Delegation Review

Delegation Review Triggers

Closed-loop ABP reassessment — sustained drift signals trigger formal review and authority boundary patching.

Table of Contents

• The Governance Loop
• Delegation Review Flow
• Triggers
• Review Policy
• Client-Side Evaluation
• Trigger States

---

The Governance Loop

The complete at-scale governance loop has three phases:

Phase	What	Status
1. Delegation Artifact	Board/leadership defines what AI can do via ABP v1 — scope, objectives, tools, data permissions, proof requirements	Complete
2. Execution Umbrella	Agents reference ABP v1; every action passes runtime validators (authority_envelope_complete, policy_hash_matches, abp_present_on_export, abp_hash_valid, drift_threshold)	Complete
3. Drift → Filter → Reassessment	Sustained drift opens Delegation Review episode — structured re-evaluation produces ABP patch	Complete

flowchart LR
    subgraph Phase1["1. Delegate"]
        BOARD["Board / Leadership"]
        ABP["ABP v1\n(authority boundaries)"]
        BOARD -->|"defines"| ABP
    end

    subgraph Phase2["2. Execute"]
        AGENT["AI Agents"]
        VALID["Runtime Validators\n(5 checks)"]
        SEAL["Sealed Runs\n(evidence)"]
        ABP -->|"governs"| AGENT
        AGENT -->|"pass validators"| VALID
        VALID -->|"produce"| SEAL
    end

    subgraph Phase3["3. Reassess"]
        DS["Drift Signals\n(8 types)"]
        DRT["DRT Triggers\n(4 conditions)"]
        REVIEW["Delegation Review\n(Reviewer)"]
        PATCH["ABP Patch\n(updated boundaries)"]
        SEAL -->|"generate"| DS
        DS -->|"filter"| DRT
        DRT -->|"threshold met"| REVIEW
        REVIEW -->|"approve"| PATCH
        PATCH -->|"rebuild"| ABP
    end

    style Phase1 fill:#e7f5ff,stroke:#1c7ed6
    style Phase2 fill:#fff3bf,stroke:#f59f00
    style Phase3 fill:#fce4ec,stroke:#e53935

Loading

Delegation Review Flow

flowchart TD
    DRIFT["Drift Events\n(functional, data, assumption,\npolicy, authority, coherence)"]

    DRIFT --> EVAL["DRT Evaluators\n(client-side in Unified)"]

    EVAL --> DRT1{"DRT-001\nRecurring fingerprint\n>= 3 repeats in 14d?"}
    EVAL --> DRT2{"DRT-002\nIrreversible blocked\n>= 5 blocks in 7d?"}
    EVAL --> DRT3{"DRT-003\nCoherence drop\n< 60 for 3 runs?"}
    EVAL --> DRT4{"DRT-004\nAssumption expiry\n>= 4 expired in 30d?"}

    DRT1 -->|"yes + auto_open"| OPEN["Open Review\n(warn)"]
    DRT2 -->|"yes + auto_open"| OPEN_CRIT["Open Review\n(critical)"]
    DRT3 -->|"yes + auto_open"| OPEN_CRIT
    DRT4 -->|"yes, manual"| QUEUE["Queue for Review\n(warn)"]

    DRT1 -->|"no"| CLEAR1["Clear"]
    DRT2 -->|"no"| CLEAR2["Clear"]
    DRT3 -->|"no"| CLEAR3["Clear"]
    DRT4 -->|"no"| CLEAR4["Clear"]

    OPEN & OPEN_CRIT & QUEUE --> REVIEWER["Reviewer\n(threshold: 1)"]

    REVIEWER -->|"approve"| OUTPUT{"Review Output"}
    OUTPUT --> PATCH["abp_patch\n(update boundaries)"]
    OUTPUT --> REPLACE["abp_replace\n(full replacement)"]
    OUTPUT --> REVOKE["abp_revoke\n(revoke ABP)"]

    PATCH --> REBUILD["Rebuild ABP\nbuild_abp.py"]
    REBUILD --> STAMP["Re-stamp EDGE\n(all 8 exports)"]
    STAMP --> GATE["Re-gate\ngate_abp.py"]

    style OPEN fill:#fff3bf,stroke:#f59f00
    style OPEN_CRIT fill:#fce4ec,stroke:#e53935
    style QUEUE fill:#fff3bf,stroke:#f59f00
    style PATCH fill:#d3f9d8,stroke:#37b24d
    style REPLACE fill:#e3f2fd,stroke:#1e88e5
    style REVOKE fill:#ff6b6b,color:#fff

