Prototype evidence · not certification

The governance decision costs 258 nanoseconds.

What costs is reading the text — and that is a bill you already pay. These are measured figures from the internal technical evidence report, on one commodity workstation, with the tail and the outliers printed rather than trimmed.

01Component timings

Where the time actually goes

Each stage timed in isolation on the stated platform. The deterministic core is not the bottleneck; embedding generation is.

Component latency and throughput by pipeline stage
StageLatencyThroughput
Embedding, single text (CPU)55.8 ms18/s
Fixed-point adapter, 768-dim9.0 µs~110,000/s
Dual-reference integer gate10.3 µs97,490/s
Friction Gate258 ns3,880,000/s
Proof binding — SHA3 + HMAC11.4 µs87,719/s
Durable 64-record group commit29,618/s
Full post-embedding decision path469 µs2,132/s

Embedding generation dominates by two orders of magnitude. Every governance stage after it — the gate, the integer comparison, the proof binding — completes in microseconds or less.

02Tail distribution

50,000 timed decisions, tail included

Percentiles from the corrected full-path stress run. An earlier set of throughput figures was withdrawn after a harness artefact was found; these replace them.

Logarithmic axis — the spread is roughly seventy-five fold, and a linear one would hide everything except the outlier. The maximum was observed during periodic Merkle-commit activity and is not a formal worst-case bound.
Tail latency percentiles across 50,000 timed decisions
PercentileLatency
p50406.9 µs
p90765.3 µs
p991.351 ms
p99.92.595 ms
max30.832 ms
03Multi-worker scaling

Where parallelism stops paying

Synchronized steady-state benchmark. Efficiency is reported against the single-worker reference, including where it falls away.

Throughput and efficiency by worker count
WorkersThroughputSpeed-up · efficiency
12,132/s1.00×
23,854/s1.81× · ~90%
46,181/s2.90× · ~72%
87,373/s3.46× · ~43%

Efficiency drops to roughly 43% at eight workers, which is hyper-thread saturation on this machine rather than a property of the architecture. A separate, more instrumented four-worker harness measured about 4,800/s under a different evidence-recording configuration; the two are not directly comparable and neither is a distributed system rate.

04Determinism

The same input, the same bits

The claim the rest of the platform rests on, stated as it was measured.

1,000 / 1,000

Identical digests across repeated runs on the same input and pinned environment.

10,000 → 0.195 s

Prototype hash-chain replay: ten thousand ledger records verified in under a fifth of a second.

What this is not

The boundary of these figures

Carried over from the report's own limitations section, unedited. A number is only worth what its scope statement allows.

  • One commodity workstation — not a distributed or multi-node measurement.
  • Prototype stage. Not a certification, accreditation, or production authorisation.
  • No FIPS-validated cryptography or hardware-backed WORM media.
  • No accredited red team or independent semantic-layer assessment.
  • Production WORM and multi-node authority remain outstanding.
  • Not independently reproduced; the verification script becomes an audit only when a third party runs it.

The rest of the evidence

The conformance study carries the policy-accuracy side of the same engine, and the security page states the current assurance position in full — including what we do not hold.