Salpida Foundation · Proxy Benchmark Track

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.

Research-stage
Non-diagnostic
Non-therapeutic
Non-clinical
Not Sal-Meter
Benchmark support only

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.

Core Track

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
Proxy Benchmark Track

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
Boundary sentence

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.

Reason 01

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.

Reason 02

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.

Reason 03

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.

ECG HRV EDA PPG Temperature IMU EEG Eye / gaze Webcam markers Keystroke / interaction timing Event markers Simulator state

Reference implementation direction

The initial build should remain local, lightweight, auditable, and replayable. Complexity comes after synchronization, metadata, and leakage control are stable.

Layer 01 Acquisition / sync

LSL + BrainFlow for synchronized device ingestion and timestamp discipline.

Layer 02 Real-time loop

Timeflux-style local streaming logic for feedback, event markers, and state windows.

Layer 03 Modeling

Python, scikit-learn, or PyTorch for baseline models, error analysis, and replayable notebooks.

Layer 04 Vision / interaction

Webcam + OpenFace-style feature extraction, with optional later Pupil or eye-tracking upgrades.

Layer 05 Dashboard / storage

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.

First 30–60 days

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
First 90 days

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
Data boundary

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.

Not Sal-Meter

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.

Not clinical

No diagnosis or therapy

This track does not diagnose, treat, prevent, or clinically interpret any condition. It is a research-stage benchmark support track.

Not certified

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.

PBEE

Biosignal / edge engineers

Best fit for people who can connect wearable sensors, stream time-series data, stabilize real-time loops, and document device ingestion.

MDE

ML / data engineers

Best fit for people who can design schemas, prevent leakage, build baseline models, manage holdouts, and create reproducible evaluation notebooks.

HSOPM

Human-session operations

Best fit for people who can manage consent, session scheduling, documentation, participant flow, and non-public human-data handling.

Public boundary note. This page is a public helper and routing surface for the Human-State-Aware AI Interaction Proxy Benchmark Track. It does not replace DOI-registered canonical records, does not define Sal-Meter, does not grant CAIS compliance, and does not announce that a broad external Sal-Meter competition is currently open. The track is research-stage, non-diagnostic, non-therapeutic, non-clinical, and benchmark-support only.