Per-record sensitivity propagation
Measured against a 4-arg typed DAG, warm cache, single-record perturbation. Methodology and reproducer in the bench file.
bench · commit 46a6328When the Heartbeat says 87 ns per sensitivity, this is the file the number comes from. When we say bit-identical, this is the test that checks it. When we say 100/100 rubric, this is what the rubric grades on.
Measured against a 4-arg typed DAG, warm cache, single-record perturbation. Methodology and reproducer in the bench file.
bench · commit 46a6328Cold + warm compile sweep across 50 / 200 / 500 / 1000 drivers. 1,024-driver case runs under one second cold.
bench · scale-warmSix dimensions, 100 points total, 90-point institutional pass threshold. Production exports score 100/100 on the SaaS, DTC, and LBO institutional test models.
rubric · six-sheet test →The compile pipeline is a pure function over the IR: same input, same output, every machine, every run. No floating-point drift across cold and warm caches; no platform variance between macOS arm64 and Linux amd64 in our CI matrix.
determinism guaranteeMCP endpoint at api.flatlandfi.com/mcp · REST endpoint at api.flatlandfi.com/api · same engine, two protocols. Live discovery at /.well-known/oauth-protected-resource.
API reference →50 answers per calendar month per API key. No card required. Two free API keys per account. Resets on the first of the calendar month.
pricing detail →50-driver compile, warm cache, runs in single-digit milliseconds on commodity hardware. Cold compile takes longer because the parser and type system warm up. The bench script in the repo is reproducible.
bench · 50-driverPer-key authentication boundary. Each API key is a logical tenant. Cross-tenant access is architecturally impossible; no shared model store exists that two keys can both read from.
security posture →