Self-Found Needs in a Mixed Stream / V100.1–V100.4
Given only a mixed stream of 3,000 verified outcomes, EMMA found its own needs, met each by reusing a library component, and improved without forgetting: +2.8 points (79.6% to 82.4%) in V100.3 and +2.7 (78.9% to 81.6%) in V100.4 on fresh sealed sets of 3,000, with no regime below the core. V100.1 and V100.2 failed by finding too few needs.
SCOPE + LIMITATIONS
- One task family and the estimate kind; needs are found only where existing candidates can help. Approval followed a standing direction recorded for every attachment.
- This public edition reviews saved evidence; experiments were not rerun for publication.
CONTROL COMPARISON
Give the system 3,000 verified outcomes from six regimes, shuffled, with no needs supplied. It must find regions where it can improve, propose a course, attach what is approved, and leave everything outside unchanged. Each version uses a fresh stream and a fresh sealed set of 3,000. V100.1 to V100.3 each change only how needs are found; V100.4 adds creation without training as a course.
V100.3 and V100.4 passed all five gates. What fixed the finder was looking for where something measurably helps rather than where errors are frequent. In V100.4 every self-found need could be met by reuse, so meeting a need that only creation can meet was not established here.
- V100.1
- 1 need found against at least 2; +2.6 points; failed
- V100.2
- No need found; failed
- V100.3
- 2 needs found; 79.6% to 82.4%
- V100.3 · inside each need
- 59.5% to 73.5%; 65.7% to 80.5%
- V100.4
- 2 needs found; 78.9% to 81.6%
- Outside each attached part
- Bit-identical in every version
Question and method
Give the system 3,000 verified outcomes from six regimes, shuffled, with no needs supplied. It must find regions where it can improve, propose a course, attach what is approved, and leave everything outside unchanged. Each version uses a fresh stream and a fresh sealed set of 3,000. V100.1 to V100.3 each change only how needs are found; V100.4 adds creation without training as a course.
Recorded results
| Condition | Recorded result |
|---|---|
| V100.1 | 1 need found against at least 2; +2.6 points; failed |
| V100.2 | No need found; failed |
| V100.3 | 2 needs found; 79.6% to 82.4% |
| V100.3 · inside each need | 59.5% to 73.5%; 65.7% to 80.5% |
| V100.4 | 2 needs found; 78.9% to 81.6% |
| Outside each attached part | Bit-identical in every version |
Comparative standing and locked gates
V100.3 and V100.4 passed all five gates. What fixed the finder was looking for where something measurably helps rather than where errors are frequent. In V100.4 every self-found need could be met by reuse, so meeting a need that only creation can meet was not established here.
Interpretation and limitations
One task family and the estimate kind; needs are found only where existing candidates can help. Approval followed a standing direction recorded for every attachment.
SOURCE PROVENANCE
Self-Found Needs in a Mixed Stream / V100.1–V100.4
LABORATORY REPORT / 2026-10-25SOURCE CHECKSUM / SHA-256
019a5c13613e3c5c4fdf86df783d0776e42b3c15c2f686afb59447127a97563bPublic 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.