0044 makes the ontology a view over what documents say

[WaylandYang]

Rewrites decision record 0044. The staged-reading draft proposed more stages around binding facts to the ontology at write time; a prototype run on 2026-09-15 found that binding at write time is where extraction errors come from.

What the record now proposes:

• Three layers. Extraction writes an open graph in the documents' own words; a small ontology is proposed by an agent and approved by people; a typed graph is computed from the open graph, with bindings decided once per signature (phrase, subject classes, object classes) and cached with the ontology version.
• Time is resolved like identity. Time mentions are facts; a document time context is carried across chunks (the document's own date, never its upload time, Upload time is used as the document date in extraction #714); the model returns an anchor, offset and granularity, and code computes the interval.
• Identity across documents is decided on evidence. Candidates, deterministic scoring with cannot-links, the adjudicator only for undecided pairs, constrained clustering, re-evaluation.
• An errata agent reviews the typed graph after extraction, through the gate.

Evidence (Re-DocRED 100 documents plus the FDA, statistics and SEC corpora; DeepSeek-V4-Flash, judged by a calibrated DeepSeek-V4-Pro): 0 of 333 open statements unstated against 4–9% of facts bound at write time; per-document slicing of a 967-property ontology keeps 86.6% of needed relations; an errata agent raised judged precision from 76.1% to 89.5%; token cost and a compact pipeline are measured in the record.

Nothing is built in the product. Stacked on #699 for 0043.

🤖 Generated with Claude Code