Self-Designed Research Cases / V83.1–V83.2
A capability designed its own training cases instead of choosing from an executed pool. Designing by uncertainty alone in V83.1 scored 73.1%, below the pool method and random cases. Weighting designs by how typical they are in V83.2 raised it to 79.1%, level with the pool method (78.6%), but only 2.3 points above random against a 5-point gate.
SCOPE + LIMITATIONS
- Own research remains qualified within a candidate pool (V83). Design by information density becomes the capability’s research method and matches the pool without needing one; a larger margin over random is later optimisation. One capability, one budget, fresh sealed sets of 1,000.
- This public edition reviews saved evidence; experiments were not rerun for publication.
CONTROL COMPARISON
Start loss_halves from its 50 shown examples and give each arm 200 more examples in four rounds of 50, relearning between rounds. The designed arm writes 2,000 candidates per round and has the 50 it is least sure about executed; V83.2 weights that choice by information density. The pool arm chooses from an already-executed pool; the random arm draws configurations. Score once on 1,000 fresh sealed configurations.
V83.1 failed both margin gates. V83.2 failed the margin over random (2.3 points against 5) and passed the pool comparison, execution of every designed case, and unchanged other answers. Uncertainty alone concentrated cases on the decision boundary, a bias the active-learning literature describes; density weighting is its standard remedy.
- V83.1 · designed by uncertainty
- 73.1%
- V83.1 · pool
- 80.8%
- V83.1 · random
- 79.4%
- V83.2 · designed, density-weighted
- 79.1%
- V83.2 · pool
- 78.6%
- V83.2 · random
- 76.8%
Question and method
Start loss_halves from its 50 shown examples and give each arm 200 more examples in four rounds of 50, relearning between rounds. The designed arm writes 2,000 candidates per round and has the 50 it is least sure about executed; V83.2 weights that choice by information density. The pool arm chooses from an already-executed pool; the random arm draws configurations. Score once on 1,000 fresh sealed configurations.
Recorded results
| Condition | Recorded result |
|---|---|
| V83.1 · designed by uncertainty | 73.1% |
| V83.1 · pool | 80.8% |
| V83.1 · random | 79.4% |
| V83.2 · designed, density-weighted | 79.1% |
| V83.2 · pool | 78.6% |
| V83.2 · random | 76.8% |
Comparative standing and locked gates
V83.1 failed both margin gates. V83.2 failed the margin over random (2.3 points against 5) and passed the pool comparison, execution of every designed case, and unchanged other answers. Uncertainty alone concentrated cases on the decision boundary, a bias the active-learning literature describes; density weighting is its standard remedy.
Interpretation and limitations
Own research remains qualified within a candidate pool (V83). Design by information density becomes the capability’s research method and matches the pool without needing one; a larger margin over random is later optimisation. One capability, one budget, fresh sealed sets of 1,000.
SOURCE PROVENANCE
V83.1: own research beyond a candidate pool (FR-D3 b)
LABORATORY REPORT / 2026-10-21SOURCE CHECKSUM / SHA-256
d29c37fc5f65c73df657a42a1391fc8290a1d4c080df5e8800bb90089b519aeePublic journal edition reviewed 2026-10-22. 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.