Human-State-Aware AI Interaction
A research-stage proxy benchmark track for synchronized multimodal human-state sensing, leakage-safe evaluation, real-time feedback loops, and future comparison against Sal-Meter inputs.
What this track is
The Proxy Benchmark Track is a parallel support track operated to build the software, dataset, metadata, and evaluation layer needed for human-state-aware interaction. It uses existing proxy signals to create a benchmark platform before the Sal-Meter molecular signal is available for comparison.
Sal-Meter signal-interface path
The core track asks whether a new molecular–electrochemical interface can produce stable, repeatable, state-relevant signal structure under the CAIS / Sal-Meter kernel program.
- External Layer-0 iodine redox / thiol feasibility
- Internal G-only and I-only kernel locking
- Twin Mini-Cell separation and leak control
- Future post-lock broader opening only after LOCK 1 / LOCK 2
Human-state benchmark platform
The proxy track does not claim to be Sal-Meter. It builds synchronized multimodal baselines, leakage-safe evaluation rules, and closed-loop demos that can later serve as a comparison lane for Sal-Meter inputs.
- Existing biosignal and behavioral signal capture
- Metadata discipline and event labeling
- Baseline models and holdout-safe evaluation
- Real-time feedback loop for AI, UI, robot, or simulator contexts
This page does not redefine CAIS, Sal-Meter, OE / RE / EE, VCE / CRI / CFI, or any certification boundary. Canonical authority remains fixed in DOI-registered records designated by SICS.
Why it exists
If the Sal-Meter core signal later enters the stack, the first serious question will be: what does it add beyond existing physiological and behavioral proxy signals? This benchmark track prepares that answer before the question arrives.
Build the feedback loop first
Human-state-aware interaction requires synchronized capture, labeling, inference, and feedback. That loop can be built now using proxy signals while the molecular core track continues separately.
Create the comparison baseline
Without a proxy benchmark, future Sal-Meter inputs cannot be cleanly compared against ECG, HRV, EDA, EEG, eye / gaze, and behavioral baselines.
Recruit builders earlier
Software engineers, biosignal engineers, ML engineers, and HCI builders can begin useful work before the Sal-Meter kernel is ready for broader external integration.
Signal families
The proxy stack may combine physiological, behavioral, visual, and interaction-level signals. The purpose is not to diagnose a person. The purpose is to create synchronized benchmark data and controlled interaction baselines.
Reference implementation direction
The initial build should remain local, lightweight, auditable, and replayable. Complexity comes after synchronization, metadata, and leakage control are stable.
LSL + BrainFlow for synchronized device ingestion and timestamp discipline.
Timeflux-style local streaming logic for feedback, event markers, and state windows.
Python, scikit-learn, or PyTorch for baseline models, error analysis, and replayable notebooks.
Webcam + OpenFace-style feature extraction, with optional later Pupil or eye-tracking upgrades.
Local web dashboard, local NAS, versioned metadata, and non-public raw human data storage.
First build targets
The first version should not chase a grand platform. It should prove that synchronized data, labels, metadata, holdout rules, and basic inference can survive contact with real sessions.
Lean benchmark spine
- Two or more synchronized proxy signals
- Session metadata schema v0.1
- Event marker and state-window logging
- Raw / interim / processed folder convention
- Local dashboard draft
Replayable benchmark v0.1
- Holdout split rule and leakage-prevention checklist
- Baseline time-series model
- Error analysis notebook
- Reproducibility pack with sample or synthetic data only
- Closed-loop demo-lite for UI, desktop, simulator, or robot feedback
Raw human data should not be placed in public repositories. Public GitHub materials should use sample data, synthetic data, schemas, documentation, and example code only.
What this is not
The boundary is the strength. The proxy benchmark track must remain useful without borrowing a name it has not earned.
No molecular core claim
ECG, HRV, EEG, EDA, PPG, eye tracking, and behavioral signals do not become Sal-Meter by being combined. They remain proxy benchmark signals.
No diagnosis or therapy
This track does not diagnose, treat, prevent, or clinically interpret any condition. It is a research-stage benchmark support track.
No compliance status
This page does not grant CAIS compliance, Sal-Meter designation, certification status, mark entitlement, or authorized-user recognition.
Who should read this
This page is for builders who can help create the benchmark layer around the Sal-Meter kernel program without confusing proxy infrastructure with the core molecular signal track.
Biosignal / edge engineers
Best fit for people who can connect wearable sensors, stream time-series data, stabilize real-time loops, and document device ingestion.
ML / data engineers
Best fit for people who can design schemas, prevent leakage, build baseline models, manage holdouts, and create reproducible evaluation notebooks.
Human-session operations
Best fit for people who can manage consent, session scheduling, documentation, participant flow, and non-public human-data handling.
Where to go next
Start with the core Sal-Meter boundary, then check the current program status before interpreting this proxy track.