The record / 6 reports
Every finding leaves the lab as an Ommny Research Report: numbered, citable, and preregistered before the runs begin. Reports are revised in place, with changelogs, the way honest datasets are.
The full account of the Ommny O-1 architecture: a single transformer trained jointly on sight, sound, touch, taste, smell, and language. We describe channel encoders, the shared 128k context, and why joint anchoring outperforms late fusion on 9 of 11 perceptual benchmarks, at 41 percent of the inference cost.
Cite as: Ommny Research Report ORR-26-04 (2026)
Anchors are the pairs that bind a sensation to its description: this texture and the word coarse, this chord and the word dusk. We introduce the anchoring scheduler used to train O-1 on 2.4 trillion tokens and show how anchor density, not raw token count, predicts downstream perceptual accuracy.
Cite as: Ommny Research Report ORR-26-02 (2026)
Force curves, slip telemetry, and motor audio, fused inside a language model, let a gripper reason about fragility the way it reasons about weight. On an open-world pick-and-place suite, grounded touch reduced damage rates by a factor of 7.4 against vision-only baselines.
Cite as: Ommny Research Report ORR-25-11 (2025)
The two senses that mainstream multimodal models ignore are also the two with the weakest naming conventions. We map volatile compound signatures and taste spectra into the O-1 token space, and report the first public benchmarks where a model describes a smell more precisely than a median human panel.
Cite as: Ommny Research Report ORR-25-06 (2025)
A model that perceives the world is a model with consequences. We propose refusal behavior grounded in physical risk, provenance obligations for sensory data, and an evaluation protocol for embodied systems that treats hallucinated perception as a safety event, not a quality bug.
Cite as: Ommny Research Report ORR-25-02 (2025)
The founding document. Why we stopped building six datasets and started building one, how consent-cleared capture works in practice, and the versioning discipline that lets a sensory dataset be cited, audited, and reproduced like code.
Cite as: Ommny Research Report ORR-24-08 (2024)
Reports are issued as capacity allows, usually four to six a year. Preregistrations open to audit partners under the Corpus Charter.
Request corpus access