Our long-term direction

上海确知智能科技有限公司 拟登记名称

尚未完成核名及注册 · 起步上海,面向世界

For AI that
understands
people.

Make human state a first-class input to AI.

Our long-term aim is to connect signals, evidence, and uncertainty so intelligent systems can respond to human state—and have their responses checked.

Today Human State Evidence Graph v0.1 · in progress

Human, interpreted.
Conceptual artwork

01 / The human dimensionHuman × intelligence

A person is more
than a prompt.

Attention. Effort.
Affect. Context.

These belong in the conversation about intelligent systems.

We aim to build the foundations for AI that can account for human state, adapt within clear boundaries, and verify whether its response helped.

Understanding begins with measurement. Responsibility begins with knowing where the evidence stops.

Human signals concept animation

02 / Our first foundation

Current work

Start with
the evidence.

Human State Evidence Graph

Our first foundation: a versioned structure linking constructs, signals, measurements, evidence, and inference boundaries.

v0.1 · in progress

The schema is settling. This is research infrastructure in development.

About the evidence graph
Human State Evidence Graphv0.1 · in progress
Five connected evidence objects Concept schematic connecting construct, signal, measurement, evidence, and inference boundary. The graph is in progress. SignalWhat we observe ConstructWhat we infer EvidenceWhat supports it Inference boundaryWhere it stops MeasurementHow we connect them
CONCEPT · v0.1 DIRECTIONDeclared. Sourced. Bounded.

What does a signal justify—and where does the inference stop?

03 / Research foundation

All projects + publications

Research,
built to compound.

Our aim is to turn research into reusable evidence, measurement protocols, models, evaluations, and software.

Selected studies / Conceptual cover artwork

01 / Ground

EviDial

A benchmark that separates agreement with depression symptom labels from the textual evidence available to support an assessment.

Submitted · ICLR 2027

A status, not a result.

BoundaryEvidence support, not diagnostic validity. Candidate references and an exploratory development audit do not establish model rankings.

02 / Verify

ANI

Affective Non-Interference studies how input construction, agent behavior, and scoring affect conclusions about emotional expression at authorization boundaries.

Submitted · ICLR

BoundaryConditional findings from synthetic tasks and saved records; clean affect-induced expansion of an unchanged authorized action set remains undemonstrated.

03 / Measure

I'm Fine

A study of how LLM self-reports and a fixed observer differ in score levels, model ordering, and temporal variation across adapted affective instruments.

Submitted · ICLR

BoundaryDescriptive results from archived model interactions. No clinical, subjective-experience, or single-mechanism causal conclusions.

04 / How we think

Our principles

Measure the state.
Respect the person.

Human-defined outcomes guide the system. A signal can support an inference; it never defines a person's value.

Long-term development loop

  1. 01

    Sense

    Name the signal and its source.

  2. 02

    Model

    State the construct being inferred.

  3. 03

    Calibrate

    Preserve context and uncertainty.

  4. 04

    Adapt

    Respond within the declared boundary.

  5. 05

    Verify

    Check the human-defined outcome.

05 / People + progress

Research + engineering

One question.
A long horizon.

AI Certainty is a founder-led research and product initiative.

Starting in Shanghai, with an international outlook.

Junjie Xu

Founder · Multimodal affective computing
and generative systems

14 first-author papers among 27 publications. The initiative is currently one person.

Personal website About the founder

Where we are today

The Human State Evidence Graph is in progress. Our next step is to make its schema and evidence records a stable reference.

06 / Collaborate

Build the
human dimension.

Join the work on evidence,
measurement, and evaluation.

Discuss collaboration