You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
Calyx is a working GPU-first (CPU-capable) Rust engine (~205k LOC, 19 crates, MCP server, ~2,830 tests) that operationalizes one idea: meaning is grounded association, and intelligence is measuring and composing those associations. Each input is viewed through many frozen 'instruments' (lenses); Calyx derives the associations between them and uses mutual information to quantify, in bits, whether a panel carries enough information about a real outcome to be trusted. A data-processing-inequality bound makes this falsifiable — you can refute an architecture before training any model.
GOALS & HOW WE'LL ACHIEVE THEM
(1) Ship an open-source 1.0 with a reproducible 'bits / panel-sufficiency' test that third parties can run. (2) Demonstrate the test correctly predicting a system's success or failure before training. (3) Release an open library of diverse frozen lenses others can extend for their own languages and domains. Achieved by hardening the existing engine, building the benchmark suite, and commissioning the lens library.
HOW FUNDING WILL BE USED
Minimum ($5k): open-source 1.0 + the reproducible bits-test. Goal ($50k): adds the open lens library and benchmark suite. Toward $100k: full-time work, compute (onnx-int8 / model2vec / candle-fp16), and broader lens commissioning. No overhead; milestone-structured, so even the minimum yields a usable public artifact.
TEAM & TRACK RECORD
Christopher Royse — independent researcher/engineer; author of two 2026 papers ('The Calculus of Association' and 'The Oracle and the Kernel') and sole builder of Calyx. Shipped, measured results: a voice reproduced at 0.961 WavLM speaker-similarity; a style model robust under prompt injection with zero-shot transfer to Golden-Age Spanish; and a predicted clean negative on SWE-bench (panel carries ~0.46 of the ~1 bit a verdict needs) — the falsifiable test working as designed. Three production panels run today (video, text/code, civic matching).
MOST LIKELY CAUSES / OUTCOMES IF IT FAILS
Main risk: making the panel-sufficiency test reproducible and general across domains proves harder than expected, or adoption is slow. Even then the public goods — the open engine, the falsifiable bits-test, and the lens library — remain usable, and negative results are themselves informative; the framework is built to produce cheap, early falsification.
FUNDS RAISED IN THE LAST 12 MONTHS
None committed to date for Calyx. Concurrently applying to Sentient Foundation, RAAIS, and Foresight Institute (AI for Science & Safety Nodes); no funding secured yet.