Creating a Capability without Training / V103–V103.1
From verified cases alone, with no training run, EMMA created capabilities that matched or beat every earlier course on two of three needs once all 420 verified cases were used: 83.4% and 82.6% on 1,000 sealed cases each, in about 11 seconds on the processor. V103 missed its 3-point margin over trained creation at 120 cases; V103.1, which builds on the existing MicroModels’ answers, passed every gate.
SCOPE + LIMITATIONS
- One task family and three needs. The predictor is an outside component used as released; no EMMA training is involved.
- This public edition reviews saved evidence; experiments were not rerun for publication.
CONTROL COMPARISON
For the three needs of V99.1, give an existing predictor, used without any training, the need’s verified cases, and score once on 1,000 sealed cases per need from configurations never among those cases. V103.1 adds the existing MicroModels’ answers as inputs, so that creation corrects the system instead of starting from nothing.
V103 failed the margin gate at 120 cases and passed the near-best and cost gates. Accuracy rises by 7 to 10 points from 120 to 420 cases, so creation is limited by data, which can grow at no training cost. V103.1 passed all three gates, the course joined the system’s options, and it later met a need by creation (V108.2).
- M1 · 420 cases
- 83.4%; best earlier course 83.1%
- M2 · 420 cases
- 82.6%; best earlier course 81.0%
- M3 · 420 cases
- 81.0%; best earlier course 83.6%
- V103 · 120 cases against trained creation
- -0.8, +2.4 and +3.0 points; gate +3 each
- Time per need
- About 11 s on the processor
- V103.1
- All three gates passed
Question and method
For the three needs of V99.1, give an existing predictor, used without any training, the need’s verified cases, and score once on 1,000 sealed cases per need from configurations never among those cases. V103.1 adds the existing MicroModels’ answers as inputs, so that creation corrects the system instead of starting from nothing.
Recorded results
| Condition | Recorded result |
|---|---|
| M1 · 420 cases | 83.4%; best earlier course 83.1% |
| M2 · 420 cases | 82.6%; best earlier course 81.0% |
| M3 · 420 cases | 81.0%; best earlier course 83.6% |
| V103 · 120 cases against trained creation | -0.8, +2.4 and +3.0 points; gate +3 each |
| Time per need | About 11 s on the processor |
| V103.1 | All three gates passed |
Comparative standing and locked gates
V103 failed the margin gate at 120 cases and passed the near-best and cost gates. Accuracy rises by 7 to 10 points from 120 to 420 cases, so creation is limited by data, which can grow at no training cost. V103.1 passed all three gates, the course joined the system’s options, and it later met a need by creation (V108.2).
Interpretation and limitations
One task family and three needs. The predictor is an outside component used as released; no EMMA training is involved.
SOURCE PROVENANCE
Creating a Capability without Training / V103–V103.1
LABORATORY REPORT / 2026-10-25SOURCE CHECKSUM / SHA-256
3218257841225fe73549bf7f4045a6f8e89d4916ac763759c0f243c7faec1913Public 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.