Multi-table proving: one proof for heterogeneous hash batches - #21

Closed

rothblum wants to merge 154 commits into

mainfrom

Commits on Jul 20, 2026

pcs: jagged assist — delegate the verifier's f̂_t evaluation (paper §5)

rothblum
and
claude
committed
b5345fd
pcs: assist optimization — Lemma 4.6 streaming prover (27x)

rothblum
and
claude
committed
71b8036
pcs: assist prover engineering pass (2.2x -> 2.5ms)

rothblum
and
claude
committed
976a143
pcs: add runtime_bits30 bench (2^30-bit packed witness, 2^12 columns)

rothblum
and
claude
committed
f872f53
pcs: jagged prover — kill the q copy and the separate round-1 message pass

rothblum
and
claude
committed
7380a84
chore: restore spartan2 competitor-bench patch files after rebase
rothblum
committed
aac8895

Commits on Jul 21, 2026

chore: cargo fmt (rustfmt 1.97 stable)

rothblum
and
claude
committed
cff0269

Commits on Jul 22, 2026

docs: add multi-table proving design doc

rothblum
and
claude
committed
fc755f2
docs: adopt the implemented chunked BatchMajor layout; drop the row-bit skip

rothblum
and
claude
committed
a8fb9a4
docs: place ring-switching before the jagged transport

rothblum
and
claude
committed
14b79c6
docs: ring-switch folds obey the declared-work discipline; claim-batching rationale

rothblum
and
claude
committed
de15a85
docs: uniform capacity; lincheck as one union-column sumcheck; two-claim transport

rothblum
and
claude
committed
d36c3e2
Merge jagged-assist: jagged kernel + verifier-assist (f̂_t delegation)

rothblum
and
claude
committed
eb19607
docs: drop the historical 'One mode' paragraph

rothblum
and
claude
committed
d42e47c
zerocheck: generalize PaddingSpec to a run-list (single-run byte-identical)

rothblum
and
claude
committed
cd2250b
core: add multi-table registry and slot-schedule types (schedule.rs)

rothblum
and
claude
committed
0676af3
docs: spell out the BatchMajor run-list mapping in §5.2

rothblum
and
claude
committed
0839eaa
core: add registry digest (flock-registry-v1)

rothblum
and
claude
committed
93bb006
docs: restructure §7 as three polynomials and two reductions

rothblum
and
claude
committed
02b2f8c
pcs: add the jagged opening path (virtual-open sumcheck + transport)

rothblum
and
claude
committed
1c7aa2b
prover: wire the jagged opening path end to end

rothblum
and
claude
committed
499cd44
prover: add m=30 direct-vs-jagged throughput test

rothblum
and
claude
committed
0c4fff9

Commits on Jul 23, 2026

docs: identify the virtual-opening sumcheck as ring-switching's own sumcheck, un-fused

rothblum
and
claude
committed
203ea2c
pcs: fuse the jagged transport into the Ligerito opening

rothblum
and
claude
committed
7059f44
docs: record the fused transport and the send-and-spot-check reduction

rothblum
and
claude
committed
693f5da
prover: throughput test times witness gen + prove, matching canonical accounting

rothblum
and
claude
committed
3966723
prover: warm the scratch pool with a full untimed prove in the throughput test

rothblum
and
claude
committed
8faae37
core: add the union-instance layer (single-type M1 plumbing)

rothblum
and
claude
committed
a53dece
prover: union prove entry + M1 byte-identity differential oracle

rothblum
and
claude
committed
c4ccbeb
core: union-column lincheck (M2) with closed-form comb verification

rothblum
and
claude
committed
2064d95
prover: drive the union entry pair through the union-column lincheck

rothblum
and
claude
committed
f753433
core: flock-mixed-v1 statement binding on the union entries (M3)

rothblum
and
claude
committed
57ffc2c
prover: first end-to-end mixed BLAKE3+SHA-256 proofs (M3 milestone tests)

rothblum
and
claude
committed
95287b6
core+prover: true dense-stack commit for the union transport (M4)

rothblum
and
claude
committed
b09cd07
prover: partial-utilization witness generation (M4 dynamic counts)

rothblum
and
claude
committed
bd913ed
prover: wire format v5 with a Mixed flavor + CLI --mix (M4)

rothblum
and
claude
committed
9373e62
docs: update the multi-table design doc to the implemented Phases 0-2

rothblum
and
claude
committed
439abb1
core: height-n_t stacking — count-proportional committed size (M5)

