IDEA-014 D3 and R4 results

Run 2026-11-04 under the spec committed at 921cc35 (deal-r4-d3-spec.md). Past data; nothing here is a skill claim. The ranker, deal-v0.1 and every earlier number are unchanged.

Short version

• D3: the acquisition label misses a lot. An estimated 44 percent of true acquisitions in the R3 pool are absent from the ledger (95 percent interval 9 to 72 percent), over the 20 percent limit. R3 stays descriptive, now with measured noise instead of "unmeasured". Its headline does not change.
• R4: stopped at its gate, no label built. 60 percent of sampled trials were clean against a 70 percent floor. The pre-registered rule fired. No R4 label was computed and no score was read, so R2 is still "cannot test".
• What R4 did teach: the names are fetched, and the exclusivity rule as written cannot pass its gate. A redesign is possible, and no R4 label has been seen, but it needs its own pre-registration.

D3: how much of R3's label is missing

Sample: 30 of 406 at T = 2020-12-31 (seed 0), 1 ledger positive and 29 negatives. Lookups: three blind searchers given names and the window only (docs/audits/idea014-d3-lookups.json). Four claimed events went to a separate adjudicator who did not see the ledger.

Result: above 20 percent, under either reading of Lannett. R3 stays descriptive. D3 could not change R3's headline: rubric minus trial count remains -0.005 (-0.079 to 0.069).

Weaknesses:

• Thirty companies cannot pin the rate down. The interval runs from under the limit to about three-quarters. The rule needed zero misses to pass; two were found.
• Lannett was a debt-for-equity restructuring, not an acquisition in the usual sense. The ledger records it as delisted_failed, so the ledger saw the exit and typed it differently. A strict reader would call that a typing difference, and one miss would remain. Its new owners are a lender group with no single registrant.
• Misses by SEC-registered acquirers, which the spec asks for separately: 1 (Panbela). Lannett cannot be placed in that split.
• Most "not acquired" answers rest on absence of evidence, some from search snippets the searcher did not open. A deal missed by the searcher would lower this estimate. HedgePath's 2022 share cancellation was flagged as arguable and was not adjudicated.
• Whether Cancer Prevention Pharmaceuticals truly belonged in a "listed" pool was not checked. It entered through a Form 10 row.

R4: did the target's asset advance after the deal

The fetch ran as specified: intervention names for 130,528 of 130,531 cached industry studies (99.998 percent, floor 95), anonymous, free, names and types only.

The linkage gate then failed. Reader's count of the 50 sampled pre-deal trials (data/processed/acquire/r4_gate_sheet.json; pool 1,275 by this count, 1,276 in the plan, difference not traced):

The rule fired: R4 stays descriptive and no label was computed. The cap was visible before the reader ran, since only 34 of 50 had any exclusive name. I checked the 16 empties for a code bug and found none; the rule does what the spec says.

Why it failed (the executor's reading, from a diagnostic listing the other lead sponsors on each emptied name; not the reader's verdict and not a tested claim):

• 9 of the 16 empties are drug trials the exclusivity rule emptied. The cause varies and was not confirmed row by row. Some look like the flaw: etrasimod is dropped because of Pfizer, tesevatinib because of "Kadmon, a Sanofi Company", MK-4827 because of Merck, which would be handovers R4 exists to follow if those trials moved or the drug was licensed. Others are not handovers: sumatriptan and dihydroergotamine are used by unrelated sponsors, and there the rule worked as meant.
• 7 of the 16 have no drug name to find: devices, observational studies or empty records. Limit 4 in the spec predicted this, but the gate's denominator still counted them.
• 3 of the 4 dirty rows are combination arm labels, such as "ngm707pluspembrolizumab". The fourth is a silicate and cellulose name the reader called a placebo alias.
• Among the 34 trials with a name, 30 are clean (88 percent). This is a post-hoc subset, and not a gate result.

Code readings, logged in decisions.md before any name was read: sponsors on the target's own pre-deal trials do not count as competing use, which loosens the spec's wording, so it cannot explain the failure; and the generic list matches by "contains", which is broader than the list as written.

What this changes

• R2 remains "cannot test" (4 positives of 69 under the old label). R4 did not replace it.
• deal-v0.1-results.md limit 2 is superseded by the D3 figures above. r3.json still prints "label noise unmeasured", because no R3 number was recomputed.
• Any negative in R3 may be a missed deal at a material rate, so a rubric or LLM AUC on that label is a weak read in either direction. This also applies to the blind LLM row.

What would reopen R4

Superseded 2026-11-04: v2 ran and stopped at its positive audit (deal-r4-v2-results.md).

A separately pre-registered v2, written before any label is counted, with a rule that does not treat a successor's trial as a competing use. No R4 label has been computed, so that is still possible, but the v2 rule would be written knowing this 60 percent, and the result would have to say so. Restricting the gate to drug-bearing trials changes the question it answers; that choice has to be stated, not slipped in.