Reflexbench

Partly. The v1 headline on the frozen 111-case public-hard cohort reports raw semantic accuracy with Jev at 73.0% against much lower small-model scores, but the repo explicitly refuses a single best-model score and reports calibration, order robustness, cardinality, failures, and latency boundaries separately.

Core is a tiny deterministic workflow-owned policy applied to the same probability response with no rewrite, adapter, or extra inference pass. The headline claim is 106 of 110 to 110 of 110 with zero extra model calls, kept as a separate research lane from raw-engine quality.

Implement the adapter contract, meet identity and provenance requirements, and open a result PR under the documented workflow. Accepted public results may use the Evaluated on ReflexBench v1 badge, which records reproducible evidence rather than endorsement.

ReflexBench covers decision-engine quality, AlignmentBench covers target-system behavior under authority and oversight constraints, and Reflex Alignment covers execution-time supervision of one consequential action. They are kept separate so one score cannot stand in for all three.

Yes. Core v1 uses only the Python standard library, and the core harness needs Python 3.11 or newer. Reproduction is PYTHONPATH=reflexbench python -m unittest discover -s tests -p 'test_*.py' -v. There is also a runner that points at any System One-compatible HTTP endpoint, and a validator that checks a Reflex recipe registry against the same provider-neutral contract.

Apache-2.0 for Brida-authored code, documentation and fixtures in the repository, while third-party datasets and models keep their own licences and raw third-party benchmark data is not redistributed unless that licence permits it. Deliberately absent: customer data, provider credentials, private evaluation corpora, hidden future gates, Brida weights, hosted routing, tenant authorization and billing. Brida's hosted control plane is a separately operated service.

It agreed on the sign and disagreed on the size. The same-response Core proof went from 96.4%, 106 of 110, to 100% with 4 rescues, 0 harms and no extra model calls, while an independent blind workflow replication moved 76.7% raw to 82.0% Core across 150 of 150 completed cases, a paired delta of +5.3 pp with bootstrap 95% of +1.3 to +10.0. Gate v1 stays a formal FAIL at 149 of 150 after one provider 503.