PASS — whole K2 timings and N4 greedy identity qualified
The sampled path replaces the selector in the complete production-weight K2 worker. On the assigned solo GPU, its measured incremental median is 0.288–0.791 ms per round across B1–B4 and output depths 4/5. All required greedy identities pass, including the older runtime’s archived N4 proposals and confidence. Production is unchanged. Physical-pair overhead and live target performance remain unmeasured.
MEASURED whole-worker cost
Both logical TP2 ranks execute on one SM86 card at 225 W. The timing fixture contains four real N4 request states, with fixed query block 5 and output depth 4/5. The table reports medians of five interleaved runs, each with 20 warmups and 100 individually event-timed rounds. Each p95 is the maximum per-run p95; neither that statistic nor the median difference is a live-pair upper bound.
| Batch | Output depth | OFF median ms | ON median ms | Increment ms | OFF p95 ms | ON p95 ms |
|---|---|---|---|---|---|---|
| 1 | 4 | 3.666944 | 3.954976 | 0.288032 | 3.777536 | 4.332960 |
| 1 | 5 | 3.686400 | 3.999392 | 0.312992 | 3.767296 | 4.157408 |
| 2 | 4 | 4.331520 | 4.822768 | 0.491248 | 4.396032 | 4.921664 |
| 2 | 5 | 4.350976 | 4.892960 | 0.541984 | 4.452352 | 4.949344 |
| 3 | 4 | 4.888576 | 5.417216 | 0.528640 | 4.979712 | 5.526976 |
| 3 | 5 | 4.898816 | 5.484656 | 0.585840 | 4.968448 | 5.576160 |
| 4 | 4 | 5.779456 | 6.264480 | 0.485024 | 5.894144 | 6.401984 |
| 4 | 5 | 6.790144 | 7.581072 | 0.790928 | 6.918144 | 8.055680 |
These measurements include the full K2 body, head, Markov conditioning, proposal selection, graph replay and request-metadata handling. They exclude physical NVLink exchange, proxy transport, target q expansion/correction and consumer wait. Both ranks run sequentially on the solo; this is not production-pair latency. The new peer exchange still needs the coordinator window.
MEASURED identity and runtime qualification
- All 120 identity cells pass: 48 all-T0 and 72 mixed-T0/T1 cells, totaling 228 T0 request comparisons across 24 unique N4/NB1 states. Mixed masks alternate both ways, covering every packed row offset. IDs and confidence are exact; meaningful q payload on T0 rows is zero.
- All 48 archived B1 proposal digests and all 48 confidence digests match the NB1 K2 receipt under the older CPython 3.12/Torch 2.10 runtime. This detects no arithmetic drift on those 24 states at either output depth. It does not prove identity for every possible state or the untested physical pair.
- Production checkpoint and Markov identities are verified against their source/extracted-payload receipts before the model loads.
- Peak host RSS was 6.95 GiB; peak Torch GPU allocation was 9.47 GiB. The full corrected run completed in 172.65 seconds. The card remained at 225 W; observed active clocks varied within the cap.
- The required CPython 3.13/Torch 2.12 CUDA 13 environment successfully imports rebuilt stock EXL3, FP32-accumulation EXL3 and P2P together. P2P has only CPU build/import qualification; no production-pair GPU was used.
full-worker.json contains the complete aggregate receipt, per-run medians, clock readings and source/binary hashes. runtime-build.json and RUNTIME.md record the dependency qualification and reproducible build instructions. No token IDs, confidence values, capture tensors, prompts, weights or generated text appear in these receipts. The archived reference contains digests only.
Harness correction and closeout
An initial loader attempt set max_rows=128, exceeding the production expert
scratch bound of 256 slots at top_k=3. It exited on the explicit guard. The
corrected max_rows=32 covers B4 × block 5; insertion buckets remain independently
captured through 128 rows. No production assertion or test was weakened.
The final source, runner and archived-reference hashes reconcile with the receipt. The solo returned to 1 MiB at 225 W, its lock is free, and the private runtime tmux server is absent. Build/failed-attempt logs and the full public object-reuse certificate are preserved as compressed campaign artifacts. All runtime scratch remains scheduled for removal by the root lane.