Automated research demo · pending scientific review
Human State Evidence Graph
A versioned, machine-readable evidence graph that answers one question: given observable human signals, what inference about a human state is scientifically justified, under what conditions, and with what uncertainty? Five object types — construct, signal, measurement, evidence, inference boundary — linked so every claim carries its provenance, and every boundary is explicit.
Explore evidence & 2D/3D graph → 中文
Research demonstrations
Separate demonstrations with different purposes and evidence collections; these are research prototypes.
- Browse evidence & 2D/3D graph →70 records · 24 sources · keyword and concept browsing
- Question-to-evidence demo →1 saved question · 6 records · 4 papers · a separate collection, not live model output
- Human signals concept animation →Synthetic illustration of signals and context · no live sensing
What it is
-
An evidence graph whose primary unit is the claim, not the paper. Studies attach to claims as evidence records, each with provenance back to its source.
-
Versioned like software: the schema, the corpus, and the confidence rubric each carry a version.
-
Contradictions and null results as first-class data — preserved, never averaged away.
-
An auditable confidence rubric: labels from
INSUFFICIENTtoVERY_HIGH, plusCONFLICTED, computed by stated rules rather than judgment calls. -
“Insufficient evidence” is a normal output, printed at full size.
What it is not
-
It never returns a verdict about a person. It returns the status of the evidence.
-
Not a diagnostic instrument. Population-level associations are not individual-level results.
-
The public explorer is an automated research demonstration. Scientific release approval remains pending.
-
The research corpus remains in progress. Automated checks are audit aids, and records remain subject to scientific review.
-
Not a substitute for the literature. Every record points back to its source.
CONCEPT · v0.1 DIRECTION
The query, walked through
Take one concrete question — the kind the graph exists to answer — and walk it through the five object types. This walkthrough is illustrative: a concept of the intended behavior, not a live query. The records behind it are v0.1, in progress.
task-evoked pupil dilation observed during a task — what inference about cognitive load does the evidence support, and where does it stop?
- Status
- EVIDENCE · INSUFFICIENT — few independent studies, small samples
- Confounders
- CONFOUNDED_BY · screen luminance · task difficulty
- Calibration
-
Group-level, task-specific, with uncertainty attached.
- Boundary
-
What may not be inferred: a verdict about an individual; a causal claim from correlational designs; anything covert. The pupil is not a lie detector.
Claims vocabulary — every status on this site uses five tiers
- Step 1 · Signal
The observation enters as a signal: pupil diameter (
SIG_PUPIL), in the eye-tracking modality (MOD_EYE). The graph separates the feature — task-evoked pupillary dilation (FEAT_PUP_DIL) — from baseline diameter, because a baseline and an evoked response are different measurements. - Step 2 · Construct
The feature points at a construct — cognitive load (
CON0030) — over the edgeoperationalized_by. The construct, not the signal, is where the claim lives. The ontology keeps cognitive load distinct from task difficulty: one is a state of the person, the other a property of the task. - Step 3 · Evidence
The edge carries claim clusters, and each cluster carries evidence records: one study's contribution, with provenance back to its source, its population, its task, and its sample. Where studies are few and small, the rubric returns
INSUFFICIENT— not as a failure, but as the answer. - Step 4 · Confounders
The same dilation answers to other constructs. Arousal moves it; surprise moves it; screen luminance moves it before anything else. Each confounder is a node in the graph, typed
confounded_by— not a footnote. - Step 5 · Boundary
The construct carries an inference-boundary card: what may be inferred (group-level, task-specific, with uncertainty attached) and what may not (a verdict about an individual; a causal claim from correlational designs; anything covert). The pupil is not a lie detector.
Group-level support here is thin, the signal's specificity is documented as limited, and the boundary travels with the answer.
Input modalities
The modality layer of the ontology — the signal families the graph can carry evidence about:
- Eye tracking and pupillometry MOD_EYE
- EEG MOD_EEG
- Behavior and interaction logs MOD_BEHAVIOR
- Peripheral physiology and wearables — heart rate, HRV, EDA, PPG MOD_PHYSIO · MOD_PPG
- Self-report MOD_SELFR
- fMRI MOD_FMRI
- Audio MOD_AUDIO
- Video MOD_VIDEO
Initial domains under exploration
Nine state families form the v0.1 scope — listed as initial domains under exploration, not as validated coverage:
- 1 Cognitive load and mental workload
- 2 Sustained attention and vigilance
- 3 Mental fatigue
- 4 Sleepiness
- 5 Stress
- 6 Affect — valence and arousal
- 7 Trust
- 8 Motivation and reward states
- 9 Mind wandering and task engagement
Current work
The ontology also carries adjacent, more sensitive constructs. Their boundaries are deliberately
conservative, and some states are excluded from inference entirely at v0.1.
Status
Current work
Status: v0.1, in progress — Rev. 2026-09. The schema is settling; claim clusters and
inference-boundary cards are being assembled for the initial domains. Under this site's claims
vocabulary, everything on this page is CURRENT WORK: a status is not a result.
What v0.1 will ship
-
A frozen schema, documented in the open, with every node type and edge predicate defined.
-
Claim clusters for the initial domains, each carrying its evidence records, its provenance, and its rubric label.
-
An inference-boundary card per construct: what may, and may not, be inferred.
-
A public changelog — the graph versions like software, and the changelog is where the versions live.
-
Contradictions and null results printed at the same size as positive findings.
No dates are attached. The roadmap is direction, not schedule.