Prover

SASH zk-inference-prover

One proof in the benchmark's sense is a sequence's whole grid of segments, every prompt and continuation position (systems/sash/README.md).

Pinned at sg-ai-safety-hub/zk-inference-prover@eb95bfa887299d1637550228d42086c27cbb2b3b

Claim card

What a verifier accept meansconfiguration defaultcard v1draft, not reviewed

not stated

Draft for layout: no SASH card is in the repository yet. Coverage and flags copy systems/sash/README.md; every other field says not stated until it is filled from upstream.

Claim class: not stated
Security notion: not stated
Security level: not stated
Model: not stated, adversary not stated
Setup: not stated
Zero knowledge: claimed not stated; hidden: not stated; not hidden: not stated; status not stated
Coverage: phase both, unit segment_grid; tokens covered: every prompt and continuation position in the grid; operations: not stated; policy: not stated
Binding: per-edge table stated upstream (docs/statement.md at the pin)
Assumptions: not stated
Residual: not stated
Availability: not stated
Forbidden flags: --unsound-skip---unsound-argmax-lax
Upstream pin: sg-ai-safety-hub/zk-inference-prover@eb95bfa887299d1637550228d42086c27cbb2b3b
Reviewed by: no review yet

Rows

One block per workload and configuration. Repetition strips use each metric's own axis.

Workload sash-l31-8b-c1024

sash / default on sash-l31-8b-c1024

profile prover-h100tier reported

Measured

setup_seconds

40.83 s

scope setupgrade measuredtier reportedmock

mean 40.83, sd 0.3512, min 40.50, max 41.20, n 3, n_failed 0

source: mock harness record

commit_seconds

63.93 s

scope commitgrade measuredtier reportedmock

mean 63.93, sd 0.4509, min 63.50, max 64.40, n 3, n_failed 0

source: mock harness record

prove_session_seconds

1,318 s

scope prove.sessiongrade measuredtier reportedmock

mean 1,318, sd 13.85, min 1,304, max 1,332, n 3, n_failed 0

source: mock harness record

prove_steady_seconds

null (not_observable) s

scope prove.steadygrade measuredtier reportedmock

source: mock harness record

whole_wall_seconds

1,382 s

scope whole_wallgrade measuredtier reportedmock

mean 1,382, sd 13.40, min 1,368, max 1,395, n 3, n_failed 0

source: mock harness record

tokens_covered_total

1,088 token

grade measuredtier reportedmock

mean 1,088, sd 0, min 1,088, max 1,088, n 3, n_failed 0

source: mock harness record

requests_count

1 count

grade measuredtier reportedmock

mean 1, sd 0, min 1, max 1, n 3, n_failed 0

source: mock harness record

proofs_count

1 count

grade measuredtier reportedmock

mean 1, sd 0, min 1, max 1, n 3, n_failed 0

source: mock harness record

proof_bytes_total

4.870e9 B

grade measuredtier reportedmock

mean 4.870e9, sd 0, min 4.870e9, max 4.870e9, n 3, n_failed 0

source: mock harness record

verify_seconds

30.17 s

scope verifygrade measuredtier reportedmock

mean 30.17, sd 0.3512, min 29.80, max 30.50, n 3, n_failed 0

source: mock harness record

verify_seconds_total

30.17 s

scope verifygrade measuredtier reportedmock

mean 30.17, sd 0.3512, min 29.80, max 30.50, n 3, n_failed 0

source: mock harness record

verify_single_thread_seconds

198.4 s

scope verify.single_threadgrade measuredtier reportedmock

mean 198.4, sd 1.250, min 197.1, max 199.6, n 3, n_failed 0

source: mock harness record

Derived (computed by the derivation code)

rate_whole_wall

0.7873 token/s

scope whole_wallgrade derivedtier reportedmock

mean 0.7873, sd 0.007639, min 0.7798, max 0.7951, n 3, n_failed 0

source: mock derivation record

rate_prove.session

0.8255 token/s

scope prove.sessiongrade derivedtier reportedmock

mean 0.8255, sd 0.008681, min 0.8170, max 0.8344, n 3, n_failed 0

source: mock derivation record

rate_prove.steady

null (not_observable) token/s

scope prove.steadygrade derivedtier reportedmock

source: mock derivation record

cost_per_1k_tokens_whole_wall

1.514 USD/1k_token

scope whole_wallgrade derivedtier reportedmock

mean 1.514, sd 0.01468, min 1.499, max 1.528, n 3, n_failed 0

source: mock derivation record

proof_bytes_per_token

4,476,000 B/token

grade derivedtier reportedmock

mean 4,476,000, sd 0, min 4,476,000, max 4,476,000, n 3, n_failed 0

source: mock derivation record

verify_seconds_per_token

0.02773 s/token

scope verifygrade derivedtier reportedmock

mean 0.02773, sd 0.0003228, min 0.02739, max 0.02803, n 3, n_failed 0

source: mock derivation record

Runs

mock-sash-c1024-r1completed
mock-sash-c1024-r2completed
mock-sash-c1024-r3completed
Superseded, failed and ineligible runs never contribute to statistics (RFC-0009 S7).

