The perception layer for Physical AI
The most valuable knowledge on Earth was never written down. We build world models that perceive it.
↑ drag to resolve
01 · Thesis
Every lab protocol, every assembly manual, every standard operating procedure is a compression of something richer: the trained hands and calibrated eyes of a person who knows how. The pressure of a pipette draw. The angle that avoids aerosolizing a sample. The half-second hesitation that means a step went wrong. None of it survives translation into text. And text is all we have been feeding our machines.
This is why robots that ace language benchmarks still fumble physical procedures. The knowledge they need was never in the training data, because it was never in words at all. Philosopher Michael Polanyi named this six decades ago: tacit knowledge: the kind we demonstrably possess but cannot articulate.
Our bet: the path to Physical AI does not run through better descriptions. It runs through world models: systems that learn the structure of physical procedures by watching them, predicting them, and representing what language cannot encode. Not tacit to explicit. Tacit to tangible.
02 · Research
Laboratories are the sharpest test bed for tacit knowledge: procedures are precisely specified in writing, yet outcomes depend on what the writing never captures. Consider a reflux that must reach steady state: actively boiling versus settled. Same equipment, same label; only the motion of the fluid distinguishes them, and a frontier VLM reading sampled frames cannot tell them apart. LabProc is our six-task benchmark structured along exactly this axis, from language-amenable state recognition to motion-only discrimination. Tacit, our 305M-parameter domain-adapted V-JEPA-2.1 encoder, is the vision-only baseline we release alongside it.
ICML 2026 · AI for Science Workshop
LabProc and Tacit: Quantifying the Visual–Textual Prior Gap in Autonomous Laboratory Perception
Across a structural axis from language-amenable to motion-only tasks, a frontier VLM leads by 41 points where its pretraining carries the signal. The gradient reverses on the hardest motion-only subset, where Tacit leads by 9.1 points despite a ~1000× parameter asymmetry. Where pretraining provides the signal, pretraining wins; where motion-state is the only distinguishing feature, dense-temporal video representations win.
03 · The Model
Tacit, a 305M-parameter domain-adapted V-JEPA-2.1 video encoder, watches physical procedures and learns their structure by representing them, not by generating captions. That distinction is the whole company. A captioning model tells you what is happening. Dense-temporal video representations capture what distinguishes this execution from the one that fails: the part that never made it into the protocol.
04 · Vision
Drug discovery runs on physical procedure, and on the scarce experts who execute it well. As labs automate, a new question becomes existential: how do you verify that an autonomous system actually performed the procedure correctly? That verification layer is where we begin.
Near term
Verification and qualification for automated laboratories: giving CROs and CDMOs an independent perception layer that watches procedures and certifies execution, accelerating the design-make-test-analyze cycle.
Long arc
Every domain where expertise lives in hands rather than documents (manufacturing, materials, medicine) needs machines that perceive procedure. We are building the models, benchmarks, and methodology to be that layer.
05 · Why we exist
Tacit World was founded by Jagrati Gogia after eleven years building ML and data systems inside high-scale operations, where the distance between what the documentation said and what skilled people actually did was, again and again, where the value hid.
Tacit World · Delaware C-Corp · Bengaluru → US
06 · Contact
Partnerships, pilots, research collaboration, or the ICML poster session. Reach out.
jagrati@tacitworld.ai