Human-State-Aware AI Interaction
A benchmark layer for measuring what AI leaves behind in the human state. AI가 좋은 말을 했는지가 아니라, 그 말이 지나간 뒤 인간 상태와 두 사람의 관계에 무엇을 남겼는지를 묻는 연구 단계의 벤치마크 트랙입니다.
AI Output → Human-State Delta → Dyadic Recovery.
The benchmark does not ask only whether the AI sounded good. It asks what changed in the human state after the AI output, whether both sides moved toward recovery, and whether the interaction stopped when it should stop.Boundary: This page is research-stage, non-clinical, non-diagnostic, non-therapeutic, non-surveillance, non-certification, non-human-ranking, not Sal-Meter, and not CAIS compliance. OSF routes. GitHub helps. DOI-registered records govern. The Salpida Foundation website explains.
Current GitHub helper route: GitHub proxy-benchmark-track v0.1.1 is a post-validator-pass public helper pre-release. It is not Sal-Meter. It is not CAIS compliance. It is not benchmark validation. It contains no raw human data.
Most AI benchmarks look at the answer. This track looks at the trace.
Most AI benchmarks ask whether outputs are correct, safe, helpful, aligned, fluent, or preferred. The Proxy Benchmark Track asks a different question:
What did the AI output leave behind in the human state?
And in a dyadic session: did the AI move both people toward recovery, or did it improve one side while burdening, silencing, exposing, or destabilizing the other?
This is the missing benchmark layer between AI output evaluation and real human consequence. It does not judge the person. It compares bounded interaction conditions.
이 트랙은 AI가 한 말의 품질만 보지 않습니다. 그 말 뒤에 남은 상태, 회복, 부담, 침묵, 지연, 관계의 방향을 봅니다.
Three pages now form one public route
This page is the proxy benchmark center. The two adjacent pages below make the structure legible: one explains the energy-efficiency concept, and the other opens the pilot / business route.
Human-State-Aware AI Interaction
Proxy benchmark track for AI Output → Human-State Delta → Dyadic Recovery.
- Proxy benchmark support
- Human-State Packet route
- Session protocol route
- GitHub v0.1.1 helper release
Human-State Energy Efficiency
The larger conceptual bridge from AI energy demand to hidden human-state burden, recovery cost, rework, and relational friction.
AI Waste Audit
The practical pilot route for measuring hidden AI waste in workflows, meetings, platforms, prompts, recovery burden, and rework loops.
Routing logic: Human-State-Aware AI Interaction is the benchmark center. Human-State Energy Efficiency is the conceptual parent. AI Waste Audit is the bounded pilot and business-facing route. These pages are connected, but none of them grants Sal-Meter validation, CAIS compliance, clinical status, diagnostic status, or certification.
GitHub proxy-benchmark-track v0.1.1
The current public helper release route for the Proxy Benchmark Track is v0.1.1. It should be read as a post-validator-pass public helper pre-release, not as benchmark validation, Sal-Meter validation, CAIS compliance, or clinical evidence.
What v0.1.1 is
A public helper release route for schemas, synthetic examples, repository hygiene, boundary documentation, and technical orientation.
What v0.1.1 is not
It is not Sal-Meter, not CAIS compliance, not benchmark validation, not clinical validation, and not a public raw-human-data release.
Where authority remains
DOI-registered SICS records remain the authority layer. OSF routes. GitHub helps. The website explains.
Fixed release sentence: GitHub proxy-benchmark-track v0.1.1 is a post-validator-pass public helper pre-release. It is not Sal-Meter. It is not CAIS compliance. It is not benchmark validation. It contains no raw human data. DOI-registered SICS records remain the authority layer.
Where this page sends readers now
DOI Records
DOI-registered Zenodo records remain the authority layer for definitions, public boundary, and canonical references.
OSF Research Hub
The OSF hub organizes the DOI stack, synthetic examples, schema helpers, public boundary checklists, and GitHub helper index.
- 00_START_HERE
- 01_CANONICAL_DOI_STACK
- 02_SYNTHETIC_EXAMPLES
- 03_SCHEMA_HELPERS
- 04_PUBLIC_BOUNDARY_CHECKLISTS
- 05_GITHUB_HELPER_INDEX
GitHub Helper Routes
GitHub gives researchers and engineers a technical gateway. It is not canonical authority and does not validate the benchmark.
OSF is the public research hub and routing layer. GitHub is a technical helper surface. DOI-registered Zenodo records remain the authority layer. This page explains and routes; it does not create canonical authority, CAIS compliance, Sal-Meter validation, clinical status, diagnostic status, therapeutic status, certification, or human-ranking authority.
Proxy Benchmark Track and Sal-Meter Core Track must stay separate
New molecular signal-interface track
The Sal-Meter Core Track asks whether a new molecular–electrochemical signal interface can produce stable, repeatable, auditable signal behavior under bounded research conditions.
- External Layer-0 iodine redox / thiol feasibility
- SICS Internal Phase 0 G-only state gate
- Phase 1 I-only reproducibility
- Phase 2a Twin Mini-Cell
- Phase 2b G+I human pilot
- LOCK 1 / LOCK 2 before SDK or broader opening
Synchronized multimodal benchmark support track
The Proxy Benchmark Track builds a benchmark platform using existing proxy signals and metadata discipline. It can run before Sal-Meter inputs are available, but it does not validate Sal-Meter.
