Derived vocabularies drop relations for rarely stated facts (the 25:50 personal best)

[antejavor]

Question

Does a derivation keep a relation for a fact that the user states only rarely? The test case is 25:50, the user's personal best.

Where it stands

The value-only pass decided in #386 fixed the typing: 25:50 is now Duration in both derived seeds. No edge lands on it, though, because neither derived vocabulary keeps a relation for it:

• S1 proposed improved_personal_best_by and race_finish_time from the one propose-sample session that mentions a personal best.
• The 60-session observe sample has no instance of either, so prune dropped both.
• S2 never proposed one.

Evidence: read_value_pass.py and read_value_pass.log on prototype/gliner2-constrained-edges (c07155a).

Why it's left open

We expect the derivation schedule (#433) to cover this, but haven't shown it. Each run reads the user's own delta, so the session that states the fact is in the data that gets observed. The held-out evaluation excludes those sessions on purpose. This issue is intentionally open to reconsider once the build lands. Ways to check:

• Re-derive with the evidence session included in observe.
• Read sparse-fact recall on the BEAM re-run (PR context-graph: BEAM eval and passage-level recall #438).

If it doesn't hold, the levers are:

• A lower evidence floor for relations whose intended tail is a core value type.
• Keeping a proposed value relation in the pool rather than dropping it.

[antejavor]

New evidence from the observer work (#448): on the real GLiNER2 model with the hygm default ontology, neither the main pass nor the value-only pass types 25:50 as Duration in "My best 5K time is 25:50 and I ran it in Berlin last May." The value pass returns only user and last May. The hand vocabulary's personal_best relation did find it. Whether the span is found at all seems to depend on there being a relation for it, not only on typing competition. This strengthens the case for checking relation coverage once derive runs on real user deltas.