What Truveil is
What is Truveil?
Truveil grades individual AI agent runs from their own structured record. Each run receives a deterministic conduct grade across four dimensions, Transparency, Accountability, Data Trust and Reversibility, with findings cited to primary regulatory text and the whole record kept on a tamper-evident, independently countersigned log.
What is conduct assurance?
Conduct assurance is evidence about what an AI agent actually did during a run, graded against the regulations that apply. It differs from testing (what the agent might do) and monitoring (watching it happen): conduct assurance produces the record a deployer shows afterwards, to a customer, an auditor or a regulator.
Does Truveil certify that my deployment is compliant?
No, deliberately. Truveil never states that a deployment is compliant, safe or approved. It states what the evidence shows, cites the regulation it relates to, and stops. The legal judgment stays with you and your advisors; the evidence for it is what Truveil produces.
Who is Truveil for?
Teams running AI agents that act: moving money, changing records, sending communications, operating infrastructure. Deployers who will be asked to prove what an agent did, builders who want governance evidence shipped with their product, and advisors who need a record they can rely on.
Is the grade a permission for my agent to act?
No. The grade is evidence about a completed run, never an authorisation gate for a pending action. Truveil does not sit in your agent's request path and cannot block, throttle or approve anything. It grades what the record shows the agent did.
How scoring works
How is a run graded?
A rule-based engine scores the run's structured record against a versioned regulatory corpus across the four dimensions, applying the emphasis appropriate to the agent's declared category. Every finding traces to a specific requirement in a specific instrument, and the engine version is pinned on every record.
Is there an AI model inside the scoring?
No language model sits in the scoring path. Scores are computed by rules, so the same record under the same engine version produces the same grade every time. A model drafts only the report's narrative prose, and it cannot move a number.
Can a grade be checked later?
Yes. Every record carries its engine version and parameters, so any grade can be re-derived exactly, later, by anyone who needs to test it. An audit conclusion that cannot be reproduced cannot be defended, so reproducibility is the design requirement, not a feature.
What happens if there is not enough evidence to grade?
The run is not graded. Truveil withholds a grade rather than describing three dimensions as four, and the report says exactly which evidence was missing. Absence of evidence renders as absence, never as a finding of misconduct and never as a pass.
Does Truveil judge my agent's intentions?
No. It grades the record, not intent. A finding is a statement about what the record contains or fails to contain. That discipline is what makes the output usable as evidence rather than opinion.
What the report covers
What does an audit report contain?
The grade and four dimension scores, the agent's category context, per-signal findings with citations to primary regulatory text, remediation guidance, an instrumentation coverage line stating how the evidence was collected, and a record reliability section carrying the chain verification and the independent countersignature.
Who is the report written for?
Three readers at once, in separate registers. Auditors get the evidence trail and the reliability section. Engineers get the instrumentation coverage, the gaps, and which integration upgrade would strengthen the evidence base. Regulators and legal teams get plain-language findings with the citation behind each one.
Does the report tell me how to improve my agent?
Not technically. Truveil never advises on models, prompts, architecture or code. Its recommendations aim at one thing: bringing the run's evidence and governance in line with the applicable guidelines, naming the artifact that is missing. Closing those gaps often improves the agent in practice, but that is your engineers' work, informed by the record.
Does Truveil ever gate, block or intervene in a run?
Never. Truveil judges the happening of a completed run and nothing else. It issues no permissions, holds no actions, and cannot stop an agent. Gating belongs to your control plane; Truveil is the independent evidence about how the whole arrangement actually behaved.
Integration
How does an agent connect to Truveil?
Five paths against one per-tenant API key: an MCP connector that works from Claude, ChatGPT or any MCP client with no code, Python and JavaScript SDKs with no dependencies, a plain HTTP API, and an n8n community node. A first agent is typically logging within the hour.
Does Truveil slow my agent down or interfere with it?
No. Truveil sits entirely outside the agent's request path, so it adds no latency to any decision and has no mechanism to interfere with one. Logging is a side channel: the agent acts exactly as it would have, and the record is written alongside.