Run anatomy

setup

commit

prove.session

prove.steady

null (not_observable)

whole_wall

verify

releases

Repetition 1, mock values, seconds from the setup call start; total 1,453 s. Scopes as RFC-0004 defines them.

Workload sash-l31-8b-c4096

sash / default on sash-l31-8b-c4096

profile prover-h100tier peer-reviewed

Measured

setup_seconds

41.40 s

scope setupgrade measuredtier peer-reviewedmock

mean 41.40, sd 0.5568, min 40.90, max 42, n 3, n_failed 1

source: mock harness record

commit_seconds

64.37 s

scope commitgrade measuredtier peer-reviewedmock

mean 64.37, sd 0.6658, min 63.80, max 65.10, n 3, n_failed 1

source: mock harness record

prove_session_seconds

5,415 s

scope prove.sessiongrade measuredtier peer-reviewedmock

mean 5,415, sd 60.57, min 5,362, max 5,481, n 3, n_failed 1

source: mock harness record

prove_steady_seconds

null (not_observable) s

scope prove.steadygrade measuredtier peer-reviewedmock

source: mock harness record

whole_wall_seconds

5,479 s

scope whole_wallgrade measuredtier peer-reviewedmock

mean 5,479, sd 61.23, min 5,426, max 5,546, n 3, n_failed 1

source: mock harness record

tokens_covered_total

4,608 token

grade measuredtier peer-reviewedmock

mean 4,608, sd 0, min 4,608, max 4,608, n 3, n_failed 1

source: mock harness record

requests_count

1 count

grade measuredtier peer-reviewedmock

mean 1, sd 0, min 1, max 1, n 3, n_failed 1

source: mock harness record

proofs_count

1 count

grade measuredtier peer-reviewedmock

mean 1, sd 0, min 1, max 1, n 3, n_failed 1

source: mock harness record

proof_bytes_total

1.980e10 B

grade measuredtier peer-reviewedmock

mean 1.980e10, sd 0, min 1.980e10, max 1.980e10, n 3, n_failed 1

source: mock harness record

verify_seconds

121.3 s

scope verifygrade measuredtier peer-reviewedmock

mean 121.3, sd 1.405, min 119.8, max 122.6, n 3, n_failed 1

source: mock harness record

verify_seconds_total

121.3 s

scope verifygrade measuredtier peer-reviewedmock

mean 121.3, sd 1.405, min 119.8, max 122.6, n 3, n_failed 1

source: mock harness record

verify_single_thread_seconds

802.6 s

scope verify.single_threadgrade measuredtier peer-reviewedmock

mean 802.6, sd 3.602, min 799.1, max 806.3, n 3, n_failed 1

source: mock harness record

Derived (computed by the derivation code)

rate_whole_wall

0.8411 token/s

scope whole_wallgrade derivedtier peer-reviewedmock

mean 0.8411, sd 0.009368, min 0.8309, max 0.8493, n 3, n_failed 1

source: mock derivation record

rate_prove.session

0.8511 token/s

scope prove.sessiongrade derivedtier peer-reviewedmock

mean 0.8511, sd 0.009488, min 0.8408, max 0.8594, n 3, n_failed 1

source: mock derivation record

rate_prove.steady

null (not_observable) token/s

scope prove.steadygrade derivedtier peer-reviewedmock

source: mock derivation record

cost_per_1k_tokens_whole_wall

1.417 USD/1k_token

scope whole_wallgrade derivedtier peer-reviewedmock

mean 1.417, sd 0.01584, min 1.403, max 1.434, n 3, n_failed 1

source: mock derivation record

proof_bytes_per_token

4,297,000 B/token

grade derivedtier peer-reviewedmock

mean 4,297,000, sd 0, min 4,297,000, max 4,297,000, n 3, n_failed 1

source: mock derivation record

verify_seconds_per_token

0.02632 s/token

scope verifygrade derivedtier peer-reviewedmock

mean 0.02632, sd 0.0003049, min 0.02600, max 0.02661, n 3, n_failed 1

source: mock derivation record

Runs

~~mock-sash-c4096-r1~~supersededreplaced by mock-sash-c4096-r4: the verifier container ran with 4 CPUs instead of 8 (mock correction).
mock-sash-c4096-r2completed
mock-sash-c4096-r3completed
mock-sash-c4096-r4completed

~~mock-sash-c4096-r5~~failedexit code 1, last lines:

upload: archive part 3 of 7: connection reset
harness: run failed in phase upload, proofs kept on the runner

Superseded, failed and ineligible runs never contribute to statistics (RFC-0009 S7).