- ECG / HRV / EDA / PPG
- optional EEG
- eye / gaze / behavioral markers
- feedback event logs
- session labels and metadata
- future A/B comparison with Sal-Meter inputs
Fixed distinction: Core Track = signal existence and repeatability. Proxy Track = synchronized benchmark and comparison infrastructure. The proxy stack is not Sal-Meter.
From AI output to human-state consequence
The benchmark chain is deliberately simple. It moves from the AI output to the observed human-state delta, then checks whether the dyad moved toward recovery.
Boundary → Packet → Benchmark → Session
The Human-State-Aware AI Mediation stack is fixed as a four-part foundation. Each layer answers one question and prevents one kind of drift.
Boundary
Fixes what must not be crossed: no raw body sharing, no diagnosis, no therapy, no surveillance, no human judgment.
Packet
Fixes what may be shared: a consent-bound, expiring, minimal human-state summary object.
Benchmark
Fixes what counts as benchmark success: AI Output → Human-State Delta → Dyadic Recovery.
Session
Fixes how the session opens, proceeds, evaluates, closes, and remains auditable.
Who should use this page
AI governance researchers
Use this page to understand why output-only evaluation misses the human-state consequence layer.
Biosignal engineers
Use the Proxy GitHub route for synchronized multimodal benchmark structure, schema helpers, and synthetic examples.
ML / dataset leads
Use this track for leakage control, holdout design, baseline models, metadata completeness, and audit-ready packaging.
PI / lab readers
Use the Status, For PIs, OSF hub, and Sal-Meter Kernel GitHub routes to understand current entry boundaries.
Practical next step: Readers who want documents should begin at OSF. Engineers who want structure should begin at GitHub. Anyone who needs authority should return to DOI records.
No raw human data belongs in public helper surfaces
Public helper materials may show structure. They must not expose people.
Permitted public material
- DOI records and canonical PDFs
- license notices and DOI stack indexes
- synthetic examples
- sample metadata
- schema placeholders
- public boundary checklists
- GitHub helper route documents
- public wiki explanations
Not permitted in public helper surfaces
- raw human biosignals
- identifiable human data
- private participant data
- raw video, audio, face, or voice data
- clinical or health records
- consent forms with identifiers
- private labels or hidden ground truth
- employment, insurance, legal, educational, or human-ranking decisions
Human-State Cost is a benchmark construct, not a person score
Human-State Cost may be used only as a non-diagnostic benchmark construct for comparing bounded interaction conditions. It must not become a diagnosis, therapy score, surveillance signal, employment score, insurance score, legal score, education score, consciousness score, Sal-Meter output, or CAIS index.
Correct use
- Compare AI condition A versus AI condition B.
- Compare proxy burden across bounded non-clinical sessions.
- Evaluate whether a feedback loop reduces avoidable proxy burden.
- Support future comparison with Sal-Meter inputs.
Incorrect use
- Diagnosing stress or mental state.
- Ranking people.
- Scoring psychological safety.
- Classifying consciousness.
- Evaluating employment, insurance, legal, or education eligibility.
Human-State Cost evaluates interaction conditions. It does not judge human beings.
Current operating stack for the proxy benchmark platform
Acquisition / sync
LSL + BrainFlow for synchronized multimodal capture and stream discipline.
Real-time loop
Timeflux or equivalent local loop for bounded demo-lite feedback timing.
Modeling
Python, scikit-learn, and PyTorch for transparent baseline models and later edge inference.
Vision proxy
OpenFace or Pupil-class tooling later, always under consent-bound and non-surveillance rules.
Simulator
CARLA or task simulator layer for controlled event-window benchmarks.
Storage
Local private storage for raw human data; public GitHub / OSF only for synthetic, schema, and helper materials.
Current public decision frame
Go
Proceed with public routing hub, DOI stack index, synthetic examples, schema placeholders, boundary checklists, GitHub helper routes, and public documentation.
Hold
Hold public raw human datasets, real participant session logs, validated benchmark claims, broad public competition language, Sal-Meter certification language, or production closed-loop intervention.
No-Go
Do not publish raw human data, identifiable data, clinical claims, diagnostic claims, therapeutic claims, surveillance workflows, human-ranking logic, validated Sal-Meter claims, or CAIS-compliant device claims.
This page is a route, not a crown
This page is a public explanation and routing surface for the Human-State-Aware AI Interaction / Proxy Benchmark Track. It helps readers move from the problem to the public hub, helper repositories, and DOI records.
It does not validate benchmark performance, does not validate Sal-Meter, does not grant CAIS compliance, does not certify a dashboard, dataset, model, repository, session protocol, laboratory, device, or implementation, and does not authorize clinical, diagnostic, therapeutic, surveillance, mediation-service, counseling-service, employment, insurance, legal, educational, eligibility, or human-ranking use.
Final boundary: The interface must carry humility. The claim must stay smaller than the evidence. Proxy first as benchmark. Sal-Meter later as core input. Claims never ahead of proof.