rothblum
and
claude
committed
e406c24
prover: per-proof PCS params + M5 area gates for height-n_t stacking

rothblum
and
claude
committed
ad754b6
docs: height-n_t stacking (M5) — count-proportional committed size

rothblum
and
claude
committed
677d385
tests: M6 byte-identity fixtures — proof bytes pinned at the pre-M6 prover

rothblum
and
claude
committed
f7b4e17
core: support-proportional zerocheck tail + count-proportional lincheck row folds (M6)

rothblum
and
claude
committed
2be53c3
core: support-proportional opening passes — virtual-open f-side, round-0 prime, batched b_tilde (M6)

rothblum
and
claude
committed
fc4543c
tests+docs: M6 low-utilization gate assertion; refresh run-list comments

rothblum
and
claude
committed
3d60337
prover: add m=30-scale mixed throughput measurement

rothblum
and
claude
committed
4614230

Commits on Jul 24, 2026

prover: add ntt_layout_probe for zero-lane-skip layout comparison

rothblum
and
claude
committed
1f61e7a
ntt+pcs: integer-lane commit (Piece 1 — encode + Merkle)

rothblum
and
claude
committed
75a81d3
prover: two-blake3-table control benchmark (multitable machinery overhead)

rothblum
and
claude
committed
b5159da
prover: per-phase breakdown for single vs two-table blake3

rothblum
and
claude
committed
81151aa
core+prover: parallel union witness scatter/compaction + scatter-vs-compact timers

rothblum
and
claude
committed
dc72394
core+prover: generate the union witness in place (kill the assembly scatter)

rothblum
and
claude
committed
c8462a0
pcs: integer-lane commit via high-bit lanes (stop committing the zero padding)

rothblum
and
claude
committed
2012181
pcs: fold L0 lane variables at block granularity (drop the lane-grid transpose)

rothblum
and
claude
committed
66d90e9
prover: integer lanes apply to identity compaction too (fix a false assert)

rothblum
and
claude
committed
1eec039
zerocheck: share the round-2 SIMD kernels with the run-list padding path

rothblum
and
claude
committed
bd0f222
jagged: fuse the blocked round-0 prime into the weight-table build

rothblum
and
claude
committed
4b2db19
core: take the union witness buffers from the scratch pool, not vec![ZERO]

rothblum
and
claude
committed
0140131
ligerito+jagged: skip the known-zero lanes in the L0 fold

rothblum
and
claude
committed
09036b5
lincheck: use the count-proportional row fold for ANY partial count

rothblum
and
claude
committed
44ff9be

Commits on Jul 25, 2026

tests: pin the factored-basis property for skipping W_rho materialization

rothblum
and
claude
committed
9b8a63d
tests: measure the JIT jagged-weight fold — 2.3 ms and 134 MB, any heights

rothblum
and
claude
committed
d06ff6f
jagged: JIT factored weight + round-0 claim (step 1 of dropping W_rho)

rothblum
and
claude
committed
23c0811
pcs: source L0's basis just-in-time — drop the 134 MB W_rho array

rothblum
and
claude
committed
82aca3e
scratch: pool the lincheck stripe (mechanism sound, gain below the noise floor)

rothblum
and
claude
committed
6bfad5f

Commits on Jul 27, 2026

tests: capacity sweep — fix the workload, grow the capacity

rothblum
and
claude
committed
383c6cd
pcs: stream the γ-combined basis — b_combined is never materialized

rothblum
and
claude
committed
f827aa3
docs: v0.5 — integer-lane commit, JIT basis, declared-work passes, streaming b

rothblum
and
claude
committed
cc112b3
ring_switch: record the rejected merged Φ-table scan

rothblum
and
claude
committed
c8d9f60
zerocheck: parallelize the sparse tail (dense-kernel structure)

rothblum
and
claude
committed
b3b8c6d
tests: sweep on the in-place path; serialize the timing tests

rothblum
and
claude
committed
0054f88
zerocheck: skip wholly-dead chunks in the sparse round 2

rothblum
and
claude
committed
ad7094d
zerocheck: engage the sparse tail at quarter utilization (gate 16 -> 4)

rothblum
and
claude
committed
68abbb7
tests: capacity sweep at the real m30 load

rothblum
and
claude
committed
649877a
zerocheck: engage the sparse tail at half utilization (gate 4 -> 2)

rothblum
and
claude
committed
3f51cfb
docs: v0.6 — the merged jagged/ring-switch reduction (design sketch)

