IDEA-014 R4 v3 results, and R4 close-out
Run 2026-11-04 under the spec committed at cd3bf9e (deal-r4-v3-spec.md). Past data; nothing here is a skill claim. The ranker, deal-v0.1 and every earlier number are unchanged.
Short version
- R4 stopped at its map gate: 12 of 20 sampled molecules were correct, against a floor of 16.
- No v3 label was computed and no scorer was read.
r4v3_label.jsonandr4v3.jsondo not exist. R2 stays "cannot test". Whether past deals worked is still untested. - R4 is closed. The spec allows no v4 and does not allow a failed gate to be repaired after seeing it.
The whole R4 record
| attempt | what stopped it | result | floor |
|---|---|---|---|
| v1, exclusive names | linkage gate | 30 of 50 clean (60 percent) | 70 percent |
| v2, majority names | positive audit | 5 of 10 true (50 percent) | 70 percent |
| v3, molecule map | map gate | 12 of 20 correct (60 percent) | 80 percent |
Map and gate
Seven Sonnet session subagents built the map from the public web: 69 targets, 155 molecules, 14 targets with none (mostly device and diagnostics companies). It was committed at 5c80937 before the gate. Sample: random.Random(3), 20 molecules. A separate Sonnet reader checked them on the web. Files: docs/audits/idea014-r4v3-map.json, docs/audits/idea014-r4v3-map-gate.json.
| failure | rows | companies |
|---|---|---|
| an alias is not the molecule's name, or has no public support | 4 | Entasis (combination label), Oyster Point (unsupported salt), Akouos (a study name), Tesaro (a single catalog page) |
| the company did not hold the molecule at the deal, or public sources cannot show it | 4 | Endocyte (not shown), Somaxon (licence ended 2009), Five Prime (partner held development), GenVec (partner held development) |
What this means
- The rights failures matter most. Half the failures are programmes a partner ran or a licence that had ended. A later trial on one of those would say nothing about the deal.
- Not every failure would have changed a label. The Akouos alias, for example, may match nothing wrong. That was not measured, and it is not a reason to pass the gate: the floor was set on the map, before any of this was seen.
- Web agents cannot yet supply "who held which rights on the deal date" at this accuracy, in one pass. A usable asset map needs primary filings read per deal, or a human.
- The floor is one reader's judgement. It marked unsettled rows as false, as instructed, and is one Sonnet agent, with no second read. Had it passed, a second read would have been needed before relying on it.
What stays true
- R3 stands as a description: the rubric shows no measurable edge over counting trials at separating acquired from not-acquired listed companies (0.578 vs 0.583, difference -0.005, 95 percent CI -0.079 to 0.069). The label misses an estimated 44 percent of true acquisitions (9 to 72, n = 30), so this is descriptive only.
- The acquisition check stays an unvalidated evidence card, not a decision tool.
Limits
- Order: the map gate ran before any label, so no outcome informed the stop. The spec, map and read code were committed before the gate; the read code has never run.
- The executor knew the v2 candidates and aliases when writing the spec, as the spec states.
- Everything here is agent-checked. No human has read the map or the gate.