JSpark3 release notes
v1.1.0 (Cadence) — 2026-09-07
GitHub release.
The sealed candidate passed final review. These are the authorized release
notes for the tagged commit; observed publication results are recorded
separately. The v1.0.0 notes below are final and historical.
Summary. v1.1 — release name Cadence — keeps the entire v1.0.1
construction and adds two measured single-stream serving features to the
recipe: a group-128 INT8 shadow for the QKV projections on small-batch pure
decode, and a request-local speculative width controller that narrows the
DFlash2 draft from seven to three tokens on high acceptance. Batches of two
or more requests and prefill fall back to the wide path. The stock indexer
workspace, the original drafter, and the sampler are retained. Weights,
checkpoints, and image digests are unchanged.
Measured scope (paired, single-stream, two independent serving starts).
Prose +16.18% and +19.17%; structured count +7.66% and +7.61% (95% CIs
exclude zero). Code +9.71% then +3.23% with the second interval spanning
zero, so the paired code gain is not replicated. The first start's sham
control failed its predeclared resolution margin, so first-start figures are
diagnostic only; the second start's predeclared sham passes. A fixed quality
battery showed candidate-only delivered-answer failures and stays
inconclusive on population-level semantic parity. A short-prompt concurrency
burst is finite-burst evidence and certifies neither sustained service nor
large-context capacity. Machine-readable source of record:
results/evidence/candidate/cadence-v11/CLAIMS.json.
Verifier hardening (F-1). The public verifier no longer requires the
stock Capturing CUDA graphs (FULL) progress bar, which the byte-pinned
Cadence capture path never renders. It now requires, on every rank, the
pinned path's full-bank capture evidence: at least 16 capture receipts whose
recorded graph-dump SHA-256 matches the on-container dump bytes, and at
least 8 non-empty serving graph dumps — the per-rank construction measured
in the archived measured-arm startup receipts. All other runtime identity,
load, memory, witness, and arithmetic gates are retained.
Integrated live verification passed. Host verifier 456a262 accepted
unchanged candidate a729583, including all-rank identity, safety, Cadence
capture, load, health, arithmetic, focused, and strict long-context gates.
The pinned long-context request reported 48,957 prompt tokens and 51
completion tokens with the codeword returned verbatim. The focused median
was 73.8718 tok/s for its fixed admission workload only; Pi slowdown remains
unresolved, and E3 is separate. Sanitized evidence and provenance
retain the earlier verifier failures and explain recipe equivalence. The
final archive has not been cold-boot tested; maximum context, sustained
concurrency, and independent-fleet reproduction remain unverified. Publication
is authorized following final review. v1.1 ships one mandatory,
measured Cadence construction: the recipe and its verifier carry no toggle
and no opt-in variant, because a second, never-measured path would be
exactly the kind of unverified claim this project refuses to ship. This is a
scoped engineering release choice — it is not a global performance or quality
claim, and all limits above stand. Whether the maintainer's own live service
later runs this public build is a separate decision outside this release. No
global or cross-recipe speed ranking is claimed, and no figure on this page
is recomputed from any post-hoc repeat.