Demand pull: no prediction test was run (IDEA-006, draft, not published)

Superseded in part, same day: the owner authorized a separate second test with a new vocabulary. It ran and found no detectable gain: docs/methodology/demand-pull-v2-result.md. This note stands as the record of the first attempt.

What we set out to do. Test whether dated buyer behaviour (industry trials by therapeutic area, from ClinicalTrials.gov) improves the lab ranker beyond the v4 features. The hoped-for result was a better prediction algorithm. It was not obtained, and the ranker is unchanged.

What happened. Before any count, we froze a 16-area vocabulary and two coverage floors. Labs passed (66.8% against 40%). Trial conditions missed: 69.16% against 70%, 85 trials short (6,901 of 9,979; the floor needs 6,986). The pre-registered consequence is that R2 is not run. The director held to it: a new vocabulary aimed at the same floor, after seeing the count, is exactly the forking path the floor exists to close. No demand number was ever joined to a label, so this says nothing either way about whether demand pull predicts anything.

What does exist.

• data/diaspore_demand.db: a dated, sourced table of industry-trial counts by area at 2014, 2019, 2021 and 2025, with a buyer subset that excludes sponsors known as startups by each cutoff. Descriptive only; every use states the 69.2% trial coverage.
• demand.pull(text, as_of): counts and provenance for IDEA-014's later rungs. No market size, value or score.
• A blind check of the vocabulary's accuracy: 94% on trials, 86% on labs. 0 of the top 50 sponsors were startups. The vocabulary is accurate; it just reaches slightly too few trials.
• A power estimate: pooled minimum detectable AUC difference of 0.0247 at S0 (a proxy from a nested 9-feature block), so a run could have detected about that and no less.

What it would take. A separately pre-registered test with a fresh vocabulary and floors fixed before its count, authorized by the owner (OWNER-QUEUE). Even a pass would earn only a v5 prospective freeze. It is not a skill claim until the prospective reads in 2028 and 2030.

Residual. The registry cache is a 2026 snapshot: sponsor, conditions and phase are 2026 state; only posted is dated at registration.