Skip to main content
Beta. pb.preflight is live and in active development — stable enough to build on, still being molded. For the patterns it unlocks, see Verify before you act.
pb.preflight answers a different question than search. Search asks “what is there?” Decision quality asks “is what’s there good enough to act on, for this purpose?” Over REST these are the /v1/preflight/* endpoints (check, audit, trace). A verdict is a disposition plus an enumerated list of findings. There are no confidence scores: a finding names a specific, structural deficiency, so you can act on it.

​
check

Return a verdict for a subject and a purpose. 
const verdict = await pb.preflight.check({
  subject: { type: "Account", id: "acme" },
  purpose: "send a renewal proposal",
});

if (verdict.safeToAct) {
  // proceed
} else {
  console.log(verdict.explain());
}
​
subject
{ type, id }
required
The entity to evaluate.
​
purpose
string
What you intend to do. The verdict is relative to this. Defaults to a general fitness assessment.
​
required
string[]
Slots that must be present for this purpose.
​
optional
string[]
Slots that strengthen the verdict but are not load-bearing.
A Verdict exposes disposition, findings, purpose, safeToAct, findingsOfKind(kind), and explain().

​
Dispositions

DispositionMeaning
SafeToActFit to act on for this purpose.
NeedsHydrationMissing or stale slots that can be filled.
BlockedA dependency prevents acting.
UnsafeActing would be a mistake.
SafeToRejectEnough to confidently decline.
NotEvaluatedNo verdict was produced.

​
Finding kinds

missing-required, needs-hydration, schema-violation, unprovenanced, stale, weak-edge, contradiction (reserved — the current engine does not yet emit it).

​
gaps

The findings that block acting, as a list: the gaps you need to fill. 
const gaps = await pb.preflight.gaps({
  subject: { type: "Account", id: "acme" },
  purpose: "send a renewal proposal",
});

​
repair

Turn a verdict’s findings into suggested remediation steps. repair names what is needed; it never generates content or writes. You run the steps through pb.* yourself. 
const verdict = await pb.preflight.check({ subject, purpose });
const plan = pb.preflight.repair(verdict);
// plan.actions: [{ action: "hydrate" | "capture" | "extract" | ..., target, reason }]

​
audit

Check whether applying a staged delta would degrade the graph. A read-only dry run; it never writes. 
const audit = await pb.preflight.audit(deltaId);
if (audit.degrades) {
  console.log(audit.explain());
}

​
trace

Check evidence coverage for a set of claims or entities. Pass a MemoryBriefing directly, or an explicit set. Each item is marked supported (backed by a stored source), inferred (semantically related but not directly sourced), or unsupported. 
const briefing = await pb.memory.synthesize({ query: "deal status" });
const trace = await pb.preflight.trace(briefing);

console.log(trace.grounded);     // true only if every claim is supported
console.log(trace.unsupported);  // claims with no backing
pb.preflight is read-only: evaluating quality never writes to the graph.