# PP3 pre-registered coordinator A/B window — NOT RUN

Decision now: STOP. This is an execution plan after missing qualification gates
close, not authorization to deploy. ESTIMATED total69minutes; hard ceiling75.
No PP3 lane access to Spark/sparkworker, serving endpoints, production pair/units or
any window directory is authorized or exercised. Coordinator owns every live
operation and the exact restore receipt. No automatic promotion.

## Admission gates

Freeze source commits/hashes and obtain independent review. Qualify the entire
production K2 TP2 sampled path under its actual interpreter/Torch/Triton and
weights; verify B1–B4 at depths4/5, T0 identity, q counts and conditioning, slot
reuse, request permutation, cancel/reset, reconnect, timeout and late reply.
Measure actual exposed added cost including eager/graph differences, Markov,
candidate exchange or full-vocab peer staging, CRC/validation and target work.
The present two-launch selector is .155–.209ms/depth5; it misses the absolute
.05ms objective and is not a speed-qualified implementation.

Reconcile both pinned image sources with live source/image hashes. Prepare
both overlays offline with prepare_verifier, review their diff/receipt, then
use a separate coordinator-approved experimental read-only mount profile.
Ordinary production staging intentionally does not deploy these files. Prove
both ABI markers and engine failure invalidation; trial missing/corrupt q on
both ranks must become target-only. Validate actual grammar/penalty/stop/EOS,
tool JSON and reasoning parser state on the exact engine. Solo histograms and
pure callback oracles do not establish those stateful properties.

Require a conservative cost estimate with positive margin at the family lower
bounds and32K proxy; fill missing terms rather than pricing them at zero.
Measure150K acceptance/cost separately in the bounded window. If any admission
gate fails or evidence is incomplete, do not open the window. The current lane
therefore hands over a NO-GO, plus this reproducible plan for a later revision.

## Frozen arms and schedule

A: current production K2 greedy implementation, worker/engine sampled switches
OFF, standard rejection, fixed depth5. B: same weights, backbone, Markov/cache,
Engram backend, target processors and source build; both explicit sampled
switches ON, top32/tau.7/dyadic24-floor-first-v1, same fixed depth5. Both disable
confidence/adaptive admission, prompt lookup, dynamic depth and depth-file
reload. A remains the existing implementation under the same fixed-depth
experimental setting; do not compare adaptive A against fixed5 B. If depth5 is
unsupported, STOP and register a new fixed4 experiment for both arms.

Keep target T1/top_p1, high reasoning effort, priority10, output budget and
context/KV profile identical. B1 is primary, short/32K/150K occupied contexts
with reasoning/output reserve. Two frozen public task families each provide
one short, one32K and one150K request (six pairs total). Hash input sources,
request/seed/target policy and ordered family assignment before any A/B data;
no use of the development/PP2 fit examples and no prompt/output text in git.
Stochastic texts need not match. T0 uses paired identical prompts/seeds; compare
emitted tokens, not undefined padding past num_sampled.

Interleave **A1/B1/B2/A2** with a separate engine session for each arm block.
Each block runs all six frozen requests in the same predeclared rotating order
with matched seeds/budgets. This uses real restarts with recorded cold/warm
TTFT; changing environment variables in a running engine is not an arm switch.
Do not claim per-round randomized interleaving. Begin every block only after
its worker/engine identity, real warmup rounds and effective target parameters
are attested; no fabricated switch or launcher flags.

ESTIMATED budget69minutes:

| Stage | Minutes | Stop condition |
|---|---:|---|
| Drain/ownership, exact snapshots and rollback preflight |7|No idle/restore ownership receipt: abort|
|A1/B1/B2/A2, each start+warm4, prefix priming4, six bounded decode runs3|44|Any block over11min: truncate the experiment, preserve raw partials|
|Paired greedy and failure/structured smoke within qualified test processes|3|Any semantic mismatch: reject B|
|Restore exact baseline and verify usable service|12|Use existing coordinator rollback escalation|
|Receipt/statistics/close|3|Incomplete data: no promotion|

Decode runs are at most30s each after prefix priming; delivered tokens and
complete target rounds are counted without requiring a fixed token count to
finish after a deadline. Record any truncated full-task observation as censored.
At elapsed50min stop admitting experiment requests and begin restore if any
stage has overrun; preserve12min for restore. Hard75min includes restoration,
not an excuse to leave B running. Longer full-task/extra-family/B4 work requires
a separately registered window; do not expand this window opportunistically.

## Endpoints, measurements and frozen decision

Only coordinator-reviewed experimental endpoints/processes are used, after
production drain/ownership and HOLD checks from COMMON. Exact endpoint and
launcher commands come from the coordinator's live registry; this plan invents
none. Prefix priming is identical across arm blocks, with cold start/prefill
reported separately and exact occupied token counts verified. No new context
limit, GPU power/clock or concurrency setting is permitted.

Collect per-request/source-family/block and per-round IDs: actual delivered
tokens, accepted drafts/round, complete ms/round p50/p95/p99, target verifier and
q costs, TTFT, first visible answer, answer/reasoning counts, bytes, watchdogs,
queue/staging and both-rank synchronization. Include canceled/failed/timeout
requests in the disposition; never drop slow samples. Save unprofiled headline
rates separately from CUDA-event profiling. Check memory/serving health and
request backlog throughout; abort on unexplained failures/pressure.

Pre-registered rule: **REJECT** for any q/target-law, greedy-identity, grammar,
stop, lifecycle, corruption or recovery failure, any lost request, or target
policy/weight mismatch. **NO PROMOTION / INCONCLUSIVE** if incomplete, measured
throughput gain<5%, source-cluster paired90% bootstrap lower bound<=0, or>5%
TTFT/p95-round regression in any context/family. Bootstrap2000 source-cluster
resamples with seed4103; report both task families and the small-family limit,
never treat thousands of rounds as independent sources. Report accepted/round
as a diagnostic; it cannot override delivered throughput or TTFT. Candidate
only if every gate passes, point gain>=5%, lower bound>0 and per-context guards
pass. Even then, independent coordinator approval and promotion are separate.