Evidence bench
what this page isThe measurement bench is a private test environment, not a product installation path. Public summaries retain its results, dates, scope and limitations. What is proven, observed, preview, and planned is dated per row on the availability matrix.
bench record: clean-room verifiedThe sealed reproduction document was run end to end on a fresh machine at v0.1.26; the redacted record and the attestation are public, and the verbatim bytes are attested by hash.
redacted record: clean-room run · release commit 8846dbd6f7ce · pinned file set: versions-lock-public
The evidence bench is the private aarch64 machine the bench series were measured on: one Apple Silicon host running the engine under its gates. It is not the only place a published number comes from. The x86_64 fleet carries its own gates, and a further family of results is canaried against the deployed service; which measurement family any number belongs to is set out on the methodology page.
The bench is not a product, not an onboarding path, and not reproducible from public bytes. This page exists so reviewers know exactly what the bench is, what its gates prove, and how to verify what we publish about it.
What the bench proves
Tests compare observed behavior with explicit acceptance criteria. Negative controls check whether the test detects a deliberately defeated property. A passing infrastructure test does not prove model intelligence, task acceptance or performance on other workloads.
The public evidence summaries preserve dates, functional results and scope. Internal scripts and operating records remain private.
The sealed quickstart, by hash
The historical reproduction record concerns the laboratory environment, not the customer quickstart. Its original reproduction material was sealed privately. A public attestation identifies what was checked; it does not supply the private material or establish that any reader can reproduce it.
The original failed/aborted attempts remain historical facts. The later v0.1.4 reproduction test passed under its amended acceptance rule. See the reproduction summary and historical block summary. A revised summary has its own digest and does not inherit a signature over different bytes.
Verify any published artifact
Download the public summary using the link on its evidence page, compute SHA-256 over the exact downloaded bytes, and compare it with that page’s digest. Matching a digest establishes byte identity, not correctness, current deployment or permission to read an internal record.
What the bench is not
It is not the hosted service (sessions run on the KeepState fleet; see how sessions run). It is not a second deployment model. And a number measured on it is a number about it: bench scope is stated wherever bench numbers appear, and fleet-side numbers publish when the fleet’s own gates run, not before.