Output Mode End to End / V86
Sixty exact questions went through the full operating path, half in a form only the learned reader can read. All 60 were committed and right, none was wrong and none was withheld, passing both gates.
SCOPE + LIMITATIONS
- Clean exact questions on four operations; requests with a soft mode or no matching operation are not tested here. Sixty requests cannot show a wrong-value rate below about 5%. One kind of text disturbance and development data.
- This public edition reviews saved evidence; experiments were not rerun for publication.
CONTROL COMPARISON
Use the live demonstration engine as it serves visitors: the qualified Stem Decision, the Mode Decision by the operation table with the learning node, the staged Context path and the independent checker on every exact value. Send 60 fresh configurations, each followed by one exact question about parameter count, weight bytes, forward multiplications or hidden activations; half are reordered so that only the learned reader can read them.
Both gates passed. Answers read by the learned reader state that the values are verified for the fields as read and that the reading itself is not verified.
- Parser-readable · 30
- 100% committed and right
- Reordered, learned reader · 30
- 100% committed and right
- Wrong committed values
- 0; gate none
- Withheld
- 0%; gate at most 25%
- Mean time per request
- 5.9 s
Question and method
Use the live demonstration engine as it serves visitors: the qualified Stem Decision, the Mode Decision by the operation table with the learning node, the staged Context path and the independent checker on every exact value. Send 60 fresh configurations, each followed by one exact question about parameter count, weight bytes, forward multiplications or hidden activations; half are reordered so that only the learned reader can read them.
Recorded results
| Condition | Recorded result |
|---|---|
| Parser-readable · 30 | 100% committed and right |
| Reordered, learned reader · 30 | 100% committed and right |
| Wrong committed values | 0; gate none |
| Withheld | 0%; gate at most 25% |
| Mean time per request | 5.9 s |
Comparative standing and locked gates
Both gates passed. Answers read by the learned reader state that the values are verified for the fields as read and that the reading itself is not verified.
Interpretation and limitations
Clean exact questions on four operations; requests with a soft mode or no matching operation are not tested here. Sixty requests cannot show a wrong-value rate below about 5%. One kind of text disturbance and development data.
SOURCE PROVENANCE
V86: output-mode choice measured end to end through the operating path (FR-C)
LABORATORY REPORT / 2026-10-20SOURCE CHECKSUM / SHA-256
bf8067424937966d707740a3fc04e0b11a7ee9555bb25625138c715faa5d443cPublic 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.