Check Write extension · v1

Nuanced Policy Gates

Judgment-heavy policies evaluated inside check_write before any write. Prefer Pause over silent Allow when confidence is low. Full evidence on every decision.

Where it runs

Inside the existing Check Write step (called gate), after hard deterministic rules (idempotency, unknown key, requireApproval, stale authority), before mismatch / final Allow. Block and Pause never write. SDK guardedWrite unchanged.

Policy types (v1) + example thresholds

Type	Default Pause when
`bias_fairness`	Proxy contribution > 15% or proxy rank in top 3; protected attribute / unapproved alt data
`intent_semantic_drift`	Goal embedding drift > 0.28 from playbook goal
`contextual_appropriateness`	Open Sev-1, active renegotiation, or owner changed in last 7 days
`explanation_quality`	Explanation completeness < 0.65 or key context fields missing
`second_order_effects`	≥ 3 downstream automations, or forecast / commission impact
`value_alignment`	Optimizes a metric while violating documented business intent
`novelty_ood`	OOD / novelty score above threshold (default 0.4)
`reversibility_cost`	Irreversible write or compensation cost above configured threshold

Thresholds live in each policy config and can be changed via API without a code deploy. Both rule thresholds and model scores (recommended / confidence) are supported (signalMode: hybrid).

Decision merge

Any policy recommends Block → Block
Else any policy recommends Pause → Pause
Else → Allow (deterministic Check Write continues)

Risky but unclear cases should Pause. Do not silent-Allow when confidence is low.

Emit signals on Write Intent

Orchestrator (or an upstream scorer) attaches nuancedSignals. No signals → nuanced gates do not change the decision (hard rules still apply).

await protect.guardedWrite({
  intent: {
    system: "Salesforce",
    objectType: "Opportunity",
    objectRef: "006xx",
    field: "StageName",
    newValue: "Closed Won",
    authorityValue: "Eligible",
    authoritySource: "billing_sor",
    nuancedSignals: {
      context: { recentComplaint: true, confidence: 0.6 },
      explanation: { qualityScore: 0.35, missingContext: ["tenure"] },
      reversibility: { irreversible: false, cost: "medium", confidence: 0.7 }
    }
  },
  write: () => sf.sobject("Opportunity").update({ Id, StageName: "Closed Won" }),
});

Evidence trail

Every Pause or Block returns policyEvidence (schema nuanced_policy.v1) with, per fired policy:

Policy name and version (plus id / type)
Key signals that fired
Confidence / signal strength
Human-readable reason
Timestamp
Agent identity (agentIdentity or agent name/id)

Pause / Block also append a Trace event with policyEvidence in metadata (same auditId as Check Write).

Create, version, attach, disable (no code deploy)

Dashboard session actions on /agentops-api.php:

list_nuanced_policies — catalog + defaults, or agent policies (agentId)
set_nuanced_policies — policies[] replace, policy{} upsert, resetToDefaults, or disableAll

// Upsert one versioned policy attached to Opportunity updates
{ "action": "set_nuanced_policies", "agentId": "ag_…",
  "policy": {
    "type": "intent_semantic_drift",
    "version": "2",
    "enabled": true,
    "attach": { "objectTypes": ["Opportunity"], "actionTypes": ["update"], "agentRoles": ["*"] },
    "config": { "goalDriftThreshold": 0.28 }
  }
}

// Roll back to platform defaults
{ "action": "set_nuanced_policies", "agentId": "ag_…", "resetToDefaults": true }

// Disable all nuanced gates (hard rules still run)
{ "action": "set_nuanced_policies", "agentId": "ag_…", "disableAll": true }

Default when unset: all eight types enabled for *. Hard deterministic gates are unchanged either way.

Out of scope (v1)

Auto-remediation or automatic policy rewriting
Full fairness metric suite or model retraining
Dedicated real-time human review UI (use Pause + kill switch + Trace)

Check Write design → Write Intent schema Policy picker

Status: Live in hosted check_write (TC Protect™). Emit nuancedSignals to enforce. Prefer Pause when confidence is low.