?
“Why did the agent do that?”
Logs often miss context, approver, and prior codebase state.
Provenance
Every line traces to its decision
Approver, brief, model, and diff on one audit trail.
The category
Git blame stops at the commit. Provenance answers who approved, which brief, which context, and which model produced the change, as a signed record.
Every agent run snapshots brief, context blocks, codebase state, and the approving human.
Model ID, version, and parameters recorded.
Each diff line mapped to the agent step that produced it.
Append-only, signed records you can query, replay, and export.
The gap
The audit question that stalls AI programs
Boards, auditors, and clients ask who is accountable for what the AI did. Most tooling cannot answer.
EU AI Act
Compliance regimes demand traceability
EU AI Act, NIST AI RMF, ISO 42001.
Liability
When the AI is wrong, who pays?
Provenance gives root cause instead of blaming whoever was on call.
Stale
Git blame stops at the commit
It misses prompt, brief, context, model version, and approval chain.
What provenance captures per decision
A typed, queryable, append-only record of how the work came to be.
Who approved
Named human approval with role, timestamp, and SSO identity.
Stays in the workspace with plan and checks
What brief drove it
Versioned brief or ticket snapshot you can replay after later edits.
Stays in the workspace with plan and checks
Which context was used
Context blocks, retrieved files, and meeting snippets as a graph.
Stays in the workspace with plan and checks
Model + version + parameters
Provider, model ID, version, temperature, and overrides.
Stays in the workspace with plan and checks
The diff, attributed
Each line mapped to the agent step that produced it.
Stays in the workspace with plan and checks
Approvals and overrides
Overrides captured with reviewer rationale.
Stays in the workspace with plan and checks
Tamper-evident
Append-only storage with cryptographic signatures.
Stays in the workspace with plan and checks
Queryable
Query by path, quarter, or context block. Real query layer.
Stays in the workspace with plan and checks
Provenance vs standard audit logs
Standard audit logsArvad provenance
Arvad sits where the thread keeps the plan, staging, spend, and checks.
| Capability | Standard audit logs | Arvad provenance |
|---|---|---|
| Captured granularity | API calls | Agent decisions, approvals, context |
| Human approval | Maybe | Required, named, SSO-bound |
| Brief at time of run | Not captured | Versioned snapshot |
| Context graph | None | Every block, every file |
| Model version | Sometimes | Always, with parameters |
| Tamper-evident | Depends on infra | Signed, append-only |
| Queryable | grep | Real query layer |
| Mapped to controls | Manual | Mapped toward SOC 2, ISO 42001, EU AI Act |
Provenance is the default
Every agent run is wrapped, signed, and stored. Replay any decision or export the bundle.
Every agent run is wrapped, signed, and stored. Replay any decision or export the bundle.
01Every agent run is wrapped
02Records are signed and stored
03Replay any decision
04Hand it to whoever asks
Why provenance becomes a buying trigger
EU AI Act
Article 12, record-keeping
Supports traceable, automatically generated logs for AI-mediated changes.
SOC 2
CC7, CC8 controls mapped
Change evidence auditors can review, not screenshots.
ISO 42001
AI management system aligned
Supports transparency and traceability controls until certified.
Straight answers
Limits stated plainly. No demo theater.
One question
When the build finishes, can you show a teammate the plan, preview, tests, and deploy status without digging through chat?
If you cannot, you still have a chat log. Arvad keeps that work in one workspace.
Keep exploring
When the question comes, have the trail
Brief, context, approval, diff, and model on one trail.