IDEA-014 R4 v2 results

Run 2026-11-04 under the spec committed at 42b8b19 (deal-r4-v2-spec.md). Past data; nothing here is a skill claim. The ranker, deal-v0.1 and every earlier number are unchanged.

Short version

• The linkage gate passed: 38 of 50 fresh trials clean (76 percent, floor 70), and a second reader reached the same 38.
• The positive audit failed: 5 of 10 candidate positives are true (50 percent, floor 70). The spec's rule fired: R4 stays descriptive.
• No scorer AUC was read under this label. The read code was removed unrun. R2 stays "cannot test". Whether past deals worked is still untested.
• Why it failed: the rule tracks drug names, not molecules. Four "advances" were molecules already at phase 3 before the deal under a different name.
• This line is closed under its own stopping rule. A further attempt needs new data, namely a molecule-level asset map, and new owner approval. **Update 2026-11-04: that attempt ran as v3 and stopped at its map gate; see deal-r4-v3-results.md.**

Linkage gate

Sample: 50 NCTs, random.Random(1), from the 1,225 pre-deal NCTs not in v1's sheet. 16 of the 50 belong to companies that also appear in v1's sheet.

• Among the 43 trials with a kept-type intervention (non-binding): 38 clean, or 88 percent.
• v1 scored 60 percent on a different sample with a different rule, so the two figures are not a clean before-and-after. The v2 rule was tuned on v1's sheet.
• Why there are two reads: the backtest-auditor flagged the first file because its reasons were nearly identical across rows and it was written quickly. The rule that both reads must reach 35 was set before the second read ran, and it is stricter than the spec. The second reader was told to identify each name row by row. It agreed with the first on the counts and on every dirty and empty row they each listed. Both are Sonnet agents, so their errors may be correlated.
• Order, stated plainly: the label and candidate counts were computed straight after the first read passed, as the spec allows. So they were already seen when the second read ran; had it failed, this order would have mattered. No score was read. The claim that the two-read rule came before the second read rests on this note and decisions.md, not on git, since nothing was committed between those steps. The rule could not have loosened anything: both reads reached 38.
• The margin is 3 trials. Files: docs/audits/idea014-r4v2-gate-verdicts.json, docs/audits/idea014-r4v2-gate-verdicts-second.json.
• The 4 dirty rows show what the majority rule still lets through: comparators and background drugs that few others registered under that exact string, such as timololmaleateophthalmicsolutionbid.

Label and audit

Of 69 uncensored targets, 16 have no asset name, 25 have a later study naming an asset, and 10 are candidate positives.

Audited: 5 true of 10, 50 percent. Below the 70 percent floor. Minimum detectable AUC at 5 of 69 is 0.878, above the 0.85 line, so "cannot test" would also have applied. Without Alder, 4 true.

Exception list: docs/audits/idea014-r4v2-exceptions.json. Agent audit: docs/audits/idea014-r4v2-positive-audit.json. Pre-read audit and review: docs/audits/idea014-r4v2-review.md. For Akcea and Constellation, the alias equivalence is the executor's outside knowledge, supported by the titles on the local rows.

How the audit got from 9 true to 5

This matters because the first audit passed and the outcome rests on the corrections.

1. The audit agent called 9 of 10 true, and it applied the higher-phase test to the name's own maximum, as its packet showed it.
2. The backtest-auditor, reading before any score, flagged Tesaro and Spark as aliases of molecules already further along.
3. The executor then checked all nine by listing every alias of each molecule across cached pre-deal studies. Aliases came from the executor's own knowledge. That found Akcea and Constellation, and it confirmed the other five reach no higher than the phase the rule used. Each removal rests on a dated registry row cited above.

Limits on this:

• The spec's audit wording says "the same molecule at a higher phase", so the removals apply the spec as written. No score had been read when they were made. The audit is still by agents: nobody human checked it.
• The alias check covered only candidates. It can only remove positives, as the spec requires.
• False negatives were never audited. Trials re-sponsored by an acquirer, and the 16 targets with no asset name, can hide real advances. The true positive rate on this population is therefore unknown; 5 of 69 is not an estimate of it.

What this means

• The gate measured whether the names are the target's own. It did not test whether a name's phase history is the molecule's. Both are needed, and the second is what failed.
• Two of the four false advances (Akcea, Constellation) were cases where another sponsor's phase 3 used a different name. That is the same 2026-sponsor and handover problem v2 set out to fix, in a new place.
• A usable label needs one row per molecule, with its aliases and its pre-deal maximum. That is new data, and the stopping rule requires new owner approval for it.
• The read code was written, never run and never committed, which git cannot show; data/processed/acquire/r4v2.json does not exist. This result says nothing about whether the rubric or any baseline predicts asset progress.