Every Insider document separates four things that are usually conflated: what was observed, whether the observation completed, how certain it is, and what it is allowed to mean. This page gives the vocabulary and the reading order so an agent does not promote one axis into another.

Reading order

  1. Schema and identity. Check schema is the one you expected and, for content-addressed documents, that id equals "sha256:" <> integrity.content_digest and that the appropriate validate_with_* function accepts the document with its exact inputs. Standalone validation proves structure and content identity only.
  2. Operational status. Did the operation run to completion? Protocol results: ok, error, timeout, cancelled, unsupported. Confirmation records: operational_status of completed, cancelled, error, timeout. CLI exits: 0, 1, 124, 130, 3. A non-completed operation can still carry a true partial finding, but never a success.
  3. Domain outcome. The thing the operation was asked about. Oracles: changed / unchanged / inconclusive (or increased / unchanged / decreased / inconclusive). Confirmation: matching_failure, nonreproduced, repaired, unrelated_failure, infrastructure_failure. Verification: verified or invalid. Minimizer and bisection: completed, exhausted, inconclusive, with a reason.
  4. Bounds and truncation. truncated, truncation_reasons, dropped, requested_limits versus effective_limits, consumption, bounds, *_truncated, *_observed. A bound that was hit narrows every conclusion to what fit inside it.
  5. Cleanup. not_required, complete, partial, uncertain, unknown, or for searches complete / incomplete / unknown. Only complete and not_required mean nothing was left behind.
  6. Confidence and uncertainty. confidence (certain, high, medium, low, unknown) and the uncertainty list of named facts the component could not establish. Read every entry; they are the component telling you where its claim stops.
  7. Controls and limitations. Fixed per schema. controls say what the artifact cannot do; limitations say what it cannot know. They do not vary per run and are not a verdict; they are the contract you are reading under.

Words that are not success

WordWhat it meansWhat it does not mean
candidatea static signal worth investigatinga defect, a cause, an impact
syntactican identity read from source without macro expansiona runtime fact
changeda global differed between the two captures of one runwhich test changed it, that it matters, that it reaches production
inconclusivea bound or read failure prevented comparisonunchanged, clean, or safe
operational_failurethe artifact never observed anythingany statement about the target
nonreproducedthe scenario ran and the expected finding was not observedthe defect is absent; the scenario recipe ran, which is all completion means
matching_failurethe expected finding was observed on a disposable targetcausality, production impact, or a repair direction
repaireda configured positive oracle reported repairrepair of anything outside the declared bounded scenario
failed (regression)two fresh, distinct runs both reproduced on the same baselinethat the record's asserted execution occurred; a trusted runtime must have produced the live records
completed (minimizer)no single remaining whole step can be deleteda globally smallest reproducer
adjacent_boundary_isolatedone adjacent nonreproduced/reproducing pair on the search pathearliest reproducing target, or cause
not_proventhe search exhausted its budget; the best bracket so farfailure of the search
verifiedboth after-runs passed with exact provenance and complete cleanupgeneralized bug freedom; proof of the declared scenario only
invalidat least one verification reasonthat the candidate is wrong; it may be a cleanup or provenance problem
declared_unverifiedcandidate bytes bound to a handoffanything about the candidate's effect
eligible (proposal)the evidence graph recomputespermission to publish; publication=blocked stands
provider_reported_createdwhat the test provider saidindependently verified remote state
passed (load)every structural gate held on this hosta production budget or capacity
no_canary_hitsno declared token surfaced from the supplied itemsapplication coverage or redaction approval
mock: true, not_attemptedthe observability mapping is deterministic datadelivery or integration

Fingerprints and identities

  • sha256: identities are content, not authenticity. Two equal digests mean equal bytes; they do not say who produced them. Trust comes from the configured adapter, provider, or host that validated the inputs.
  • Oracle hmac-sha256: fingerprints use a random per-run key that is never exported. Compare them only inside the report they appear in.
  • Provenance has five components (code, configuration, dependencies, runtime, target), each available, unavailable with a reason, or redacted under a named policy fingerprint. An unavailable component is not a mismatch; it is a documented gap, and the manifest keeps unavailable and redacted distinct.
  • context in an oracle report is whatever the caller passed in. It is echoed, never verified, and every report says so.

Binding evidence to the world

No oracle, inventory, or load report binds itself to the test you ran, the revision you ran it on, or the ticket you are working. The inventory gives you source_fingerprint and candidate id; the oracle gives you context; the load report binds git_sha and tree state but only for the host it ran on. You must record the command, the exact selected test, the revision, and the inventory identity next to the report. caller_must_bind_test_and_revision appears in every oracle report for this reason.

When to stop

Stop and report rather than proceed when:

  • a 70 exit, a missing oracle line, or infrastructure_failure appears: the environment is broken, not clean;
  • an inventory's completeness is anything but complete (exits 3 to 5) and you were about to treat a missing candidate as absence of risk;
  • scanner_revision differs between inventories you are comparing;
  • cleanup is anything but complete or not_required on a step whose next step assumes a clean executor;
  • uncertainty names the exact thing your next inference depends on;
  • a syntactic identity has not yet been validated against the real module environment;
  • a record validates standalone but you have not run its validate_with_* with the exact inputs.

An unconfirmed finding does not authorize a repair, and a verified repair does not authorize publication. Those are the two sentences the whole chain is built to keep true.