Claim Basis View#

claimBasisView(claim, evidence) answers "how was this claim's status established?" as a short line a renderer shows after the status chip, for example:

Verified   Extracted from a source · 2 entail the claim · 1 cited only
Pending review   Model-derived · 1 not evaluated
Disputed   1 contradicts the claim · Checked against expectations
No evidence   Basis not recorded

It does not restate or upgrade the status and never collapses the basis into a single confidence value. Surface owns the summary rules so every renderer shows the same line; renderers (for example the @kontourai/ui trust basis primitive, kontourai/ui#87) take the view as-is.

import { claimBasisView, missingClaimBasisView } from "@kontourai/surface/display";

const view = claimBasisView(claim, bundle.evidence); // only evidence linked to the claim (claimId === claim.id) is used

@kontourai/surface/display imports no Node-only modules, so it can be bundled for the browser. The label tables below are also exported from the root entry; claimBasisView is only on ./display.

Shape#

type TrustBasisView =
  | {
      state: "recorded";
      facets: { field: TrustBasisFacetField; code: string; label: string; caveat: boolean }[];
      detail?: { label: string; value: string }[]; // inspector rows
    }
  | {
      state: "not-recorded" | "restricted" | "unavailable" | "not-available";
      label: string; // "Basis not recorded" / "Basis restricted" / "Basis unavailable" / "Basis not available"
      detail?: { label: string; value: string }[];
    };

field is one of derivationMethod, result, supportStrength, counterevidence, method or reviewerAuthority. code is the wire enum or derived state (model, not-evaluated, cited, counterevidence, failed-not-blocking, extraction, entails, operator, …). A renderer can put field, code and caveat on data attributes.

Rules#

1. Only linked evidence bears on a claim. Not every tool call is evidence. The view uses only evidence whose claimId equals claim.id (supporting, cited, or counterevidence). Execution-trail tool calls and other claims' evidence never reach the line or its counts, whether they succeeded, failed, or were retried; they stay inspectable at the raw-execution layer.
2. Never blank. A claim with no linked evidence, no derivation edges and no reviewer (other than none) is not-recorded. A missing claim is not-available. restricted and unavailable come from the host, which knows about permission denials and failed reads: use missingClaimBasisView(state) for the canonical label.
3. At most 3 facets, except that caveats are never dropped: when there are more than 3 caveats the line holds every caveat and nothing else.
4. Caveats first, in this order:
   • Model-derived (a model derivation edge)
   • N could not run (evidence carrying inconclusive, schemaVersion 9)
   • N not evaluated (absent passing, not inconclusive)
   • N cited only
   • N contradict(s) the claim: exactly Surface's isStandingCounterevidence (entailing evidence with passing: false and blocking not false)
   • N failed (reason): every other passing: false item, labelled by why it is not counterevidence. supportStrength: "cited" gives "cited only" (checked first, since cited evidence never counts whatever blocking says); otherwise blocking: false gives "not blocking". Both reasons share one slot: "2 failed (cited only)", "1 failed (not blocking)", or "3 failed (1 not blocking, 2 cited only)". A failure is never left off the line.
5. Then method, support, review. Method is the evidence methods in Surface enum order (no depth ranking exists); more than one becomes Extracted from a source + 2 more methods. With no evidence, the non-model derivation methods are used instead, or Derived from N inputs when edges state no method. Support is N entail(s) the claim, then N with support not stated, counting only evidence that did not fail and is not inconclusive, so one item never reads as both contradicting and supporting the claim, and an attempt that never ran is not counted as cited (the inspector's Support row keeps the full partition). Review is the confidenceBasis.reviewerAuthority label when it is not none. It is producer-asserted, like the rating below, but stays on the line because a recorded review is first-class provenance.
6. Results come from passing / blocking only. Absent passing is never a pass; it is counted as not evaluated. execution.isError and a non-zero exitCode mean the check ran and failed (producers set isError from the exit code or the MCP tool result), so they are classified like any other result, exactly as status derivation does. Only an explicit evidence.inconclusive record (schemaVersion 9) means the check could not run: its result is could-not-run whatever else the item records. A schema-valid inconclusive item is cited with no passing, so status derivation never sees it either.
7. Absent supportStrength is shown as "with support not stated", as in the trust panel's evidence rows, even though status derivation treats it as entailing (the gloss says so).
8. Unrecognized wire values are shown as Unrecognized method (value) / Unrecognized reviewer (value), never as the bare string. A confidenceBasis.reviewerAuthority or evidenceStrength that is not a non-empty string (null, a number, an object, "") is treated as absent: no facet and no detail row. Every facet code is a string.
9. Producer-supplied values stay in the inspector. confidenceBasis.evidenceStrength appears only as a Producer rating detail row ("Strong support (producer-rated)"), and conclusionConfidence only as a Calibrated confidence (producer-supplied) row, worded as the probability that the conclusion is correct (its interval bounds that probability, not the claim's value). A value or interval bound that is not a finite number in [0, 1] is left out. Neither is ever a facet.

Detail rows, when there is data for them: How, Support, Results, Derived, Review, Producer rating, Calibrated confidence (producer-supplied), Collected by, Estimate, Sources. Collected by counts collectedByKind values and names undeclared items as "Collector not stated"; it appears only when at least one item declares a kind. Estimate reads the basis-annotations profile's claim.metadata.estimate: "Estimated · low–high (basis)" for a well-formed one, and "Estimate recorded but malformed; bounds not shown" otherwise.

Label tables#

All in src/display-names.ts, as *_DISPLAY_NAMES (label and one-line gloss) with *_LABELS label-only projections:

The trust panel keeps inline copies of the support, result, collector-kind and could-not-run reason labels (it ships as one bundle under a size budget); tests/display-names.test.ts fails if they drift from these tables.