Run anatomy

setup

commit

prove.session

prove.steady

null (not_observable)

whole_wall

verify

releases

Repetition 1, mock values, seconds from the setup call start; total 5,629 s. Scopes as RFC-0004 defines them.

Controls

All controls

sash-l31-8b-c1024 profile prover-h100

7 passed0 failed3 void, not applicable or error

sash-l31-8b-c4096 profile prover-h100

8 passed0 failed3 void, not applicable or error

Reference rows on the same workloads

RFC-0015

ref-recompute on sash-l31-8b-c4096

profile prover-h100tier reported

Measured

whole_wall_seconds

9.827 s

scope whole_wallgrade measuredtier reportedmock

mean 9.827, sd 0.1106, min 9.710, max 9.930, n 3, n_failed 0

source: mock harness record

tokens_covered_total

4,608 token

grade measuredtier reportedmock

mean 4,608, sd 0, min 4,608, max 4,608, n 3, n_failed 0

source: mock harness record

Derived (computed by the derivation code)

rate_whole_wall

469.0 token/s

scope whole_wallgrade derivedtier reportedmock

mean 469.0, sd 5.289, min 464.0, max 474.6, n 3, n_failed 0

source: mock derivation record

cost_per_1k_tokens_whole_wall

0.002541 USD/1k_token

scope whole_wallgrade derivedtier reportedmock

mean 0.002541, sd 2.860e-5, min 0.002511, max 0.002568, n 3, n_failed 0

source: mock derivation record

Runs

mock-rr-c4096-r1completed
mock-rr-c4096-r2completed
mock-rr-c4096-r3completed
Superseded, failed and ineligible runs never contribute to statistics (RFC-0009 S7).

ref-serve on sash-l31-8b-c4096

profile prover-h100tier reported

Measured

whole_wall_seconds

3.420 s

scope whole_wallgrade measuredtier reportedmock

mean 3.420, sd 0.03000, min 3.390, max 3.450, n 3, n_failed 0

source: mock harness record

tokens_covered_total

4,608 token

grade measuredtier reportedmock

mean 4,608, sd 0, min 4,608, max 4,608, n 3, n_failed 0

source: mock harness record

Derived (computed by the derivation code)

rate_whole_wall

1,347 token/s

scope whole_wallgrade derivedtier reportedmock

mean 1,347, sd 11.82, min 1,336, max 1,359, n 3, n_failed 0

source: mock derivation record

cost_per_1k_tokens_whole_wall

0.0008844 USD/1k_token

scope whole_wallgrade derivedtier reportedmock

mean 0.0008844, sd 7.758e-6, min 0.0008767, max 0.0008922, n 3, n_failed 0

source: mock derivation record

Runs

mock-rs-c4096-r1completed
mock-rs-c4096-r2completed
mock-rs-c4096-r3completed
Superseded, failed and ineligible runs never contribute to statistics (RFC-0009 S7).

Figures the system publishes upstream

Real data from systems/sash/external.yaml, grade external

Copied at the pin, never recomputed or rounded in the record. They are context, never inputs to a measured number (RFC-0004 INV-MAP-1). The mapping from the upstream scope is marked approximate where the definitions differ.

Upstream figures
Metric	Workload	Value and labels	Upstream scope	Note
rate_prove.steady	sash-l31-8b-c4096	1.251 token/s scope prove.steadygrade externaltier reported source: upstream	loop mapping approximate	positions/s over 4,608 positions (4,096 prompt and 512 output); run 2026-09-09-headline-c4096-h100
rate_whole_wall	sash-l31-8b-c4096	0.8401 token/s scope whole_wallgrade externaltier reported source: upstream	whole wall mapping approximate	positions/s over 4,608 positions (4,096 prompt and 512 output); run 2026-09-09-headline-c4096-h100
prove_steady_seconds	sash-l31-8b-c4096	3,683 s scope prove.steadygrade externaltier reported source: upstream	loop mapping approximate	the 5 segment-loop processes' own wall-clock time added up; run 2026-09-09-headline-c4096-h100
whole_wall_seconds	sash-l31-8b-c4096	5,485 s scope whole_wallgrade externaltier reported source: upstream	whole wall mapping approximate	trace generation, interaction generation, stwo total and serialization, added up per segment; run 2026-09-09-headline-c4096-h100

Upstream pin sg-ai-safety-hub/zk-inference-results@11decec3c0b62c707618aab5deb82750b355d202

Reply from the system maintainers

No reply for this release. System maintainers may write up to 300 words, shown here unedited (RFC-0009 §Right of reply).