Evidence
What an AI evidence record actually contains
Vendors talk about "full audit trails" and then never show you a field list. Here is ours, with each field marked as shipped, pilot work or design. Take it to whoever reviews you and check it against their list before you talk to us.
Every capability on this page carries its real status: shipped code, work we do inside a pilot, or a design we have not built yet.
The record, field by field
Fields marked live exist today in the router we run on our own AI spend. Fields marked design do not exist — they are the direction, and calling them shipped would be the exact lie this page is against.
| Field | The question it answers | Status |
|---|---|---|
timestamp | When the call happened.Stored per record, UTC. | Live today |
actor | Which user or service account made the call.Nullable for unauthenticated internal jobs — the record says so rather than guessing. | Live today |
source | Which surface or subsystem originated the call.Free-form, indexed — the closest thing to "which feature" that exists today. | Live today |
operation | What the call was for (classification, summary, translation…).Indexed, used to group cost by kind of work. | Live today |
provider | Which vendor answered.Resolved per call across the configured adapters. | Live today |
model | Which model actually ran — not which one was requested.This distinction matters the moment a fallback fires. | Live today |
input_tokens / output_tokens | How much went in and came out.Reported by the provider; absent when the provider does not return counts. | Live today |
latency_ms | How long the call took.Useful for the "was a human waiting on this" question. | Live today |
cost_usd | Estimated cost of this single call.Estimated from token counts and a price table — an estimate, not the invoice line. | Live today |
request_meta | Structured extras attached at call time.JSON. Where per-pilot fields live until they earn a column. | Live today |
team / project | Which team and initiative the call belongs to.No such columns today. Per-team attribution is design, not a setting. | Design intent |
policy_result | Whether a rule allowed, redacted or blocked the call.There is no policy engine in the product. Routing picks a provider and model; it does not adjudicate. | Design intent |
event_hash / chain | Whether the record can still be trusted months later.Not implemented. Any page telling you our records are tamper-evident today would be wrong. | Design intent |
scope / retention metadata | Which system and period a pack covers, and how long records are kept.Agreed per pilot today, because the right answer depends on your auditor and your residency rules. | Pilot scope |
One record, shaped
An illustrative record with example values — the structure is real, the numbers are not from your account or anyone else's.
{
"id": "7c1f…", // record id
"timestamp": "2026-08-05T09:14:22Z",
"actor": "u_314 (analytics service)",
"source": "kb-translation",
"operation": "translate_article",
"provider": "deepseek",
"model": "deepseek-v4-flash",
"input_tokens": 2841,
"output_tokens": 1190,
"latency_ms": 3120,
"cost_usd": "0.004310",
"request_meta": { "article_id": 812, "target_locale": "ru" }
// not present today — see the table above:
// "team", "project", "policy_result", "event_hash"
}Illustrative example. Not exported data, not a customer record.
What a pack would add on top
A pack is a scoped bundle rather than a dump. None of this is a button in the product today; it is what a pilot assembles with you.
Scope statement
Pilot scopeWhich system, which period, which decisions — written down, so a reader knows what the pack does not cover.
Field dictionary
Pilot scopeThe same table you are reading, frozen at export time, so a reviewer in a year knows what each column meant.
Records for the scope
Pilot scopeThe rows themselves, complete within the declared scope rather than sampled.
Integrity proof
Design intentA chain that makes a later edit detectable. Designed, not built — this is the single biggest gap between this page and a finished product.
What exists today
- Per-call records with timestamp, actor, source, operation, provider, model, tokens, latency, estimated cost and a JSON extras field — in our own router, covered by tests.
- A free External AI Exposure Check that finds AI use nobody registered, with a permanent link per report.
- This page: a field list you can check without talking to sales.
What we do not claim
- There is no evidence-pack export in the product. No button here produces a file, and none will until the feature exists.
- Records are not tamper-evident today. Hash chaining is design intent.
- There is no policy engine, so no record carries a policy decision.
- We hold no SOC 2 attestation and we are not a legal opinion on any regulation.
- Records only cover instrumented traffic. Whatever runs around the instrumentation leaves nothing behind.
Questions people actually ask
Can I download a sample evidence pack?
No, and that is deliberate. Export is not built, so a downloadable "sample" would be a designed artefact pretending to be product output. The field table above is the honest version of the same information.
Which of these fields would we get in a pilot?
The live ones immediately, because they already exist. Scope and retention are agreed with you. Team, project, policy result and the integrity chain are design — if your reviewer requires them, you need a different answer today, and we would rather you hear that now.
Is an estimated cost good enough for audit?
For spend attribution it usually is, and we label it as an estimate rather than dressing it up as the invoice. For financial reporting, reconcile against the provider bill.
How does this differ from our observability traces?
Traces are for debugging: sampled, short-lived, freely editable. Evidence is for review: scoped, complete inside that scope, retained deliberately. Same events, different contract.
Where does the record live?
Today, in the Postgres database behind the product. Private deployment and residency are pilot-scoping questions, not switches we can flip on this page.
Bring your reviewer's field list
Send us what your auditor, customer or board actually asks for. We will mark which fields answer it today, which are pilot work, and which we simply do not have.