rothblum
and
claude
committed
f8bbe6e
docs: the fat verifier is registry-column-bounded (Ron's catch)

rothblum
and
claude
committed
c31b2af
jagged+docs: the Frobenius decomposition resolves the twisted assist

rothblum
and
claude
committed
674884a
docs: the batched Frobenius assist, in detail

rothblum
and
claude
committed
c2ae82e
docs: merged reduction — define W's zero tail; scope and wire notes

rothblum
and
claude
committed
b6cfecf
pcs+jagged: the merged jagged/ring-switch transport (prototype)

rothblum
and
claude
committed
e1fbcd9
docs: the merged transport is a measured prototype, not a sketch

rothblum
and
claude
committed
1a89b6d
pcs+jagged: merged-transport optimization pass (-18 ms)

rothblum
and
claude
committed
f034fe0
pcs+ligerito: merged transport — integer lanes, fused prime, assist tuning

rothblum
and
claude
committed
22201c7
docs: capacity-free ring switching as a standalone paper draft

rothblum
and
claude
committed
9d3337b
docs: tighten the paper to a research note

rothblum
and
claude
committed
2b0f795
docs: note — spell out the z_r/z_c positional split in the lemma

rothblum
and
claude
committed
f964909
docs: note — assume Σ h_y = 2^m (padding column), drop the tail cases

rothblum
and
claude
committed
f22eb16
docs: note — make the claim index explicit

rothblum
and
claude
committed
268b13c
docs: note — single-claim presentation, multi-claim as a closing remark

rothblum
and
claude
committed
95167a7
docs: note — f_jag's index-point argument is rho, not z_i

rothblum
and
claude
committed
de403d4
docs: note — Lemma 1 takes z_r, z_c directly

rothblum
and
claude
committed
cacf91f
docs: note — rewrite Lemma 1's proof as an explicit equality chain

rothblum
and
claude
committed
5b6772a
tests: merged transport — tamper the commitment params

rothblum
and
claude
committed
d0095e7

Commits on Jul 28, 2026

mixed: ship the merged transport — wire v6

rothblum
and
claude
committed
a6a757f
tests: m30 probe — three tiers, both arms on the shipped lane config

rothblum
and
claude
committed
4245770
jagged: quadrant-product challenge fold in the Frobenius assist rounds

rothblum
and
claude
committed
f0388d0
note: the dual form of the Frobenius decomposition (Remark 5)

rothblum
and
claude
committed
627df4b
jagged: multipoint twisted evaluation — the column-scalable assist

rothblum
and
claude
committed
7c2393a
jagged: optimize the multipoint prove — 162 -> 95 ms at 32K columns

rothblum
and
claude
committed
0777c28
jagged: strided low tables for the multipoint values — 95 -> 35 ms at 32K

rothblum
and
claude
committed
111d164
jagged: column-parallel assist dispatch for small statement counts

rothblum
and
claude
committed
93ce0da
jagged: raise the column-parallel gate to 2^14 columns

rothblum
and
claude
committed
7b21319
jagged: fused multipoint prover — shared merged-open context

rothblum
and
claude
committed
0606fe3
prover: phase timers on the merged prove + capacity-attribution probe

rothblum
and
claude
committed
afe1c11
union: padding-aware witness buffer sourcing — nu18 prove 872 -> 265 ms

rothblum
and
claude
committed
59dd098
union: PooledDirty witness mode — padding is unread, stop zeroing it

rothblum
and
claude
committed
22bca4e
tests: merged capacity probe — five tiers, 1x to 16x over-capacity

rothblum
and
claude
committed
912066e
tests: poison-pool coverage at full utilization

rothblum
and
claude
committed
8a2237a
scratch: window-preferring pool takes, class-aware eviction, take trace

rothblum
and
claude
committed
4dcd471
zerocheck: sparse-always (SPARSE_TAIL_GATE 2 -> 1) + gate env knob;

rothblum
and
claude
committed
c6393d9
zerocheck: interval-derived sparse tasks — round 2 stops scanning the padded domain

rothblum
and
claude
committed
b6c8d9a

Commits on Jul 29, 2026

scratch: union prewarm sized by buffer class, not by the padded m

rothblum
and
claude
committed
2e63bb5
zerocheck: live-span fold storage — the tail stops scattering over the padded domain

rothblum
and
claude
committed
9f74dc7
permutation: one shared eq(ρ) for the five v-openings, not five dense builds

rothblum
and
claude
committed
3120f5b
product_gkr: port the GKR grand product, then fuse its fold with the next round

