Demonstration Rehearsal with Learning in Operation / V91
On the real engine the demonstration met every applicable gate of its design for the first time: no planted attack learned or remembered, returning visitors answered from memory without leaks, and one unit promoted from outcomes the system verified itself, with the probe rising from 86.8% to 87.7%.
SCOPE + LIMITATIONS
- Scripted visitors written by the laboratory and one run. The 0.8-point probe gain is positive but not statistically resolved, and the learned unit came from the self-chosen scope mechanism before it qualified.
- This public edition reviews saved evidence; experiments were not rerun for publication.
CONTROL COMPARISON
Run the V89 script unchanged, 18 visitors, 69 turns and 12 planted attacks, with the research worker enabled, the research loop running between turns, and every readable visitor configuration queued for execution. After the visitors, keep the engine open with no one in line.
Selective intake, memory, learning in operation and no regression all passed; the Source Model phase-out gate does not apply. Against V89 on the same script the intake and memory results are unchanged and learning in operation is now shown.
- Outcomes executed between turns
- 160
- Selective intake
- 0 of 12 planted messages learned or remembered
- Memory
- Both returning visitors given the verified value; no name leaked
- Unit promoted
- At 1,880 outcomes; validation 81.2% against 74.4% frozen
- Probe
- 87.7% against 86.8% frozen on 600 configurations; 280 answers outside scope bit-identical
Question and method
Run the V89 script unchanged, 18 visitors, 69 turns and 12 planted attacks, with the research worker enabled, the research loop running between turns, and every readable visitor configuration queued for execution. After the visitors, keep the engine open with no one in line.
Recorded results
| Condition | Recorded result |
|---|---|
| Outcomes executed between turns | 160 |
| Selective intake | 0 of 12 planted messages learned or remembered |
| Memory | Both returning visitors given the verified value; no name leaked |
| Unit promoted | At 1,880 outcomes; validation 81.2% against 74.4% frozen |
| Probe | 87.7% against 86.8% frozen on 600 configurations; 280 answers outside scope bit-identical |
Comparative standing and locked gates
Selective intake, memory, learning in operation and no regression all passed; the Source Model phase-out gate does not apply. Against V89 on the same script the intake and memory results are unchanged and learning in operation is now shown.
Interpretation and limitations
Scripted visitors written by the laboratory and one run. The 0.8-point probe gain is positive but not statistically resolved, and the learned unit came from the self-chosen scope mechanism before it qualified.
SOURCE PROVENANCE
V91: rehearsal of the live demonstration on the real engine, with learning in operation
LABORATORY REPORT / 2026-10-20SOURCE CHECKSUM / SHA-256
43516a41357aad3a7b231a8dc47d683a76e5afb84badfe76eb1b20fb3ff95652Public 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.