Identity
Immutable source SHA, release tag, engine and Guard versions, OCI digest, artifact hashes, schemas, and compatibility profile.
Measured claims only
Performance claims start with reproducible evidence.
Every published result binds the source release, proof profile, workload, host, command, raw measurements, proof digest, and verifier result. Independent reproduction is recorded separately as an advisory metric.
No production performance claim is asserted here yet. The thresholds below are release gates, not measured customer promises. Candidate reports appear only after review and identity verification.
Fixed-host qualification matrix
| Rows | Workloads | Required evidence |
|---|---|---|
| 1,048,576 | Fibonacci and Poseidon2 | Conventional baseline and bounded run; bounded peak RSS at least 4× lower; bounded wall time no more than 3× baseline |
| 16,777,216 | Fibonacci only | Bounded peak RSS no more than 2 GiB; observed scratch within 10% of preflight; official verifier acceptance. Poseidon2 at this size remains a transparent post-GA capacity target. |
| External (advisory) | Three non-reference workloads | Not completed and not checkout-blocking; future results must use the same public adapter and customer-owned compute. |
Reference host
Linux x86-64, 8 CPUs, 15–17 GiB resident memory, swap disabled, cgroup v2 enforcement, local NVMe, and at least 500 GB free scratch. Reports from other hosts must be labeled separately and never substituted for the release matrix.
Evidence bundle
Resources
Whole-process peak RSS, elapsed time, CPU time, read and write bytes, scratch peak and estimate, exit class, and cgroup settings.
Correctness
Proof digest, conventional/bounded equality where required, official verifier outcome, crash-resume cases, and integrity failures.
Publication rules
- Component measurements are labeled as component measurements.
- Targets are never presented as achieved results.
- Every result names hardware, profile, workload, source, command, and raw report.
- Results are not generalized to unsupported AIRs, platforms, fields, hashes, forks, or row counts.
- Independent reproduction and specialist review are transparent, nonblocking advisory metrics.