Reframing the problem
If we only ask: Could a model interpret multimodal data directly?
The more useful question: Can import, alignment, and analysis each be checked before AI helps organize the work? Build inspectable steps first; validity on real data remains to be tested.
The difficulty I wanted to address
Research using several devices produces different formats, clocks, and cleaning requirements. Conversion and alignment come before the actual analysis.
I initiated PsyPhiClaw and worked on its scope, workflow, and core modules to connect some of these repeated steps.
Why separate modules
Import, normalization, analysis, visualization, and reporting have inspectable intermediate outputs. When a result is questionable, the researcher can locate the step and decide what needs repeating.
Input areas include gaze, EEG, physiology, facial expression, fNIRS, and observational coding. Language models help organize work and explanations; processing and calculation need explicit tools and checks.
Built and still unverified
The public code includes 18 behavioral-analysis modules. The main PsyPhiClaw repository contains the overview; module code is in the OpenClaw fork linked on this page.
Validation is paused. Module count describes scope, not accuracy. Further work needs real-data comparisons for formats, timing, statistics, error detection, and review cost against established analysis workflows.
This is a public prototype. It does not yet support a claim that it can be relied on for formal research conclusions.