Policy replay
AvailableTest governance changes against reality before deploying them.
Replay historical actions under a proposed policy to see what would remain allowed, require approval, or be blocked.
Product · Control Plane
Give agents authority without giving away control.
Lucidrail sits between AI agents and consequential actions. Every call is attributed to an identity, evaluated against policy and live context, and allowed, held or blocked — then verified and recorded.
- 01IdentifyWhich agent, for whom
- 02EvaluatePolicy, risk, provenance, spend
- 03DecideAllow, hold, restrict, block
- 04ExecuteOnly inside the envelope
- 05VerifyAgainst system state
- 06RecordStructured audit evidence
Identity
Identity for non-human workers.
Every agent is a first-class principal with an organization, a role, a declared purpose, a tool set, limits and a responsible human. Decisions and evidence attach to that identity — not to a shared API key.
Governance · Authorization
Policy for verbs that matter, evaluated in context.
Policies govern what agents may do with each system. The same rule can resolve differently depending on the agent’s capability, the task’s provenance, the argument values and current spend.
# illustrative policy syntax policy external-comms/contaminated-trajectory when action = email.send and recipient.domain ∉ org.domains and task.provenance = contaminated then require_approval(agent.responsible_human) evidence task.sources, draft, recipients
Approvals
Approve the action, not an abstract permission.
Reviewers see the exact proposed action and the evidence behind the hold. They can approve, reject, edit, ask the agent to explain, or inspect sources. The decision is bound to that action — and to the tool schema it was approved against.
Try the approval queue →email.sendThe actual draft, recipients and attachments.
payments.transferAmount, beneficiary, invoice and ledger evidence.
db.mutateThe statement and the rows it would affect.
github.deployDiff, CI evidence and the target environment.
Provenance
Trust follows the trajectory.
Action provenance records the information sources, tool outputs and external inputs that preceded each action, their trust state, and the downstream actions that depend on them. Lucidrail works from observable inputs and outputs; it does not claim access to a model’s hidden reasoning.
calendar.readinternal
web.fetch vendor-portalexternal
inbox/msg_8812.pdfcontaminated
draft.composeinherits
email.sendheld
Audit
Evidence, not log lines.
Every consequential action produces a structured record you can query, export and hand to a reviewer.
actoractionpolicydecisionapprovalevidenceresulttimestampenvironment
Cryptographic signing of records is applied only where a deployment’s key management supports it; we document exactly which records are signed.
Spend controls
Budgets that act before the invoice.
Per-agent budget$1,284 / $2,500
Per-team budget$31,900 / $40,000
Model spend · escalation at 80%$6,400 / $6,000
Also: tool spend, per-transaction limits, rate limits and escalation thresholds. Illustrative data.
Verification
“The agent says it worked” is not evidence.
Lucidrail records what the agent claimed separately from what the system confirmed. Mismatches are surfaced to the responsible human and count against the agent’s performance signal.
HTTP evidenceStatus, headers, response body
Browser evidenceDOM state after the action
Database stateRead-back of the affected rows
Logs & artifactsBuild output, files, hashes
Independent verifierA separate check of the outcome
Observability
Configured isn’t the same as enforced.
For each control, Lucidrail shows whether it is configured, whether it is in the path and evaluating, how often it triggered, and the most recent evidence. A control that has never produced evidence is flagged — not assumed to work.
See control effectiveness in the console →ENABLEDA setting says it’s on.
OBSERVED WORKINGEvidence shows it acted.
Mechanisms