IDEA-014 R4 v2 specification

Written 2026-11-04, before any v2 gate draw, label count or score read. Owner approval in chat, 2026-11-04: "Yes run R4 V2", on the OWNER-QUEUE row "R4 v2". No new fetch and no spend.

This is a second attempt, and it is not blind to the first. Known when this was written:

• R4 v1 failed its linkage gate at 60 percent against a 70 percent floor (deal-r4-d3-results.md).
• The v1 gate sheet of 50 trials was read in full and used as a development set for this rule, using pre-deal rows only.
• On those 50, the v2 rule leaves 42 with a name, and the executor's informal read is around 80 percent clean. That read is tuned on those rows, so it is not evidence.
• The R2 scorer AUCs under the old label are known (deal-v0.1-results.md), including the blind LLM's 0.68. The same scorers are read again here under a new label.

No R4 label has been computed under either rule, and no later study has been read for any asset.

What this can show. On past data the project has handled, so it is not a skill claim. Later trial progress is not deal value. Neither read changes the ranker or deal-v0.1.

What changes from v1

Everything not listed here stays as in deal-r4-d3-spec.md: population, fetch, name types, dose and parenthesis rules, 36-month horizon, censoring, primary and secondary label wording, gate splits, baselines and decision rule.

1. Combination labels are split before normalising. Splits on +, plus , with , and , /, ; and , (comma then space). Each part is normalised on its own. So "NGM707 plus pembrolizumab" gives ngm707 and pembrolizumab. Every study's names go through this, in the index as well as the asset set.
2. "dummy" joins the frozen generic list. Nothing else is added.
3. Majority rule replaces exclusivity. A name is an asset of the target only if the target's own trials are strictly more than half of all cached trials naming it that were posted before the deal. A trial's sponsor is not read. Exactly half is not a majority. Reason: registry sponsors are 2026 state, so v1 counted a successor's trial as a competitor. A comparator or background drug is mostly used by others and still fails this test.
4. An asset's pre-deal maximum phase is read across all sponsors. v1 could use the target's own trials because its names were exclusive. Under the majority rule another sponsor's pre-deal trial can exist, and a later study below its phase is not an advance.

Linkage gate

• Sample: 50 NCTs by random.Random(1).sample(sorted(pre-deal NCTs minus v1's 50), 50).
• Reader: a path-mode Sonnet subagent that did not write the rule and has not seen v1's sheet or this spec's reasoning. It gets only the NCT id, title, interventions and asset set, with no score or label.
• Clean: at least one asset name, and no asset name that is a comparator or background drug. Same wording as v1.
• Binding threshold: at least 35 of 50 clean, on all 50. The floor and the denominator are unchanged from v1. Trials with no drug name still count against the rule, so the ceiling is below 100 percent.
• Non-binding rows, always reported: the share among trials with at least one kept-type intervention, and how many sampled trials belong to companies that also appear in v1's 50. Trials are drawn singly, so companies seen in development can recur.
• Under 35: R4 stays descriptive, no label is computed, R2 stays "cannot test", and this line closes (see Stopping).

Label, audit and read

• Label: a study by any sponsor, posted in (deal, deal + 36 months], naming an asset at a phase strictly above that asset's all-sponsor pre-deal maximum. Dates split only through diaspore.backtest.gate. Secondary: any study naming an asset in that window.
• Targets with no asset name count as label 0, as in v1 limit 4, because the population is unchanged. Always reported: how many there are, and the same AUC table excluding them. Non-binding.
• Positive audit: before any score is read, an agent that did not write the rule checks every candidate positive against the pre-deal study and the later study. True means the later study tests the same molecule as an investigational product, not a comparator or background drug, at a higher phase. False positives go on a fixed exception list saved with the file. The list only removes positives. Under 70 percent true: R4 stays descriptive. This check is by agent and is labelled "agent-checked", not hand-checked.
• Also reported after the audit: how many positives rest only on a phase 4 study after a phase 3 maximum, since that can be post-marketing work and not advancement.
• Mode: minimum detectable AUC from audited counts by min_detectable_auc. Above 0.85: "cannot test", and no scorer AUC is read under the new label. With 69 rows that means at least 6 audited positives.
• Auditor: backtest-auditor runs after the audit and before any scorer AUC is read.
• Decision rule, unchanged: passes only if the 95 percent lower bound of AUC(rubric) minus AUC(trial count) is above 0. There is one row per company, so resampling companies is the clustered bootstrap: 10,000 draws, seed 0. Otherwise the result reads that the rubric adds nothing over trial count. Any change is deal-v0.2.
• Baselines, unchanged: random (seed 0), stage only, trial count, price rank by low edge on the priced subset, rubric without economics (identical row), and the stored blind LLM scores with their existing caveats.

Stopping

This is the last attempt at this name data under this question. If the gate, the audit or the mode check stops it, the result is reported as it stands and R2 stays "cannot test". A v3 would need new data, such as registry record history or a hand-curated asset map, not a further rule tweak. It would also need new owner approval.

Known limits, stated before the read

1. All v1 limits still apply: names are 2026 registry state, and advancing can reflect the acquirer's budget and is not value or a return.
2. Target trials link by 2026 sponsor name. If an acquirer re-sponsored most of a target's trials before our snapshot, the target loses its majority and the asset is dropped (etrasimod at 8 of 16 is an example from development). This undercounts positives, and it hits the handovers this label is meant to follow. Some of those handovers leave a shorter trace: the same asset's code name (apd334 at 8 of 15) can survive.
3. A generic reformulated by the target can pass the majority rule if few others registered it, and then unrelated later studies can match. The positive audit is the only check.
4. Splitting can leave fragments such as microporous. A fragment can match an unrelated later study. The positive audit is the only check.
5. Positive-rate selection: with the label fixed only after this rule, a pass at this gate says the links are usable. It does not say the label is complete.
6. The development read used the same data the gate draws from. Fresh trials reduce this but do not remove it, since the same companies recur.