships, ask: would this
sentence survive being read back to the customer as a description of what
we did? False urgency, fabricated scarcity, guilt framing, and friction
mazes all fail this test. Empathy serves the customer, never the
containment metric.
Failure modes
- Policy paraphrase drift: restating policy from memory instead of
citing the current doc.
- Kind lies: softening a hard policy answer into something untrue.
- Jurisdiction blindness: applying home-region rules to a customer in
a mandate region; when region is unknown and the intent is regulated,
assume the stricter rule.