Audit
The platform records every model call with its full transcript inside your tenant. Every mission is replayable from its event timeline. No third-party trace vendor sits on the path.
The question behind an audit requirement is usually: can you show me exactly what it did, and can I reproduce it?
The record is the timeline
Missions run on an event-sourced world. Every change is an appended event. The history of the mission is not a log written alongside the work. It is the work. A replay of the timeline reconstructs the state at any point.
What that gets you: you can answer "what did it see, and what did it decide" after the fact. The answer comes from the same structure the runtime used to make the decision. There is no separate audit pipeline that can drift from reality or lose events the runtime acted on.
Model calls are recorded in full
The platform records every model call with:
- the prompt messages and the completion, as an immutable transcript. The platform writes the transcript once, at first observation, and never rewrites it.
- prompt and completion token counts, and therefore per-call cost
The transcript lives in the tenant's own world model. The dashboard reads it from there.
No third-party trace vendor sits on this path. Prompts and completions do not leave your tenant to be observable. An earlier build sent traces to an external tracing service. We removed that, and the in-world transcript replaced it. If an older document names that vendor, the document is stale.
This matters more than usual here. Prompts to a security agent routinely contain the thing you protect: hostnames, findings, snippets of your own code. A tracing vendor that holds them is a data-flow that you would have to account for in your own review. There is no such flow.
Traces and metrics
OpenTelemetry traces and Prometheus metrics ship with the platform for the operational picture: latency, errors, and throughput across services. They carry operational telemetry, not prompt content.
The audit log
The daemon writes each audit record to the audit_log table of the
platform Postgres. That table is the durable copy. A Redis stream carries
a live copy for the console.
Some actions write their record before they run. A request to rewind a mission is one example. If the daemon cannot write the record, the action does not run.
Each tenant has a hash chain over its records. Each record holds its position in the chain, the hash of the record before it, and its own hash. An edit of a stored field, or a deleted record, then shows as a break in the chain. The chain has no key. An actor who can write the whole table can compute the chain again, and the chain does not show a cut of the newest records. Do not describe this table as tamper-proof.
Retention
The daemon removes audit records that are older than the retention
period. The default period is 13 months. Set a longer period with
GIBSON_AUDIT_RETENTION_MONTHS. A value under 13 is an error, and the
daemon does not start.
Retention removes only the oldest records of a chain, and only records that the export already wrote to the durable bucket. In the same step it records where the chain now starts, so the chain still verifies.
Export to the durable bucket
The daemon exports each audit record to the durable bucket of the install. On premises, the bucket is your S3-compatible store. Each object holds one range of the chain of one tenant:
audit/<tenant>/<first position>-<last position>.ndjson
Each line is one record with its chain values, so a reader can verify the chain from the object alone. Each object carries an object lock. The export runs each minute. If the daemon restarts during a write, it writes the same range again under the same name. A record is never lost and is never in the bucket under two names.
The daemon reads these variables:
| Variable | Meaning |
|---|---|
GIBSON_AUDIT_EXPORT_BUCKET | The bucket name. |
GIBSON_AUDIT_EXPORT_ENDPOINT | The URL of the store, http or https. |
GIBSON_AUDIT_EXPORT_REGION | The region of the bucket. Empty means us-east-1. |
GIBSON_AUDIT_EXPORT_ACCESS_KEY_ID | The access key of the store. |
GIBSON_AUDIT_EXPORT_SECRET_ACCESS_KEY | The secret key of the store. |
GIBSON_AUDIT_EXPORT_LOCK_MODE | GOVERNANCE or COMPLIANCE. It must match the bucket policy. |
GIBSON_AUDIT_EXPORT_LOCK_DAYS | The lock period in days, at least 1. Use at least the period that the bucket policy requires. |
A bad value stops the daemon at start. With no bucket set, the daemon
starts and logs an error. It then exports nothing, and retention removes
nothing. The metric gibson_audit_export_lag_seconds shows the age of
the oldest record that is not in the bucket.
Attribution
Every component authenticates as itself (see Identity). Every principal has a human owner (see Authorization). So an action in the record resolves to a component. That component resolves to a person.
What to check
- The mission timeline is the source of truth, not a parallel log.
- Model transcripts are immutable once written.
- Prompts and completions stay in the tenant, with no external trace vendor on the path.
- Every recorded action attributes to a component, and every component to an owner.
- Retention keeps 13 months at least, and removes only exported records.
- Each exported object carries an object lock.
Runtime
How an always-on agent runs. A persistent event-sourced loop, goal missions that never finish on their own, mission constraints as the stop button, and no human approval mid-run.
Supply chain
One versioned OCI chart that pins every first-party image by digest at package time, so a version is an exact reproducible deployable. Installs on-prem and air-gapped.