IDEA-014 R4 v3 specification

Written 2026-11-04, before any molecule map exists, any v3 label count or any score read under an asset label. Owner approval in chat, 2026-11-04: "Wrap up R4 whatever and get it done and fixed so we can publish that to the site". That is the new owner approval v2's stopping rule required. No spend: session subagents only, public web only.

Known when this was written, so this is not blind:

• v2's positive audit failed 5 of 10 because names are not molecules (deal-r4-v2-results.md). The nine v2 candidates and their aliases are known to the executor.
• R2 scorer AUCs under the old phase label are known (deal-v0.1-results.md), including the blind LLM's 0.68 and 0.64.
• No scorer has been read under any R4 label.

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. No result here changes the ranker or deal-v0.1.

What changes from v2

Population, 36-month horizon, censoring, gate splits, baselines and decision rule stay as in deal-r4-v2-spec.md. Asset names are no longer inferred from the registry by a majority rule.

1. Molecule map. For each of the 69 uncensored targets, a Sonnet session subagent that has seen no rule, label or score searches the public web and lists the investigational molecules the target held at the deal: canonical name, every public alias (code names under any owner, INN, brand) and a source URL. Input is the company name, deal date and the intervention names on its own pre-deal trials. Later names are required, because following them is the point. Aliases are label inputs only and never a feature.
2. A molecule needs a pre-deal trial. At least one cached study posted before the deal must name one of its aliases. This drops programmes the web knows about only because they later worked, which would bias the label towards success.
3. Pre-deal maximum phase is read across all aliases and all sponsors. Dates go through diaspore.backtest.gate.
4. Primary label: a study by any sponsor, posted in (deal, deal + 36 months], naming any alias at a phase strictly above that molecule's pre-deal maximum and at phase 3 or below.
5. Phase 4 leaves the primary label. v2 showed it can be post-marketing work (Alder). This can only remove positives. The count with phase 4 is reported as a secondary, non-binding row.

Gates, in order. Any failure stops the line and the result is reported as it stands

1. Map gate. 20 molecules by random.Random(3).sample(sorted(map rows), 20), read by a separate Sonnet agent with public web access, which does not see the label. Correct means every alias is the same molecule, and the target held it at the deal. At least 16 of 20.
2. Positive audit. Before any score is read, an agent that did not build the map checks every candidate against its pre-deal and later study rows. True means the later study tests that molecule as an investigational product, not as a comparator or background drug, at a higher phase than any pre-deal study. False positives go on an exception list that only removes positives. At least 70 percent true. Labelled "agent-checked".
3. Mode check. Minimum detectable AUC from audited counts by min_detectable_auc. Above 0.85: "cannot test", and no scorer AUC is read. With 69 rows that means at least 6 audited positives.
4. Auditor. backtest-auditor runs after the audit and before any scorer AUC is read.

Always reported, non-binding

• Targets with no mapped molecule, and with no molecule passing change 2. They count as label 0 because the population is unchanged. The AUC table excluding them is also reported.
• A miss check: 10 targets by random.Random(4).sample(sorted(label-0 targets with a molecule), 10), read by a web-enabled agent asking whether any mapped molecule reached a higher phase within 36 months. It estimates missed positives roughly and never changes a label.

Read, if every gate passes

• Decision rule, unchanged: passes only if the 95 percent lower bound of AUC(rubric) minus AUC(trial count) is above 0. One row per company; resampling companies, 10,000 draws, seed 0. Otherwise the result reads that the rubric adds nothing over trial count. Any rubric 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 and publishing

This is the last attempt at this question on this population. There is no v4. The result goes on the site whichever way it comes out, including "cannot test" or a stop at a gate. A failed gate is not repaired after seeing it.

Known limits, stated before the read

1. Agents wrote the map. Better-known molecules have more findable aliases, so a map can favour programmes that later succeeded. Change 2 reduces this and does not remove it.
2. The registry names cache covers industry-lead studies only. An advance run only by an academic sponsor is missed.
3. Registry names are 2026 state, and advancing can reflect the acquirer's budget. It is neither value nor a return.
4. Holding "at the deal" is judged from the web today. Licensed-out or partnered rights can be wrong; the map gate samples this but does not certify it.
5. 69 companies, one horizon. Even a pass here would be one past-data read.