Research
We publish our studies with the code and the raw logs, including the results that did not hold. Every number carries the setup that produced it. Research contracts follow the same rule: fixed scope, a report, the code, the logs.
SeabedNet Churchill Corridor Report: model-completed bathymetry and a shoal-hazard fieldA masked-completion network on the Canadian Hydrographic Service archive. 6.46 million held-out soundings predicted to 12.45 m against 18.08 m for the gravity prior. Sealed forecast, atlas, validation code.
Joules Per Task: energy measurements for learned policies on embedded hardwareEnergy per successful task for a vision-language-action policy on a Jetson Orin Nano, then four policies from 0.2M to 450M parameters on one power rail. Paper draft, harness, on-device logs.
Per-block drafter selection in speculative decoding7,309 blocks, four drafters verified on identical prefixes across two GPUs. Workload-level routing is a null at 0.6%; per-block confidence selection captures 61% of a 13.3% oracle gap at 43% of the cost. Bootstrap intervals throughout.
Hochelaga-8B: Québec French continued pretraining and task fine-tuningFull-parameter continued pretraining on a Québec corpus, task fine-tuning, QFrCoLA evaluation against zero-shot frontier models, and the experiments that did not work. Open weights in BF16 and NVFP4 with an EAGLE-3 drafter.
Three published papers reproduced on a Jetson Orin NanoRisk-aware terrain navigation (IEEE T-IV 2026) at 12.6 ms FP16 and 3.98 ms INT8 at 19.4 W; class-incremental learning across precisions over five seeds; a rover autonomy stack at 446 FPS detection.
Pylox Neuro: a brain-response encoder for videoAn encoder that predicts the functional MRI response of a human brain to a video, built on public recordings and evaluated on concepts held out of training, so the score cannot come from memorising the stimuli. Private weights, in development; research partnerships by arrangement.
Connectome topology as a reservoir substrateFourteen connectomes across four phyla as reservoir computers, with phylum-mean baselines and 10 to 20 seeds per condition. One positive result and two nulls; three apparent wins died under the null models. Write-up in preparation.