Defensive shield wrapping a flow's input/output

Primitives used: persona · flow · shield · anchor · axonendpoint

// A shield is defensive composition: it scans for breaches before
// and after the protected primitive runs. `on_breach:` is a closed
// catalog — {deflect, escalate, halt, quarantine, sanitize_and_retry}.

persona Reviewer {
    domain: ["review"]
    tone: precise
    confidence_threshold: 0.85
    cite_sources: true
}

anchor SafeReply {
    require: source_citation
    confidence_floor: 0.7
    unknown_response: "I cannot answer that safely."
    on_violation: raise AnchorBreachError
}

shield InputOutputShield {
    scan:       [prompt_injection, pii_leak, data_exfil]
    on_breach:  sanitize_and_retry
    // v2.69.0 (axon-T952): `sanitize_and_retry` must name what it masks.
    redact:     [email, phone]
    severity:   high
    compliance: [SOC2]
}

type Query  { text: String }
type Reply  { text: String }
type ReviewRequest { q: Query }

flow Review(q: Query) -> FlowEnvelope<Reply> {
    step Think {
        given: q
        ask: "Review the input and produce a safe reply."
        output: FlowEnvelope<Reply>
    }
    return Think.output
}

axonendpoint ReviewAPI {
    method:     post
    path:       "/v1/review"
    body:       ReviewRequest
    execute:    Review
    output:     FlowEnvelope<Reply>
    shield:     InputOutputShield
    backend:    auto
    compliance: [SOC2]
    retries:    1
    timeout:    10s
}