Will instrumentation change how my agent behaves?
No. Capture is observational: Truveil cannot block or alter a run, and instrumentation works in shadow, so an agent can be evaluated without touching production behaviour.
What happens if Truveil is unreachable during a run?
Your agent keeps working. You get a gap in the record, not an outage, and the gap is visible in the evidence rather than papered over. Truveil is designed so that its absence can never become your incident.
Do reports say how the evidence was collected?
Yes. Every report carries an instrumentation coverage line: which path produced the evidence, how many events were logged, and what that path structurally cannot observe. Steps reconstructed from sparse evidence are marked as inferred, distinct from steps backed by logged events.
Data and privacy
What does Truveil capture?
Structured signals: action names, declared decision types, risk as declared, checkpoint requests and their resolutions, reversibility flags, run boundaries, and the one-line descriptions your agent writes for them. Server arrival times are minted by Truveil and cannot be set by the caller.
What does Truveil never capture?
Truveil has no fields for prompts, payloads or model outputs, and it never requests them. Capture is structured signals plus short free-text descriptions, and what goes into a description is under your control, so keep customer content out of it. Designed this way, the audit layer cannot become a second copy of your customers' personal data.
How long are records retained?
Logs are retained indefinitely at present, with erasure honoured on request. A defined retention schedule will be published before the retention and erasure provisions of India's DPDP Rules, 2025 take effect. A lawful deletion reads as an erasure on record, not as tampering.
Who owns the records?
The tenant. Your runs, your records, your erasure rights. Truveil's role is custody and grading, not ownership.
Evidence and trust
How do I know a record has not been edited?
Audit logs are hash-chained and tamper-evident: verification replays the chain and halts at the first entry it cannot account for, naming it. Each account's chain head is also countersigned by an independent RFC 3161 timestamping authority, so the record's integrity does not rest on Truveil's word.
Can the agent game its own grade?
It can write flattering prose; it cannot author the evidence that matters. Every piece of evidence is weighed by who established it: an agent's own attestations about human oversight are capped whatever they claim, server-minted timestamps can falsify impossible claims, and harness-recorded capture earns standing an agent's self-reporting never does.
Can Truveil itself quietly change a record?
The design makes that detectable rather than deniable: the chain is independently countersigned, records are re-derivable, and database-level guards make log rows append-only. Honest wording matters here: tamper-evident, not tamper-proof. Nothing is beyond attack; everything is beyond silent attack.
Can I approve my agent's actions inside Truveil?
No, and that is a boundary, not a gap. Truveil is an independent third party to your agent's conduct, and an independent party cannot also be a participant in the events it countersigns. Approvals live in your infrastructure; Truveil records who approved and when, on its own clock, and weighs that evidence by who recorded it.
How is human approval actually evidenced?
Your agent raises a checkpoint before a consequential action, a person in your organisation resolves it in your systems, and the resolution is logged with its authorship stated. Evidence recorded by your infrastructure earns full standing; an agent's own claim that a human approved is capped, whatever it says.
Does Truveil detect bias?
No. Truveil's capture has no fields for prompts or model outputs, so there is nothing to inspect for bias. What it records is the governance around the decision, including whether a bias check was declared on it. Evidence that the check happened, not a substitute for the check.
Working with Truveil
What does the beta cost?
Nothing during the pilot. Pricing follows once the records have proven they are worth paying for. What we ask back is feedback and, where the records earn it, case-study rights.
What does getting started look like?
One agent instrumented in shadow, graded records the same day, and a short conversation about what they show. A mutual NDA with a data-handling annex is available at the point a real agent is instrumented.
Which regulations does Truveil score against?
Six frameworks are scored, with fourteen regulatory instruments cited, including India's DPDP Rules 2025 and the EU AI Act, and the corpus is versioned so reports state exactly which text they applied. Where a provision is not yet in force, reports carry its effective date rather than asserting a present-tense breach.
Where can I see a real report?
Sample reports, published as they came out of the engine, are at truveil.app/samples. Nothing was selected for its grade.