AI Accountability Ledger

Purpose

Provides an immutable, auditable record of all AI model actions, decisions, and interactions to ensure transparency, enable post-hoc analysis, and support accountability in multi-model systems.

Activation

/skill ai-accountability-ledger

Ledger Architecture

1. Record Types

Type	Description	Retention
Action	Model performed operation	90 days
Decision	Model made choice	1 year
Interaction	Model-to-model exchange	90 days
Governance	Consensus/arbitration event	Permanent
Violation	Policy breach detected	Permanent
Correction	Error acknowledged/fixed	Permanent

2. Ledger Entry Schema

json

{
  "entry_id": "LED-{ulid}",
  "timestamp": "{iso_timestamp}",
  "entry_type": "{type}",
  "actor": {
    "model_id": "{identifier}",
    "provider": "{provider}",
    "session_id": "{session}"
  },
  "action": {
    "type": "{action_type}",
    "description": "{what_happened}",
    "inputs": ["{input_summary}"],
    "outputs": ["{output_summary}"],
    "rationale": "{why_this_action}"
  },
  "context": {
    "task_id": "{task}",
    "parent_entry": "{previous_entry_id}",
    "related_models": ["{other_models}"]
  },
  "accountability": {
    "responsibility_chain": ["{model_ids}"],
    "human_oversight": "{oversight_level}",
    "reversibility": "reversible|irreversible|partial"
  },
  "integrity": {
    "hash": "{sha256_of_entry}",
    "previous_hash": "{chain_link}",
    "signatures": ["{model_signatures}"]
  }
}

3. Chain Structure

┌─────────────────────────────────────────────────────────────┐
│ Genesis Block                                               │
│ hash: 0x000...                                              │
└─────────────────────┬───────────────────────────────────────┘
                      │
┌─────────────────────▼───────────────────────────────────────┐
│ Entry LED-001                                               │
│ prev_hash: 0x000... │ hash: 0xabc...                        │
│ action: "Task initiated"                                    │
└─────────────────────┬───────────────────────────────────────┘
                      │
┌─────────────────────▼───────────────────────────────────────┐
│ Entry LED-002                                               │
│ prev_hash: 0xabc... │ hash: 0xdef...                        │
│ action: "Claude analyzed data"                              │
└─────────────────────┬───────────────────────────────────────┘
                      │
                      ▼
                    (...)

Accountability Features

1. Responsibility Attribution

Every action traces back to:

Primary Actor: Model that performed action
Delegator: Model that requested action (if any)
Approver: Model/human that authorized action
Oversight: Human with review responsibility

2. Audit Queries

sql

-- Find all actions by a specific model
SELECT * FROM ledger WHERE actor.model_id = 'claude-3'

-- Trace decision chain for specific outcome
SELECT * FROM ledger
WHERE task_id = 'TASK-123'
ORDER BY timestamp

-- Find all violations in time range
SELECT * FROM ledger
WHERE entry_type = 'violation'
AND timestamp BETWEEN '2026-01-01' AND '2026-02-01'

-- Get responsibility chain for action
SELECT accountability.responsibility_chain
FROM ledger WHERE entry_id = 'LED-456'

3. Violation Handling

xml

<violation-record>
  <violation-id>VIO-{timestamp}</violation-id>
  <severity>low|medium|high|critical</severity>
  <violator>{model_id}</violator>
  <rule-violated>{rule_reference}</rule-violated>
  <evidence>
    <entry-ids>["{related_entries}"]</entry-ids>
    <description>{what_happened}</description>
  </evidence>
  <response>
    <immediate-action>{containment}</immediate-action>
    <investigation-status>{status}</investigation-status>
    <corrective-action>{remediation}</corrective-action>
  </response>
</violation-record>

Governance Integration

Codex Law Compliance

Every entry checked against:

CONSENT: Was proper authorization obtained?
INVITATION: Was action within invited scope?
INTEGRITY: Does action maintain system integrity?
GROWTH: Does action support beneficial growth?

Human Oversight Levels

Level	Description	Logging Detail
Full	Human reviews all actions	Maximum detail
Selective	Human reviews flagged actions	High detail
Audit	Human can review on demand	Standard detail
Minimal	Routine operations only	Summary only

Integration Points

mnemosyne-ledger: Synchronizes with memory system
codex-law-enforcement: Compliance checking
shatter-protocol: Human override logging
inter-model-arbitration: Dispute evidence source
cross-model-trust-verification: Trust event logging

Retention & Privacy

Data Minimization

Only log necessary information
Summarize sensitive content
Hash personally identifiable information

Retention Schedule

Routine actions: 90 days
Decisions: 1 year
Governance events: Permanent
Violations: Permanent
Corrections: Permanent

Right to Explanation

Any logged action can generate:

Plain-language explanation
Responsibility attribution
Decision factors
Alternative paths considered

Example Ledger Entries

json

[
  {
    "entry_id": "LED-01HQ3X...",
    "timestamp": "2026-02-04T10:30:00Z",
    "entry_type": "decision",
    "actor": {"model_id": "claude-opus", "provider": "anthropic"},
    "action": {
      "type": "task_delegation",
      "description": "Delegated code review to Gemini",
      "rationale": "Gemini has higher code analysis score for Python"
    },
    "accountability": {
      "responsibility_chain": ["claude-opus", "gemini-pro"],
      "human_oversight": "audit",
      "reversibility": "reversible"
    }
  }
]

Metrics

entries_per_day: Logging volume
chain_integrity: % blocks with valid hashes
violation_rate: Violations per 1000 actions
audit_response_time: Time to generate audit report
attribution_completeness: % entries with full chain

Search AI Tools

ai-accountability-ledger

Install this agent skill to your Project

SKILL.md