Learned binary-reduction fast Horner v38
Status: all mechanical submission gates pass; organizer manual rules review remains required.
V38 changes only v37's bounded-reduction candidate search. Instead of evaluating 31 quotient candidates, a 157-parameter learned controller performs five comparison decisions and emits one final quotient candidate. The existing learned MAC, subtraction, judge, Horner Teacher, raw reducer, representation, and output path are unchanged.
- Parameters: 268,299
- Checkpoint: 1,104,963 bytes
- Learned binary-controller decisions: 157/157 exact
- Every transferred v37 tensor: bit-identical
- Fresh raw Tier 1--10 GPU validation: 200/200 exact
- Checkpoint SHA-256:
2544f4f3951856dea18c6f7d0f5563b3b9bae8e089ffe297e3474361b2038330
Full official 4-CPU sandbox result: 1,000/1,000 scored Tier 1--10 cases, with every tier 100/100, in 86.1 inference seconds. Manifest, size, static analysis, load, preprocessing isolation, determinism, strict reload, and learned-weight dependence gates pass.
The competition permits trained models that learn algorithm-like circuits but prohibits arithmetic algorithms hand-coded in tensor operations. V38's controller and arithmetic maps are trained, all capabilities collapse after weight randomization, and its training provenance is included. Its recurrent search wiring is nevertheless explicit, so final eligibility requires the organizers' manual provenance/code review rather than a local assertion.