Loading

Triggers

Four concrete triggers are defined in the ABP's delegation_review.triggers array:

ID	Name	Condition	Severity	Auto-Open
DRT-001	recurring_drift_fingerprint	Same drift fingerprint appears >= 3 times within 14 days	warn	true
DRT-002	irreversible_action_blocked	Gate blocks >= 5 irreversible actions within 7 days	critical	true
DRT-003	linkage_score_sustained_drop	Overall coherence index < 60 for >= 3 consecutive runs	critical	true
DRT-004	assumption_expiry_breach	>= 4 expired claims/assumptions within 30 days	warn	false

DRT-001: Recurring Drift Fingerprint

What it detects: The same root cause keeps triggering drift signals — a structural issue, not a transient one.

Condition:

{
    "fingerprint_repeats": 3,
    "window_days": 14
}

Data sources: Drift signals from hiring/bid localStorage data, fingerprint deduplication.

DRT-002: Irreversible Action Blocked

What it detects: The ABP is blocking too many actions — the boundary may be too restrictive for current operations.

Condition:

{
    "blocked_count_gt": 5,
    "window_days": 7
}

Data sources: Gate failure counts from buildHostRollup().

DRT-003: Linkage Score Sustained Drop

What it detects: Organizational coherence is degrading consistently — not a blip, but a trend.

Condition:

{
    "overall_coherence_lt": 60,
    "consecutive_runs": 3
}

Data sources: Composite score history tracked in ds_delegation_review_state_v1 localStorage key.

DRT-004: Assumption Expiry Breach

What it detects: Too many assumptions have expired without being reviewed — stale logic is accumulating.

Condition:

{
    "expired_claims_gt": 4,
    "window_days": 30
}

Data sources: Expired assumptions from bid/boe localStorage data.

Review Policy

When a trigger fires, the review policy governs what happens next:

Field	Value	Description
approver_role	"Reviewer"	Who must approve the review
threshold	1	Number of approvals needed
timeout_ms	604800000	7-day timeout window
output	"abp_patch"	Default output type

Output Types

Output	Description
abp_patch	Update specific ABP sections (objectives, tools, permissions, etc.) — preserves continuity
abp_replace	Full ABP replacement — new ABP ID, new hash, fresh start
abp_revoke	Revoke current ABP — requires new ABP before operations resume

Client-Side Evaluation

In EDGE Unified's Delegation tab, triggers are evaluated client-side using four evaluator functions:

Function	Trigger	Data Source
evaluateDRT001(trigger)	DRT-001	Drift fingerprint repeats across hiring/bid localStorage
evaluateDRT002(trigger)	DRT-002	Gate failure count from buildHostRollup()
evaluateDRT003(trigger)	DRT-003	Composite score history in ds_delegation_review_state_v1
evaluateDRT004(trigger)	DRT-004	Expired assumptions from bid/boe localStorage

The evaluators are registered in the DRT_EVALUATORS map:

const DRT_EVALUATORS = {
    'DRT-001': evaluateDRT001,
    'DRT-002': evaluateDRT002,
    'DRT-003': evaluateDRT003,
    'DRT-004': evaluateDRT004,
};

renderDelegationPanel() iterates over all triggers from the ABP, runs the corresponding evaluator, and renders cards with progress bars showing threshold proximity.

Trigger States

Each trigger card in the Delegation tab displays one of three states:

State	Visual	Meaning
clear	Green pill	Well below threshold — no action needed
armed	Yellow pill	Approaching threshold — warning level
tripped	Red pill	Threshold met/exceeded — review needed

The state is determined by comparing the evaluator's current count against the trigger's condition threshold.

Schema Validation

The ABP JSON schema (enterprise/schemas/reconstruct/abp_v1.json) enforces:

• Trigger IDs match pattern ^DRT-[0-9]{3}$
• Severity must be "warn" or "critical"
• auto_open is boolean
• Review policy requires approver_role, threshold (>= 1), timeout_ms (integer or null), output (one of abp_patch, abp_replace, abp_revoke)
• The entire delegation_review section is optional — ABPs without it still validate