[Acentelles]

This reduces Akita verifier work inside a zkVM guest and pins Jolt to 4e55c9a5f7033d1b88254df051e0bd8c4a4b0ec3. Against main (038db2d1), total verifier rows fall 18.794% at inner $2^{12}$ and 21.171% at $2^{20}$, with unchanged Akita proof bytes. The pin and unconditional changes account for most of the combined reduction; the optional-feature saving is reported separately below.

Changes

The pin commit updates all Jolt dependencies, patches the profile workspace to this Akita checkout to keep one field identity, and handles the profile host's verify returning Result. A separate commit records the three regenerated lockfiles.

The 11 retained changes, including their supporting tests:

1. Signed sparse challenge evaluation uses Field::signed_sum for coefficients of magnitude one or two (unconditional).
2. ont-static-base uses generated D64 operator-norm base tables, with the existing builder as fallback.
3. Factored compression weights evaluate aligned event families by shared factors instead of individual events (unconditional).
4. compression-weights-crosscheck checks grouped weights against expanded events and dense/roundtrip tests.
5. The cross-check event-sum type gets a name so diagnostic Clippy passes.
6. Reduced setup rows sharing one functional use weighted_dot_base_rows (unconditional).
7. Reduced shared-row tests import Ring for their field constructors.
8. tensor-no-carry separates tabulated tensor residuals into high/low contractions.
9. tensor-eq-runs accumulates tensor equality weights over power-of-two runs.
10. Tensor runs use power-of-two shifts and test Fp128 boundaries; the main-compatible dispatch replaces the earlier formatting workaround.
11. setup-sop-structured batches the structured E/T/Z contractions through dot products and sums of products.

Two reworks make this independent of the reverted row kernels: factored weights call main's OffsetEqWindow::eval, and reduced rows keep main's eval_ring_at_pows_fast in the nonshared path. Tensor paths decline streamed residuals and preserve main's family-scalar semantics. The cross-check roundtrips also pass the current witness-chunking argument. The redundant ctutils bump is omitted.

Measurement

Jolt main plus the recursion example from #1911, with a host input-sizing patch, built against Akita main (BASE) or 6c25733560c72aca671a2a079991c8643310e807 with the four optional features off/on. All builds use Jolt field-inline; Akita hashing runs in software. AMD Ryzen 9 9950X; one run per arm and size. Counts are total real plus virtual verifier rows, not timing claims.

Inner size	BASE rows	OFF rows	ON rows	Akita bytes, every arm
$2^{12}$	174,970,923	145,186,867	142,087,101	80,689
$2^{20}$	357,355,426	296,255,058	281,699,170	84,460

Derived from those measured totals:

Row change	$2^{12}$	$2^{20}$
OFF minus BASE: Jolt pin plus unconditional code changes	-29,784,056	-61,100,368
ON minus OFF: enabling the four optional features	-3,099,766 (2.135% of OFF)	-14,555,888 (4.913% of OFF)
ON minus BASE: combined	-32,883,822	-75,656,256

There is no pin-only control: the pure pin benefit and the total contribution of all retained changes are not isolated. The outer Jolt measurement tree is held fixed and supplies its workspace jolt-field. Complete serialized proof files are byte-identical across arms at each size; Akita proof-byte deltas are zero.

Validation and limits

The latest commit moves the Jolt pin from e1d8e02b to 4e55c9a5 (a16z/jolt main, which adds range checks on direct RAM loads); it changes only dependency pins and lockfiles. At the new head the x86 CI mirror below was re-run and every gate passed, and the rows above were re-measured. The test counts, the flipped-bit rejection and the clean rebuild reproduction were measured at the previous head.

The x86 CI mirror passed format/TOML/typos, line/proof-stream/field-identity guards, all three CI Clippy configurations, dependency hygiene, docs, proof-channel roundtrips, six schedule-drift checks, both transcript-semantics suites, doc guardrails, portable x86/AVX2, profile smoke, and cargo-deny/audit. Nextest passed 2,051 tests (17 skipped); optional-feature nextest passed 2,061 (17 skipped). Feature-specific tests, three additional Clippy configurations, static-table equality/fallback tests, and compression cross-check unit/e2e tests passed.

The nightly fuzz check failed with debug-link relocation overflow on both main and the subject. AArch64 NEON, emulated AVX-512 IFMA, fuzz campaigns and the CI timing report were not run. Schedule regeneration was omitted because planner, schedules and artifacts are untouched. This is mirror evidence; GitHub CI must validate the pushed head.

All six honest proof traces accepted. A flipped opening bit was rejected by small ON with the same ELF. Size, feature-map, dependency-lock, proof-identity and byte gates passed; BASE/OFF/ON ELFs were distinct. The clean small ON rebuild reproduced rows and ELF exactly.

Risks and switches

The pin changes dependency resolution. Signed evaluation, factored weights and reduced-row batching affect every verifier; they have no feature switch and require reverting their commits to disable. Their test-only fixes have no production switch. Reverting the pin and locks also requires reverting code that uses the new field APIs. Generated tables must match the builder; grouped-weight and tensor fallback boundaries need continued coverage. Tensor edits are not line-preserving on main; structured sums preserve feature-off statement lines but add macro complexity. Row counts can vary with crate source identity and link layout.

All optional switches default off. Remove the following features from every enabling dependency edge to disable them:

Change	Feature
Static tables	akita-challenges/ont-static-base
No-carry contraction	akita-algebra/tensor-no-carry
Equality runs, including run shifts	akita-algebra/tensor-eq-runs
Structured sums	akita-verifier/setup-sop-structured and its akita-pcs forward
Compression diagnostic and its event-sum type	akita-types/compression-weights-crosscheck and its akita-pcs forward (disables checking, not factored evaluation)