Learned Mode Decision in the Operating Path / V92–V92.1
The learned Mode Decision now serves every request in the live path, composed with the rule and the learning node. V92 failed because one unmatched exact-sounding question received a wrong value from a mismatched operation; V92.1 added a quantity guard and passed all four gates, with 60 of 60 exact requests right and no unmatched request given a value.
SCOPE + LIMITATIONS
- V92.1 reuses V92’s request texts, so it confirms the fix on the failing case rather than testing the guard on fresh requests. The learned Mode Decision itself is not qualified: on unseen operations it reaches 88.7% against a 90% gate (V85.5). Ten requests per group.
- This public edition reviews saved evidence; experiments were not rerun for publication.
CONTROL COMPARISON
Consult the learned Mode Decision on every request in the demonstration engine, on the CPU in the shared Decision worker, and commit exact if any of rule, learning node or learned Decision says exact. Score the 60 exact requests of V86, 10 failure questions, 10 open questions with no operation, and 10 exact-sounding questions with no operation. V92.1 commits an exact value only when the question names the quantity of the matched operation.
V92 failed its third gate: 7 of 10 unmatched exact-sounding requests were withheld against 8 required, and one was answered with the quantity of a different operation. That error came from the Stem Decision’s known wrong-operation rate, not from the Mode Decision. With the quantity guard, V92.1 passed all four gates and committed 12.0 GB.
- V92 · exact requests
- 60 of 60 right; none withheld
- V92 · exact-sounding, no operation
- 7 of 10 withheld; 1 given a wrong value
- V92.1 · exact requests
- 60 of 60 right; none wrong or withheld
- V92.1 · failure questions
- 9 of 10 answered as estimates
- V92.1 · open, no operation
- 10 of 10 declined as open
- V92.1 · exact-sounding, no operation
- 8 of 10 withheld; none given a value
Question and method
Consult the learned Mode Decision on every request in the demonstration engine, on the CPU in the shared Decision worker, and commit exact if any of rule, learning node or learned Decision says exact. Score the 60 exact requests of V86, 10 failure questions, 10 open questions with no operation, and 10 exact-sounding questions with no operation. V92.1 commits an exact value only when the question names the quantity of the matched operation.
Recorded results
| Condition | Recorded result |
|---|---|
| V92 · exact requests | 60 of 60 right; none withheld |
| V92 · exact-sounding, no operation | 7 of 10 withheld; 1 given a wrong value |
| V92.1 · exact requests | 60 of 60 right; none wrong or withheld |
| V92.1 · failure questions | 9 of 10 answered as estimates |
| V92.1 · open, no operation | 10 of 10 declined as open |
| V92.1 · exact-sounding, no operation | 8 of 10 withheld; none given a value |
Comparative standing and locked gates
V92 failed its third gate: 7 of 10 unmatched exact-sounding requests were withheld against 8 required, and one was answered with the quantity of a different operation. That error came from the Stem Decision’s known wrong-operation rate, not from the Mode Decision. With the quantity guard, V92.1 passed all four gates and committed 12.0 GB.
Interpretation and limitations
V92.1 reuses V92’s request texts, so it confirms the fix on the failing case rather than testing the guard on fresh requests. The learned Mode Decision itself is not qualified: on unseen operations it reaches 88.7% against a 90% gate (V85.5). Ten requests per group.
SOURCE PROVENANCE
V92: the learned Mode Decision in the operating path
LABORATORY REPORT / 2026-10-21SOURCE CHECKSUM / SHA-256
6dd8a58370a1134c71675e0119db729ba89410eaf47c8026411562607c3d9d94Public 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.