Factorized Operational Learning / V63
A shared 354-parameter predicate and 42-parameter combiner scored 512/512 on sealed interval-cancellation tasks. A better-budgeted flat control reached 487/512. About 97% of sealed slot inputs had appeared before; this tests new composition geometry, not broad predicate novelty.
CONTROL COMPARISON
512 sealed interval-cancellation tasks; approximately 97% of slot inputs had appeared before.
This tests new pair geometry rather than mostly novel component facts. The candidate remains unpromoted; an exact rule remains a competent control.
- Factorized candidate
- 512/512
- Better-budgeted flat control
- 487/512
Task and protocol
Learn whether two typed interval effects cancel at a queried position. Query-relative coordinates expose the relevant geometry to a shared 354-parameter predicate and 42-parameter combiner. Labels are independently checked against actual program executions. Training, development, calibration, older-task checks, sealed tasks, and retention are separately defined.
Results
The factorized candidate scored 128/128 on development and 512/512 on sealed tasks. A better-budgeted flat comparison reached 487/512, approximately 95.1%, showing that V62’s training budget contributed to failure. The passing candidate remains an unpromoted library candidate.
Interpretation and limitations
Approximately 97% of sealed per-slot inputs had appeared before. Evaluation therefore tests unseen pair geometry rather than mostly unseen component facts. An exact rule remains a competent control. These results do not establish unrestricted operational learning, production repair, or superiority over deterministic solutions.
SOURCE PROVENANCE
EMMA V63: per-effect predicates and owned composition
LABORATORY REPORT / 2026-10-07SOURCE CHECKSUM / SHA-256
d5f050ca88033b085f8cdc2cbe35b5cee2f33ce0d9314d84d3c5092ecf5c55a9Public journal edition reviewed 2026-10-08. 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.