BankGPTEvidence & storage
Where step logs, transcripts, screenshots, and artifacts live — the engine SQLite DB.
Every run (discovery or replay) leaves a complete, redacted audit trail.
Evidence lives in the engine's own SQLite database
(apps/engine/data/engine.sqlite, override with ENGINE_DB_PATH) —
separate from the frontend auth DB — so nothing touches the filesystem and
everything is served back over the engine HTTP API.
What a run stores
One row per file in the run_files table, keyed by (runId, name):
| File | Contents |
|---|---|
steps.jsonl | One redacted JSON object per step: action, target, the model's/recorded reason, duration, and result. The canonical step store. |
transcript.json | The full model transcript (discovery only), redacted. What makes a discovery run verifiable as genuine. |
result.json | The final structured run result, redacted. |
control.json | The session ownership log (runs with a live session): every pause/cede/resume and human action. |
failure-step-N.png / .yml | Screenshot + accessibility snapshot captured when step N failed. |
handoff-step-N.* | The same capture when a stuck run escalates to a human. |
Approval and dialog steps are logged inline in steps.jsonl (with step
index -1 for approvals), so the decision trail — who approved, what the
model saw, why it acted — is complete in one place.
Redaction
Everything written passes through policy redaction first. redactText /
redactValue scrub credential-shaped and PII-shaped substrings before
anything is persisted, and every JSON response leaves the engine through
the same redaction. Secrets and PII never reach an artifact, a run log, or
a screenshot caption.
The schema
| Table | Contents |
|---|---|
capabilities | One row per stored artifact version: id, version, name, risk, reviewed, createdAt, and the full artifact JSON. |
runs | One row per run: id, kind (discovery/replay), capabilityId, status, goal, targetUrl, timestamps, and the structured result JSON. |
interventions | Approval and stuck requests: kind, status, reason, context, requestedBy / decidedBy, the approval token, and the consuming run. |
run_files | Evidence blobs keyed by (runId, name). |
Rows store validated JSON documents as TEXT and evidence as BLOBs;
querying happens on a few indexed columns only. This keeps the schema
small and the artifact schema (artifact.ts) the single source of truth
for shape.
Reading evidence back
Over the engine API:
curl http://127.0.0.1:4011/runs/<runId>/evidence # parsed step log
curl http://127.0.0.1:4011/runs/<runId>/files # list files
curl http://127.0.0.1:4011/runs/<runId>/files/result.json # one fileor in the console at /admin/runs/:id, which renders the result, the step
evidence, and the failure/handoff captures.