Governance

Audit log

Antraft records every decision. The tamper-evident logger hash-chains each record so any edit, reorder, or deletion is cryptographically detectable — a requirement for regulated and compliance use.

Tamper-evident logging

audit.py

from antraft.audit.tamper_evident import TamperEvidentAuditLogger
 
log = TamperEvidentAuditLogger("audit.log", secret="signing-key")
# attach to a runtime:
runtime.auditor = log
 
await runtime.run()
print(log.verify())   # VerifyResult(ok=True, count=N, ...)

Note

Each record stores the SHA-256 of (previous_hash + record). With an optional HMACsecret, an attacker who rewrites the whole file still can't forge a valid chain.

Verifying integrity

verify.py

from antraft.audit.tamper_evident import verify_chain
 
res = verify_chain("audit.log", secret="signing-key")
# res.ok          -> bool
# res.count       -> records verified
# res.broken_index-> first bad record (if any)
# res.detail      -> human-readable reason

What gets recorded

Event	Contains
`action`	The action, the decision (allow/deny/pause), and the reason.
`cognitive`	Agent thoughts / reasoning steps.
`drift_warning`	Anomalous behavior (e.g. tight loops).
`tool_error`	A tool/validation/RBAC failure (non-fatal).
`run_complete`	Terminal summary: actions, spend, status, final answer.

Session recording & PII redaction

Capture full input/thought/action/output traces for replay, with PII scrubbed:

record.py

Antraft.agent(model, tools, task="...")
    .record_session("session.jsonl", redact_pii=True)
    .build()

← Previous

RBAC & tenancy

Human-in-the-loop