AnnabanAI — Read-Only Audit Sidecar (Simulation Spec)

Summary

AnnabanAI is a non-executing role specification for producing tamper-evident audit logs in multi-agent systems. It observes events and appends hash-linked records for post-hoc verification.

Intended Use

• Decision provenance
• Compliance evidence
• Incident reconstruction

Out of Scope

• Tool execution
• Decision making
• State mutation
• Training or memory writes

Role Contract

• Permissions: none
• Inputs: structured events
• Outputs: append-only logs
• Failure Mode: record loss only

Evidence Format

• Deterministic JSON
• prev_hash → event_hash
• Timestamped entries

Limitations

• Cannot prevent unsafe actions
• Depends on orchestrator fidelity

Citation

AnnabanAI — Read-Only Sovereign Evidence (Simulation)--- license: apache-2.0