Emitted-q acceptance requalification
PASS: emitted dyadic24 q preserves PP2's acceptance benefit. MEASURED expected retained tokens per nonterminal round rise from 3.091813 to 3.305515, a 6.912% increase before added time. Paired source-cluster 95% interval for the gain is 0.175370–0.254435 tokens/round. This is acceptance qualification, not an observed throughput gain or permission for a production experiment.
The complete replay receipt and paired summary bind the 1,710 frozen anchors, 24 requests/source clusters and both heldout task families to the exact production weights and consumed capture payloads. Actual fused support and q match PP3's reference producer on 8,550/8,550 rows. Own-greedy-path full-logit argmax parity passes on 8,550/8,550 positions. Current greedy source/context prefix aggregates equal frozen PP2; that aggregate check alone cannot prove per-anchor identity across the two runtime versions.
| View | Greedy retained/round | Sampled retained/round | Delta / paired 95% source interval |
|---|---|---|---|
| All 1,710 anchors / 24 sources | 3.091813 | 3.305515 | 0.213702 /0.175370–0.254435 |
| Chronology 802 / 12 | 3.133416 | 3.369527 | 0.236111 /0.163163–0.301578 |
| Motion outline 908 / 12 | 3.055066 | 3.248976 | 0.193909 /0.162188–0.233094 |
| 16–32K 545 / 12 | 3.223853 | 3.376376 | 0.152523 /0.103830–0.200527 |
Every view with an available PP2 interval remains inside that interval. Pooled change from PP2's float64 nominal q estimate is +0.000000613 retained tokens/round. The >=32K view still has only 66 anchors from one source and no interval. Maximum anchor context remains 47,084; no 150K result is claimed.
MEASURED on x86host's bc62df2a at 225 W, under its exclusive flock, eight reserved threads/nice 19: full scoring 50.260 s after loading, peak RSS 8.69 GiB and Torch GPU allocation 9.35 GiB. This duration includes diagnostics and is not serving latency. Required miniconda CPython 3.13.12, Torch 2.12.1/cu130, Triton 3.7.1 used the separately rebuilt stock/facc extensions recorded in the receipt. The solo was verified 225 W / 1 MiB with a free lock before handing it to the runtime sub-item. No serving request or physical-pair execution occurred.
The primary policy remains top32/tau 0.7, fixed depth 5, as frozen in PP2
e1befc9aaa081a4f6cccbb11423a747fcd392503. acceptance_frozen.py checks the
SHA256 of PP2's selection, heldout aggregates, summary, numeric scorer and
source-bootstrap implementation before using them. It can extract those
public-code/aggregate files with git show only into PP4 scratch. There is no
refit or new anchor selection.
acceptance_replay.py reloads the same production K2 weights and captured
heldout shards, verifying checkpoint tensor payloads, source/weight identity,
manifest, sidecars, token histories and every consumed shard. Both logical
TP2 shards run on x86host's assigned solo GPU. All 1,710 scheduled anchors from
the 24 chronology/motion_outline requests are required. Their grouping,
source connectivity, context bins and counts must match the frozen PP2
aggregate receipt before a summary is admitted. Greedy prefix aggregate drift
is reported explicitly, with separate paired comparisons to frozen and current
greedy baselines. Aggregate equality cannot prove per-anchor numeric identity.
The current CPython 3.13/Torch 2.12 environment differs from PP2's
CPython 3.12/Torch 2.10 environment. Terminal suffix
censoring, 128-token schedule and available 128-row windows are unchanged.
For each anchor, the new FusedContext GPU producer emits dyadic24 q on all
five rows from production base logits and forced target-history previous IDs.
It computes the actual fused Markov bias and candidate selection. PP3's
Top32 producer runs separately on materialized BF16 conditioned rows as a
diagnostic, counting any differing support/q rows. The scorer consumes
those exact FP32 probabilities, checks integer ticket mass and positive draw
membership, and never renormalizes or reapplies temperature. The replay hashes
the transient sparse IDs/q and saves no logits, captured tensor or generated
text. Source hashes distinguish both producers. This qualifies the fused
selector on these heldout rows; it does not replace whole-worker graph,
physical-pair or lifecycle qualification.
The frozen target normalization/storage-error model is reused verbatim.
Prefix survival is cumprod(min(1,q(c)/p(c))) on the captured target-sampled
continuation. Expected retained tokens per nonterminal round are 1 plus the
sum of those survival values; this is a one-round counterfactual on fixed
target-history anchors, conditional on captured teacher/generation agreement.
It does not measure changed future round starts or throughput.
The exact PP2 bootstrap uses 2,000 paired connected-source-cluster resamples, seed 4102, separately reporting both heldout task families and context views. Each new point and paired 95% delta interval is compared to PP2's original point/interval. The >=32K view retains its unavailable interval because it has one source. Neither 150K nor broad task-family generalization is supported.
Reproduction uses /home/operator/miniconda3/bin/python3 throughout. Stage the
code and frozen aggregate files only under PP4's approved scratch paths. Set
PP4_SOURCE_ROOT to staged PP4 code and PP4_EXL3_PATH to separately qualified
CPython 3.13 stock/facc modules. Then run acceptance_run_locked.sh --out ...
inside PP4's named detached x86host tmux. The wrapper waits for the sole assigned
UUID's exclusive flock, enforces 225 W, taskset 12–15,28–31, nice 19 and eight
threads; the replay guards resident memory after every anchor. No fallback
interpreter, serving endpoint or service operation is present.
CPU validation: 20 tests pass, covering sparse/dense frozen scoring equality, the reverse-prefix dependence, invalid q refusal, exact frozen-bootstrap parity, source/population drift, missing full replay and frozen-byte tampering. Strict mypy passes for all four new Python files, ruff and shell syntax pass. These checks establish the estimator/admission logic and reconcile committed source/receipt identities. The separate GPU receipt establishes the heldout result above.