rothblum
and
claude
committed
81d0552
product_gkr: s_id is the index, so stop building a table for it

rothblum
and
claude
committed
ccafc90
product_gkr: point the fold probe at the kernel that actually runs

rothblum
and
claude
committed
2e82e8b
product_gkr: verifier review — pin the s_σ trust boundary, stop panicking on shape

rothblum
and
claude
committed
e8b865e
Merge main into multitable: selectable Merkle/FS hashes

rothblum
and
claude
committed
9a4c7bf
element_r1cs: the large-field table type and its two PIOP phases

rothblum
and
claude
committed
5339bee
element_r1cs: wire the phases into a single-table proof, tamper matrix

rothblum
and
claude
committed
bc6bab0
element_r1cs: pin the dummy-row scope boundary instead of implying it

rothblum
and
claude
committed
5974bd6
blake3: word-align every I/O region, move the const pin to the end

rothblum
and
claude
committed
f95dfbb
element_r1cs: collapse the lincheck's row half, dropping n_log rounds

rothblum
and
claude
committed
c87a067
schedule: class-tag the table types, keep the boolean digest byte-identical

rothblum
and
claude
committed
f980da5
union: element slots through the layout, boolean PIOP on its prefix subcube

rothblum
and
claude
committed
c3d89f0
element_r1cs: the element-region PIOP, one zerocheck and one collapsed lincheck

rothblum
and
claude
committed
dc0a6ca
union: mixed boolean+element proofs through one jagged opening

rothblum
and
claude
committed
21c0901
union: fix the M_bool = 0 underflow, make both build profiles meaningful

rothblum
and
claude
committed
09918ed

Commits on Jul 30, 2026

union_element: cost probe for the element class, arms interleaved

rothblum
and
claude
committed
c6a00bb
pcs: skip the combine's dead blocks, exploiting the classes' disjointness

rothblum
and
claude
committed
b289f39
prove_union: PCS_TRACE phase breakdown, and a fair witness path in the probe

rothblum
and
claude
committed
c1e7e10
element_r1cs: count-proportional row rounds, bit-identical to the dense ones

rothblum
and
claude
committed
137a5db
docs: circuit-wiring design v0.2 — circuits, large-field gates, recursion

rothblum
and
claude
committed
6eb7b69
docs: spell out the copy-constraint mechanism in the wiring section

rothblum
and
claude
committed
eb8bbbb
docs: disambiguate "copy constraints" from R1CS constraints

rothblum
and
claude
committed
74cf2de
docs: the capacity-tax masking puts 1's in the leaf vector, not the trace

rothblum
and
claude
committed
2823aa8
docs: dead-address audit and mitigation plan for the two-subcube layout

rothblum
and
claude
committed
20200de
docs: worked example for the wiring cell space

rothblum
and
claude
committed
292a30d
docs: the IO schema is a per-type choice — chaining vs IV-pinned sha2

rothblum
and
claude
committed
2689b84
docs: taxonomy-of-emptiness remark — dead addresses vs padded rows

rothblum
and
claude
committed
47cb140
product_gkr: the two sanctioned hardening items

rothblum
and
claude
committed
7040cf5
docs: the boundary run tree — assist cost and measurements

rothblum
and
claude
committed
7fd4500

Commits on Jul 31, 2026

pcs: packed-direct intake for the merged transport

rothblum
and
claude
committed
11c1973
prover/verifier: mixed-class proofs over the merged transport

rothblum
and
claude
committed
ed7ab89

Commits on Aug 2, 2026

prover/pcs: defer the element eq tensors — merged never reads them

rothblum
and
claude
committed
0a0b0b4
pcs/prover: element-only and identity-compaction fixes on the merged path

rothblum
and
claude
committed
c113c12
tests: merged-transport byte pins, tamper walk, and the final A/B record

rothblum
and
claude
committed
fa84ab0
tests: retarget the union suites jagged -> merged

rothblum
and
claude
committed
3bb2539
prover/verifier/pcs: delete the jagged transport

rothblum
and
claude
committed
76f1e86
prover/verifier: unify the shipped merged entries with the shared body

rothblum
and
claude
committed
b0cfc1a
prover/verifier/pcs: drop the jagged/merged name fossils

rothblum
and
claude
committed
600f75e
pcs/jagged: collapse shared-row scalar claims in the merged open

rothblum
and
claude
committed
8e8c1a3
pcs/jagged: one-hot fast paths for Boolean column points

rothblum
and
claude
committed
ab2aa08