STOP — PP3 is an OFF correctness prototype, not a serving-speed candidate
| Brief item | Delivered | Result / remaining gate |
|---|---|---|
| 1 Fused producer/K2 | dyadic24 top32 primitive, keyed Philox, T0 RNG bypass, optional eager K2 hook, B1–B4/depth4–5 measurements | Two launches/position, not one. B1depth5 .155443ms; B4 .208589ms. Misses absolute .05ms goal. Synthetic conditioned-logit selector only; actual K2/emulated model/pair costs unqualified. |
| 2 Async q/verification | Negotiated extension, request seed/T mask, sparse q+CRC, both proxy modes/shared pools, FP32 full-vocab q, pinned standard verifier+failure mask overlays | 100K GPU toy-law and100K T0 comparisons PASS; exact PP1 law using actual GPU q PASS; CPU thread/process mixed-T/failure paths PASS. Engine grammar/state and real pair/GB10 qualification remain. |
| 3 Estimate/window | Hash-bound PP2 freeze, .182813ms known proxies, explicit unknown costs, short/32K/150K sensitivities,69min ABBA plan capped75min | Live rate remains null.150K acceptance unmeasured. No window opened or recommended for current prototype. |
| 4 Validation/handoff | CPU/GPU evidence, strict typed new runtime components, source/hash reconciliation, reproduction notes, cleanup/report | See campaign PP3.validation.log and PP3.report.md for final counts, commits and cleanup. |
Production remains unchanged. Neither switch is enabled by the commits. Ordinary stage_overlay.sh does not deploy the prepared verifier/runner files. The sampled startup guard requires both ABI markers and refuses incompatible policies.
The required /home/operator/miniconda3/bin/python3 on x86host isCPython3.13 with Torch2.12.1/cu130 and Triton3.7.1, but no exllamav3_ext_facc. Production's EXL3 binaries areCPython3.12. PP2 used a cached separateCPython3.12 interpreter; PP3 did not substitute it or install/rebuild dependencies. Therefore no claim of full production-K2 TP2 emulation or paired greedy identity is made here.
The eager hook still computes the existing greedy selector, recomputes Markov logits and copies full vocabulary shards GPU-to-GPU before drawing. The single-launch/candidate-exchange/graph work must be finished and measured before any promotion decision. Graph-replayed selector costs and SM86 sampler costs are not exposed engine timings; PP2's float64 nominal top32/.7 acceptance is not new dyadic24 replay evidence. COST.md keeps these distinctions.
Coordinator next action: review the prototype and choose a follow-on optimization and interpreter qualification lane. Close full K2/pair/GB10/state gates, fill all missing costs and independently review before considering the pre-registered LIVE-AB-PLAN.md. This handoff requests no service mutation.