Evidence

Evidence has to stay attached to the live system underneath it.

This page keeps the public story tied to actual cohorts, governance, support structures, model audits, and evidence-linked reporting.

Model evidence

The evidence base now includes how the engine behaves.

A route is more credible when the system can explain what supports it, what weakens it, and what kind of signal the model is actually using.

Stability evidence

Routes are checked for validation strength, missingness sensitivity, participant-level error spread, interval behavior, and feature-space complexity.

Dependency evidence

Grouped ablation separates wearable sleep, activity, physiology, self report context, portable domains, and quality signals so the narrative stays attached to what the model uses.

Confidence evidence

Outputs can carry a research confidence score that lowers trust when quality, validation, uncertainty, portability, or dependency patterns are weak.

Evidence Ledger

The public evidence story should read like a traceable ledger.

Instead of presenting evidence as a pile of claims, this layer shows what each source is doing inside the research stack and where it is allowed to support interpretation.

Source Open cohorts

Used to train, stress-test and compare route behavior before any public-facing interpretation is trusted.

active
Mechanism Route audits

Stability checks, dependency checks and confidence scoring explain why a route is promoted, watched, capped or held back.

review
Boundary Evidence attachment

Each output keeps source, quality status, claim level, uncertainty and audience context attached to the visible result.

bounded
Next step External validation

Future clinical or diagnostic ambition would require prospective data, clinical ground truth, ethics and independent validation.

future

Filter

Read the evidence base by lane.

This view helps a public reader isolate the strongest stress evidence, the weaker non-stress anchors, or the explicitly portable pieces without having to parse every row at once.

Anchor map

Each visible layer rests on a different evidence base.

The point is not to pretend that all layers are equally mature. The point is to show what anchors each one: open cohorts, support cohorts, transport probes, governance, and evidence linkage.

Night Layer Open cohorts, route governance, held-out evaluation
open cohortsheld-outroute governance
Event Layer Portable LOCO, challenge cohorts, event-specific governance
portable LOCOchallenge cohortsevent governance
Daily Challenge Layer Daily self-report cohorts, temporal challenge probes
daily self-reporttemporal probeschallenge layer
Non-Stress Daily Layer Support cohorts, same-instrument anchors, daily probes
support cohortssame-instrumentdaily support
Non-Stress Portable Families Family transport probes, support cohorts, outcome mapping
outcome mappingsupport cardstransport probes

Live attachment

The evidence base should stay attached to a visible signal.

Each row ties evidence posture back to a representative visible signal and its anchor tokens, so the page does not drift into abstract background material.

Night Layer Open cohorts, route governance, held-out evaluation
Representative visible signal Novartis D0 fatigue contextual interpretation
open cohortsheld-outroute governance
Held-out R² 0.4350
Event Layer Portable LOCO, challenge cohorts, event-specific governance
Representative visible signal Portable family stress event
portable LOCOchallenge cohortsevent governance
R² 0.3164
Daily Challenge Layer Daily self-report cohorts, temporal challenge probes
Representative visible signal Boston anxiety D0 temporal
daily self-reporttemporal probeschallenge layer
R² 0.1685
Non-Stress Daily Layer Support cohorts, same-instrument anchors, daily probes
Representative visible signal Novartis depression D0 temporal
support cohortssame-instrumentdaily support
R² 0.3454
Non-Stress Portable Families Family transport probes, support cohorts, outcome mapping
Representative visible signal Anxiety
outcome mappingsupport cardstransport probes
Portable mean R² 0.0202

Foundation

The public surface is downstream of a deeper validation stack.

These blocks keep the site aligned with the real system instead of letting it drift into a generic health narrative.

Open cohorts

The surface rests on open nightly, daily, and event cohorts, plus support cohorts used to keep anxiety, depression, and physical pain anchors scientifically legible.

nightdailyevent

Governance

Every layer preserves its own accepted, experimental, or support status. Nothing rises to the top level just because it looks good in one cohort or one benchmark.

tiersheld-outgates

Evidence linkage

The public surface sits on top of evidence-linked reporting, route cards, support cards, and transport probes rather than free-form interpretation.

reportscardstransport probes

Recent Research Direction

The non-stress picture changed in a specific way.

The mixed depression family is still blocked by instrument mismatch, but the strict `PHQ-9 only` depression subtrack is now materially cleaner. Anxiety remains the strongest non-stress portable family overall, and somatic burden remains support-heavy rather than transport-ready.

Depression

  • Mixed familyStill blocked by questionnaire mismatch across cohorts.
  • PHQ-9 only clean subtrackPortable mean R² `0.0168`, replay `0.0469`, now the correct research direction for depression.
  • StatusResearch-only, not promoted into route governance.

Somatic burden

  • Direct anchor`EPISONO pain 2016` keeps physical pain scientifically anchored.
  • Family postureStill `proxy_only` until transport is shown explicitly.
  • InterpretationUseful as somatic burden support, not as clean portable pain.