Finding, Connecting or Creating Capabilities / V99–V99.1
For each need, EMMA measured whether to reuse a library component, link existing MicroModels, or create a new part, proposed one, and waited for a recorded decision before attaching it. V99’s proposals were not reliable and it failed. With a proposal rule that accounts for noise and prefers reuse, V99.1 chose the best course for all three fresh needs, gaining 10.3, 11.2 and 24.7 points on 1,000 sealed cases each, and passed every gate.
SCOPE + LIMITATIONS
- One task family and three needs per version. Decisions in V99 were made by the person in charge of the laboratory; V99.1 followed a standing direction recorded for each decision.
- This public edition reviews saved evidence; experiments were not rerun for publication.
CONTROL COMPARISON
For each need, fit every applicable course on executed cases, measure each on held-back cases (80 in V99, 300 in V99.1), propose one with intervals, record the decision, and only then attach. Score every course on 1,000 fresh sealed cases per need. V99.1 changes only the proposal rule.
V99 failed its fix and soundness gates and passed the core-unchanged and approval gates. V99.1 passed all four: each attached course fixes its need and is the best or tied for best on sealed cases, the core is unchanged, and nothing was attached without a recorded decision. Creating a new part was the weakest course every time, 7 to 15 points below the best.
- V99 · N1
- Attached and fixed (70.4% to 76.7%), but not the best course
- V99 · N2
- Rejected; nothing attached
- V99 · N3
- 64.7% against 64.0% for the majority answer; not fixed
- V99.1 · M1, reuse
- 72.8% to 83.1%
- V99.1 · M2, link
- 69.8% to 81.0%
- V99.1 · M3, reuse
- 58.9% to 83.6%
- Attach before a decision
- Refused in every case
Question and method
For each need, fit every applicable course on executed cases, measure each on held-back cases (80 in V99, 300 in V99.1), propose one with intervals, record the decision, and only then attach. Score every course on 1,000 fresh sealed cases per need. V99.1 changes only the proposal rule.
Recorded results
| Condition | Recorded result |
|---|---|
| V99 · N1 | Attached and fixed (70.4% to 76.7%), but not the best course |
| V99 · N2 | Rejected; nothing attached |
| V99 · N3 | 64.7% against 64.0% for the majority answer; not fixed |
| V99.1 · M1, reuse | 72.8% to 83.1% |
| V99.1 · M2, link | 69.8% to 81.0% |
| V99.1 · M3, reuse | 58.9% to 83.6% |
| Attach before a decision | Refused in every case |
Comparative standing and locked gates
V99 failed its fix and soundness gates and passed the core-unchanged and approval gates. V99.1 passed all four: each attached course fixes its need and is the best or tied for best on sealed cases, the core is unchanged, and nothing was attached without a recorded decision. Creating a new part was the weakest course every time, 7 to 15 points below the best.
Interpretation and limitations
One task family and three needs per version. Decisions in V99 were made by the person in charge of the laboratory; V99.1 followed a standing direction recorded for each decision.
SOURCE PROVENANCE
Finding, Connecting or Creating Capabilities / V99–V99.1
LABORATORY REPORT / 2026-10-24SOURCE CHECKSUM / SHA-256
a528a25e5799e7105b6925c803343e36fa03d3e0ffe43060ec19dd7d838ae45bPublic journal edition reviewed 2026-10-28. Source documents and saved evidence were inspected; experiments were not rerun for this edition. Proprietary implementation code, model binaries, private infrastructure, and detailed machine records are not published here. Journal identifiers are editorial references. Catalog inclusion does not imply qualification or runtime promotion.