Whole Live Demonstration, Every Part Switched On / V98–V98.1
The live demonstration ran on the real engine with every built part of the operating path switched on and, in V98.1, met every applicable gate of its design: none of 12 planted attacks was learned or remembered, returning visitors were answered from memory without leaking a name, every exact answer was right, and a unit learned from self-verified outcomes was promoted during operation. V98 failed one gate first and remains recorded.
SCOPE + LIMITATIONS
- The script and probe were seen in V98, so V98.1 confirms the fix rather than serving as a fresh test. Visitors are scripted by the laboratory, the attacks are not an outside red team, and the run is a single execution on one desktop computer. The probe gain of 0.6 points is positive but not statistically resolved. The AdaptiveWeight unit for the reader stays out of the path until a format router exists.
- This public edition reviews saved evidence; experiments were not rerun for publication.
CONTROL COMPARISON
Run the demonstration engine behind its visitor line with intake, the learned readers through the block store, the qualified Stem Decision, the learned Mode Decision with its quantity guard, the three stems through the Context lanes with the focus stage, the independent checker, memory layers, and the research worker learning between turns. The script has 21 visitors and 97 turns: the 18 visitors and 12 planted attacks of the earlier rehearsal, a proposal question from every honest visitor, and three visitors asking 15 requests outside the operation set. The runbook was locked before each run. V98.1 changes only how the research worker is created.
V98 passed every applicable gate except learning in operation. The cause was a race in which the background research thread and the first visitor each created a research worker; the engine now creates exactly one under a lock, and a four-thread regression test covers it. V98.1 then passed all seven applicable gates. The Source Model phase-out gate does not apply, because the engine makes no Source Model calls while serving. No gate was changed.
- V98 · selective intake
- 0 of 12 planted attacks learned or remembered
- V98 · learning in operation
- Failed: the research loop executed no case
- V98 · cause
- Two research workers created at the same moment; visitors fed one, the loop watched the other
- V98.1 · selective intake
- 0 of 12 learned, 0 remembered; no name leaked
- V98.1 · learning in operation
- 2,440 self-verified outcomes; one unit promoted at 2,420
- V98.1 · probe
- 86.2% to 86.8%; all 371 answers outside the unit’s scope bit-identical
- V98.1 · exact answers and proposals
- Every exact answer right; 10 of 10 proposal questions answered with a proposal
- V98.1 · peak commit
- 13,242 MB against a 13.6 GB limit; 3.24 s mean per turn
Question and method
Run the demonstration engine behind its visitor line with intake, the learned readers through the block store, the qualified Stem Decision, the learned Mode Decision with its quantity guard, the three stems through the Context lanes with the focus stage, the independent checker, memory layers, and the research worker learning between turns. The script has 21 visitors and 97 turns: the 18 visitors and 12 planted attacks of the earlier rehearsal, a proposal question from every honest visitor, and three visitors asking 15 requests outside the operation set. The runbook was locked before each run. V98.1 changes only how the research worker is created.
Recorded results
| Condition | Recorded result |
|---|---|
| V98 · selective intake | 0 of 12 planted attacks learned or remembered |
| V98 · learning in operation | Failed: the research loop executed no case |
| V98 · cause | Two research workers created at the same moment; visitors fed one, the loop watched the other |
| V98.1 · selective intake | 0 of 12 learned, 0 remembered; no name leaked |
| V98.1 · learning in operation | 2,440 self-verified outcomes; one unit promoted at 2,420 |
| V98.1 · probe | 86.2% to 86.8%; all 371 answers outside the unit’s scope bit-identical |
| V98.1 · exact answers and proposals | Every exact answer right; 10 of 10 proposal questions answered with a proposal |
| V98.1 · peak commit | 13,242 MB against a 13.6 GB limit; 3.24 s mean per turn |
Comparative standing and locked gates
V98 passed every applicable gate except learning in operation. The cause was a race in which the background research thread and the first visitor each created a research worker; the engine now creates exactly one under a lock, and a four-thread regression test covers it. V98.1 then passed all seven applicable gates. The Source Model phase-out gate does not apply, because the engine makes no Source Model calls while serving. No gate was changed.
Interpretation and limitations
The script and probe were seen in V98, so V98.1 confirms the fix rather than serving as a fresh test. Visitors are scripted by the laboratory, the attacks are not an outside red team, and the run is a single execution on one desktop computer. The probe gain of 0.6 points is positive but not statistically resolved. The AdaptiveWeight unit for the reader stays out of the path until a format router exists.
SOURCE PROVENANCE
V98: the whole demonstration rehearsed with every part switched on
LABORATORY REPORT / 2026-10-21SOURCE CHECKSUM / SHA-256
70d47b0c9dc97f0490ddf53f76b6f5d78425f7b8f3be9a8dffd59f4c00ed7763Public 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.