A quotes-route rollout grounded more claim uses without earning claim identifiers

Type: kb/types/note.md · Tags: kb-maintenance

Commonplace's 2026-08 grounding rollout checked source-dependent claim uses in notes against their ingested sources in successive cohorts. Cohorts 01–10a used a normalized claims ledger: each ingest entry held a paraphrased claim, exact extracts, and authored scope, confidence, and limitation fields, and a review compared the note's wording with the paraphrase. Cohort 10b used the direct model adopted by ADR 073: the note's use is judged against retained verbatim quotes or against the exact pinned snapshot. The unit of comparison is one source-dependent claim use in a target note, not a note, ingest, quote pair, or link.

Terminal dispositions

Population Claim uses Grounded, n (%) Narrowed Contradicted or repaired Retained local delta Literature handoff False positive Unavailable
Normalized claims ledger, cohorts 01–10a 205 62 (30.2%) 93 15 24 9 1 1
Quotes or snapshot, cohort 10b 59 44 (74.6%) 2 1 5 7 0 0

Cohort 01's table was reconstructed from committed artifacts and review state after its workshop closed.

What this establishes

Under the direct model, a much larger share of claim uses ended grounded and a much smaller share ended narrowed. The rollout also recorded no case that needed a stable claim identifier, semantic deduplication of quotes, or reconciliation between overlapping entries. The normalized-ledger cohorts showed reuse and denser sections; cohort 08 was the nearest counter-signal, where partly overlapping premises created modest reconciliation pressure, but exact paraphrase plus scope and limitation still selected every use. In cohort 10b, six atomic recoveries appended 51 distinct verbatim source and location pairs; similar passages with different local contexts coexisted and no accepted pair required an identifier. A deterministic scan on 2026-08-25 found 19 link occurrences carrying the (snapshot required) marker across seven of 68 frozen target notes, each with its exact-byte snapshot present, so the snapshot route was used rather than merely available.

For a corpus of this size and shape, the direct model did not earn claim identifiers, deduplication, or reconciliation machinery. That is the inference this record supports.

What this does not establish

The cohorts were successive and non-random, differed in content and execution phase, and ran under non-equivalent review protocols. The gap between 30.2% and 74.6% is descriptive, not a causal estimate of the representation's effect. Neither rate tests whether a grounded use preserves the source's proposition. The record supplies no threshold at which a larger, more repetitive, contested, or multi-writer corpus would need identifier or reconciliation machinery, and does not show that such a corpus never would.

Relevant Notes: