A Stream of Needs Met without Forgetting / V100
Five needs of three kinds arrived in order. EMMA met three of them, with a reused component, a remembered fact and a linked composition, and correctly declined two where no course would help. Average sealed accuracy over the five rose from 77.6% to 88.1%, and every earlier answer stayed bit-identical after the whole stream. All five gates passed.
SCOPE + LIMITATIONS
- The needs were supplied, not found; finding them is tested in V100.1 to V100.4. One task family; 5,680 executed cases in all.
- This public edition reviews saved evidence; experiments were not rerun for publication.
CONTROL COMPARISON
Five needs arrive in a fixed order, each of a kind named by its signal: an estimate, a new question, or a repeated fact. For each, the system measures its possible courses on 300 held-back cases, proposes one or nothing, attaches what is approved, and is scored on 1,000 sealed cases.
All five gates passed: each decision sound, no forgetting, each part placed in the kind its signal named, 3 parts for 5 needs, and nothing attached without a recorded decision.
- S1 · deep runs fail
- 71.9% to 78.3%; reuse
- S2 · runs diverge
- Nothing proposed; no course beat the current 93.3%
- S3 · the same run asked again
- 71.9% to 100.0%; memory
- S4 · large-init runs fail
- Nothing proposed; no course beat the current 88.2% by the margin
- S5 · runs stall
- 62.8% to 80.7%; linked composition
- Average over five needs
- 77.6% to 88.1%
Question and method
Five needs arrive in a fixed order, each of a kind named by its signal: an estimate, a new question, or a repeated fact. For each, the system measures its possible courses on 300 held-back cases, proposes one or nothing, attaches what is approved, and is scored on 1,000 sealed cases.
Recorded results
| Condition | Recorded result |
|---|---|
| S1 · deep runs fail | 71.9% to 78.3%; reuse |
| S2 · runs diverge | Nothing proposed; no course beat the current 93.3% |
| S3 · the same run asked again | 71.9% to 100.0%; memory |
| S4 · large-init runs fail | Nothing proposed; no course beat the current 88.2% by the margin |
| S5 · runs stall | 62.8% to 80.7%; linked composition |
| Average over five needs | 77.6% to 88.1% |
Comparative standing and locked gates
All five gates passed: each decision sound, no forgetting, each part placed in the kind its signal named, 3 parts for 5 needs, and nothing attached without a recorded decision.
Interpretation and limitations
The needs were supplied, not found; finding them is tested in V100.1 to V100.4. One task family; 5,680 executed cases in all.
SOURCE PROVENANCE
A Stream of Needs Met without Forgetting / V100
LABORATORY REPORT / 2026-10-25SOURCE CHECKSUM / SHA-256
9f506b4ad375b7385fe8293b7e5d672e380ab0b31c66b78690d3b594726d5985Public 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.