Live Demonstration Serving / V88–V88.1
The demonstration now serves a line of visitors one conversation at a time, with session limits and wait estimates. Sharing one Decision worker in V88.1 removed 5.4 GB of commit and halved the working set with identical answers, but neither run met its memory budget, so both failed that gate.
SCOPE + LIMITATIONS
- The budgets were set from a single-worker figure the real engine does not meet. Most of the remaining commit is inferred to be graphics-driver accounting rather than data the process holds. The engine needs about 13.5 GB of free commit, and the hardware preflight now refuses to start it with less. Simulated visitors on one machine, one run each.
- This public edition reviews saved evidence; experiments were not rerun for publication.
CONTROL COMPARISON
Hold a line of up to 40 visitors in front of the engine; limit each session to 30 turns or 10 minutes and release it after 2 minutes without a message; serve one request at a time. Five simulated visitors join at once and each sends a configuration, an exact question and a failure question; one goes quiet. Memory is sampled every two seconds. V88 uses one worker per Decision weight set; V88.1 uses one shared worker.
Both runs failed the memory gate: V88 against 6,300 MB and V88.1 against 11,360 MB. The ordering, release, correctness and identity gates passed. The shared worker is kept as the default.
- Peak commit · V88 / V88.1
- 18,932 MB / 13,532 MB
- Peak working set · V88 / V88.1
- 12,558 MB / 6,421 MB
- Mean seconds per turn · V88 / V88.1
- 3.1 / 2.4
- Visitors served in order; waiting visitor refused
- Yes in both
- Exact answers right
- 4 of 4 in both
- Shared-worker answers identical, 40 probes per weight set
- Yes
Question and method
Hold a line of up to 40 visitors in front of the engine; limit each session to 30 turns or 10 minutes and release it after 2 minutes without a message; serve one request at a time. Five simulated visitors join at once and each sends a configuration, an exact question and a failure question; one goes quiet. Memory is sampled every two seconds. V88 uses one worker per Decision weight set; V88.1 uses one shared worker.
Recorded results
| Condition | Recorded result |
|---|---|
| Peak commit · V88 / V88.1 | 18,932 MB / 13,532 MB |
| Peak working set · V88 / V88.1 | 12,558 MB / 6,421 MB |
| Mean seconds per turn · V88 / V88.1 | 3.1 / 2.4 |
| Visitors served in order; waiting visitor refused | Yes in both |
| Exact answers right | 4 of 4 in both |
| Shared-worker answers identical, 40 probes per weight set | Yes |
Comparative standing and locked gates
Both runs failed the memory gate: V88 against 6,300 MB and V88.1 against 11,360 MB. The ordering, release, correctness and identity gates passed. The shared worker is kept as the default.
Interpretation and limitations
The budgets were set from a single-worker figure the real engine does not meet. Most of the remaining commit is inferred to be graphics-driver accounting rather than data the process holds. The engine needs about 13.5 GB of free commit, and the hardware preflight now refuses to start it with less. Simulated visitors on one machine, one run each.
SOURCE PROVENANCE
V88 and V88.1: the real-engine demonstration on this PC, with a line for several visitors
LABORATORY REPORT / 2026-10-20SOURCE CHECKSUM / SHA-256
517f9446d2a37ca296ab9082d43aa2f0431fa8db040e68e395b75525f81c7ec0Public 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.