Features
Promptless Evidence
Promptless evidence records governance decisions without storing prompt text. It gives operators and auditors proof of policy behavior through safe metadata, identifiers, statuses, counts, approvals, exports, lineage, and replay jobs.
Use promptless evidence when the organization needs accountability for AI and endpoint policy decisions but should not retain raw prompts, completions, lookup values, matched PII, or backend responses in the evidence system.
Evidence APIs
| API | Purpose |
|---|---|
/api/v1/llm/governance/exports | Create and retrieve safe export artifacts. |
/api/v1/llm/governance/exports/from-history | Build an export from retained feature decisions and explicitly authorized endpoint history. |
/api/v1/llm/governance/lineage | Record safe lineage events for policy decisions and artifacts. |
/api/v1/llm/governance/replay-jobs | Track replay jobs for policy and evidence validation. |
/api/v1/llm/governance/feature-records/{feature_kind} | Store feature-specific records such as simulations, approvals, DLP scans, evals, and dashboard views. |
Feature Records
Supported feature kinds include:
| Feature kind | Use it for |
|---|---|
policy_simulation | Preview stored policies against safe request metadata. |
approval_request | Capture human approval or attestation state. |
evidence_pack_template | Define reusable evidence export templates. |
entitlement_sync_job | Track IdP, SCIM, or import sync jobs. |
resource_control | Store resource labels and allowed-group controls. |
model_eval_job | Track model evaluation or drift jobs. |
data_quality_check | Track endpoint or template quality checks. |
dashboard_view | Store curated dashboard views. |
dlp_scan | Store DLP scan metadata and safe counts. |
Evidence Model
| Evidence field | Example |
|---|---|
| Subject identifiers | User UUID, agent UUID, auth group UUID, gateway key UUID. |
| Resource identifiers | Endpoint UUID, template UUID, route alias, model/provider ID. |
| Policy identifiers | Profile UUID, binding UUID, boundary UUID, policy version. |
| Decision metadata | Allow, audit, redact, block, skipped, error, approval pending. |
| Counts | Match counts, term counts, route counts, token counts, row counts. |
| Lineage | Export ID, replay job ID, approval record ID, feature record ID. |
Promptless evidence is intentionally reconstructive rather than verbatim. It should let an auditor answer which policy applied, what action Eden took, and which safe artifacts prove it, without needing sensitive content.
Build Evidence From Retained History
POST /api/v1/llm/governance/exports/from-history builds and persists a hash-addressed compliance or readiness manifest:
{
"export_kind": "soc2_readiness",
"range": "7d",
"feature_names": ["auth_groups", "content_governance"],
"endpoint_uuids": ["11111111-1111-4111-8111-111111111111"],
"sample_limit": 250
}| The caller needs organization `AUDIT | READ`. Every supplied endpoint also |
|---|
transaction history; it does not broaden the request to every endpoint.
The manifest includes source coverage, exact total counts, bounded summaries, bounded event references, truncation flags, and hashes for unsafe service or command dimensions. It does not include prompts, document text, transaction keys or arguments, client IP addresses, user UUIDs, connection IDs, or raw backend responses. The persisted export response is marked Cache-Control: no-store.
Review Workflows
Evidence records support policy simulations before rollout, approval requests for sensitive actions, export packets for reviewers, lineage records tying decisions to artifacts, and replay jobs for validating updated policy against safe historical metadata.
Safety
Evidence records must remain promptless. Store identifiers, policy versions, statuses, counts, and safe metadata. Do not store raw prompts, raw lookup values, raw matched PII terms, or raw provider responses.