FAQ

Does Provantix replace Langfuse or LangSmith?

No, it sits above them. Observability platforms record traces and administrative activity: what your agent did technically. Provantix is aimed at a different record: verified business actions tied to authority. It binds those actions to real organizational identities and packages the result into verifiable evidence for your audit trail. We import from Langfuse (LangSmith import is rolling out) and also accept direct telemetry, so most teams run both.

What exactly does an event prove?

Three things, bound together: the workload identity that acted, the human or team that owns it, and the credential that authorized it, plus a chain-of-custody hash linking the event to every related event. An auditor re-verifies the chain from an export on their own machine, and checks our signature on the export against our published keys.

If a check passes, does that mean the record is true and complete?

No, and it is worth being exact about the two limits.

  • A valid signature proves who signed, not that the content is true. It shows an export is what Provantix issued, and that a record is what the emitter signed. It does not show the agent reported what really happened.
  • Integrity is not completeness. Intact records and an unbroken chain show nothing was altered or removed inside what was recorded. They do not show every action was captured. Missing batches are reported as gaps, and a statement about coverage is only as wide as what was actually checked.

Both limits apply to every export. See Trust and How verification works.

How does Provantix check that an action really happened in the payment system?

For actions in a reconciled scope, Provantix reads the payment system's own records (today Razorpay, for refunds) and compares them with what the agent claimed. Each action then has one of five states, with the reason: destination-confirmed, conflicting, pending, unverified or agent-reported. A refund the system never recorded is unverified, not a failed action, and a destination that is down is shown as health, never as a fault of the agent. This is limited to the scope, the period and the window that were actually collected, and it does not show that the payment system's record is itself true. See Destination confirmation.

What if my agent can't reach the internet?

The SDK queues signed batches in a local spool directory and drains when connectivity returns. Payloads are signed at capture, so delayed delivery can't be tampered with in transit. Gaps are surfaced explicitly as custody-gap findings; missing evidence is never silently smoothed over.

Can Provantix block or stop an agent?

No, and that is deliberate. Provantix is read-only: it observes, attributes and packages evidence, never sitting in the execution path of your agents. Inventory states like "quarantined" are labels for your own workflow, with no side effects.

Which compliance frameworks does this map to?

The evidence model is designed to support SOC 2 change-management and monitoring controls (CC7, CC8), ISO 27001 audit-logging and access-control clauses, NIST AI RMF Govern and Map functions, and EU AI Act Article 12 record-keeping. The honest phrasing is "designed to support": Provantix produces the evidence, your auditor makes the attestation.

What data does Provantix store?

Only what your agents emit and what identity connectors read: event metadata (action, target, timestamps, credential), identity records, and findings. Prompts and model outputs are included only if you choose to send them. Embeddings of event text are stored for semantic search, inside your organization's own database.

Does Provantix tell me if my agent's answer was right?

No. Nothing in the record grades content quality — see what Provantix does not tell you. outcome: success means the call completed without an error; an agent that returns cleanly after doing the wrong thing still records success.

Do you store my prompts?

Not by default. Prompts and completions are stored as hash-only fingerprints — enough for you to prove a match against the original you hold, never the content itself — unless you explicitly choose to send full content. Tool arguments and results are redacted and truncated into erasable summaries, sealed with the record, and erasable on request.

Are cost figures billing-accurate?

No — they are estimates. Cost is computed at ingest from the token counts the emitter reported and the prices in your organization's price table (each with a source and effective date). Editing a price restates the estimates of the records it priced, and nothing reconciles the figures against the provider's invoice. Treat cost as directional for chargeback and trends, not as a billing record.

What Provantix does not tell you

Two limits are worth stating plainly:

  • It does not judge whether your agent's answer was right. Outcome fields are technical: success means the call completed without raising an error, not that the task succeeded, and nothing in the record grades content quality.
  • Content is hash-only on the hosted OTLP intake. Prompts and completions are stored as hashes (so a customer holding the original can prove a match) and the hosted intake never keeps the text.

Do you have a Python SDK?

On the roadmap. Until it ships, Python services can POST batches directly (see Direct HTTP API), or shell out to the Node SDK. If Python is blocking a rollout, tell us at [email protected] and we'll bump it.