Content Moderation scaffold

A complete program, not a fragment: it compiles as written. Copy it, rename the placeholder identifiers to your domain, and run axon check — the compiler will tell you what your renaming broke. An MCP client can also generate it through the axon.compose tool by naming the content_moderation domain.

// AXON Content Moderation scaffold — multi-tier content review with
// reflex (immediate response) + heal (human-in-loop) chain.
//
// The pattern: user-generated content arrives; the immune system
// (continuous baseline) detects anomalies; reflex fires immediate
// auto-actions (drop/quarantine); heal escalates ambiguous cases to
// human reviewers under bounded patch budget.

// ── Types ─────────────────────────────────────────────────────────

type UGCSubmission {
    submission_id:  Text
    author_id:      Text
    body:           Text
    media_refs:     Text
    submitted_at:   Text
}

type ModerationVerdict {
    submission_id: Text
    decision:      Text
    rationale:     Text
    reviewer_path: Text
}

type ModerateRequest { sub: UGCSubmission }

// ── Resource binding for the immune to watch ──────────────────────

resource ModerationSignals {
    kind:      https
    endpoint:  moderation.signals.base
    lifetime:  persistent
}

fabric ModerationFabric {
    provider:  aws
    region:    "us-east-1"
    zones:     2
    ephemeral: false
}

manifest ModerationManifest {
    resources:   [ModerationSignals]
    fabric:      ModerationFabric
    compliance:  [SOC2]
}

observe SignalStream from ModerationManifest {
    sources: [signal_collector]
    quorum:  1
    timeout: 5s
}

// ── Immune + reflex + heal chain ──────────────────────────────────

immune ContentVigil {
    watch:       [SignalStream]
    sensitivity: 0.88
    baseline:    learned
    window:      500
    scope:       tenant
    tau:         600s
    decay:       exponential
}

reflex QuarantineToxic {
    trigger:  ContentVigil
    on_level: believe
    action:   quarantine
    scope:    tenant
    sla:      500ms
}

shield ModerationShield {
    scan:       [toxicity, prompt_injection, pii_leak]
    on_breach:  quarantine
    severity:   high
    compliance: [SOC2]
}

heal ReviewBorderlineContent {
    source:       ContentVigil
    on_level:     speculate
    mode:         human_in_loop
    scope:        tenant
    review_sla:   2h
    shield:       ModerationShield
    max_patches:  2
}

// ── Identity + grounding ──────────────────────────────────────────

persona ContentModerator {
    domain: ["content-policy", "trust-and-safety", "moderation"]
    tone: precise
    confidence_threshold: 0.85
    cite_sources: true
}

context ModerationReview {
    memory: session
    language: "en"
    depth: deep
    max_tokens: 2048
    temperature: 0.15
}

anchor PolicyCited {
    require: source_citation
    confidence_floor: 0.8
    unknown_response: "Policy citation insufficient — routing to senior moderator."
    on_violation: raise PolicyCitationError
}

// ── Flow ──────────────────────────────────────────────────────────

flow ModerateSubmission(sub: UGCSubmission) -> FlowEnvelope<ModerationVerdict> {
    step Scan {
        given: sub
        ask: "Scan the submission for policy violations; cite the policy section + severity."
        output: FlowEnvelope<ModerationVerdict>
    }
    return Scan.output
}

// ── HTTP boundary ─────────────────────────────────────────────────

axonendpoint ModerationAPI {
    method:      post
    path:        "/v1/content/moderate"
    body:        ModerateRequest
    execute:     ModerateSubmission
    output:      FlowEnvelope<ModerationVerdict>
    shield:      ModerationShield
    backend:     auto
    compliance:  [SOC2]
    retries:     0
    timeout:     5s
    requires:    ["content.moderate"]
}