import type { TStageContext } from "../../../../lib/pipelines/index.js";
import type { TClaimMentionExtractionOutput, TSegmentationOutput } from "../stages/schemas.js";
import { type TIngestionSourceAnchor } from "../source-anchors.js";
/**
 * Where each segment actually begins in the input.
 *
 * The segmentation prompt requires `segment.text` be copied verbatim, so
 * the true offset is recoverable by searching — and searching is what
 * this does, because the model's own `span.start` drifts — off by one on
 * several segments of the recorded corpus. Small, and invisible until a
 * quote repeats, at which point the drifted hint picks the wrong
 * occurrence.
 *
 * Segments are emitted left to right and cover the input, so the scan
 * carries a cursor rather than searching from zero: that keeps a segment
 * whose text repeats earlier in the document from collapsing onto the
 * earlier copy. Among the candidates at or after the cursor, the one
 * nearest the model's reported start wins — the model's number chooses
 * between verified positions rather than supplying one.
 */
export declare function resolveSegmentStarts(inputText: string, segmentation: TSegmentationOutput | undefined): Map<string, number>;
/**
 * Non-fatal note codes for anchor resolution. None stops assembly: an
 * unresolved quote yields no anchor; an ambiguous one yields the
 * occurrence nearest the reported position when the quote is verbatim,
 * and no anchor when it is only approximately in the input; an approximate
 * one yields the nearest passage. Every one of them is reported because
 * silence here is indistinguishable from success — a model that starts
 * paraphrasing would otherwise take anchor coverage to zero with no signal.
 * A claim the mention stage produced no mention for was never looked up at
 * all; it is reported too, so "we did not look" is not read as "we found
 * nothing" or as success. So is the reverse: a mention the stage produced
 * that no claim references, whose passage is linked to no claim. A mention
 * id the stage emitted more than once is resolved from its first copy only,
 * so it gets at most one resolution note; later copies are ignored.
 */
export declare const SOURCE_ANCHOR_NOTE_CODES: {
    readonly unresolved: "SOURCE_ANCHOR_UNRESOLVED";
    readonly ambiguous: "SOURCE_ANCHOR_AMBIGUOUS";
    readonly approximate: "SOURCE_ANCHOR_APPROXIMATE";
    readonly inputUnavailable: "SOURCE_ANCHOR_INPUT_UNAVAILABLE";
    readonly notAttempted: "SOURCE_ANCHOR_NOT_ATTEMPTED";
    readonly mentionUnclaimed: "SOURCE_ANCHOR_MENTION_UNCLAIMED";
};
/** Resolve every mention to a verified anchor in the input, by id. */
export declare function buildAnchorByMentionId(args: {
    ctx: TStageContext;
    inputText: string;
    segmentStartById: Map<string, number>;
    segments: TSegmentationOutput["segments"] | undefined;
    mentions: TClaimMentionExtractionOutput | undefined;
}): Map<string, TIngestionSourceAnchor>;
/** Anchors for the mentions of one claim, in order, deduped by range. */
export declare function claimAnchors(mentionIds: readonly string[], anchorByMentionId: Map<string, TIngestionSourceAnchor>): TIngestionSourceAnchor[];
/**
 * The raw text this pipeline was given, or `""` for any input shape that
 * does not carry one.
 *
 * `TStageContext.input` is `unknown` and this assembler is public API
 * for consumers composing their own pipelines, whose `inputSchema` need
 * not be `{ text }`. Anchors are the only thing that reads the input, so
 * an empty string degrades exactly into the documented no-anchor
 * behavior — every match misses — rather than throwing out of finalize
 * on the happy path, after every LLM call has been paid for.
 */
export declare function readInputText(input: unknown): string;
/** The `mentionIds` on a canonical claim record, defensively read. */
export declare function readMentionIds(record: Record<string, unknown>): string[];
//# sourceMappingURL=source-anchor-resolution.d.ts.map