Janus learns how a professional works with AI over time.
Most AI systems optimize the task in front of them: better answers, faster work, more automation.
Janus also learns the human side of the workflow, because repeated AI assistance can change what people keep doing themselves.
It works with the same professional over time, learns from how they reason with AI, and uses that history to adapt the next interaction.
The aim is not to preserve human involvement for its own sake. It is to keep human judgment active where it still has value — and automate where it does not.
Learn from ordinary work
Janus is designed as a persistent professional counterpart, not a recurring assessment.
A professional can ask a question, test a hypothesis, explain why evidence matters, disagree with the model, revisit a past case, or reason through uncertainty.
Across that history, Janus builds a private, revisable model of how this professional works with AI. It can learn:
Ordinary work produces evidence
- ask a question
- test a hypothesis
- explain why evidence matters
- disagree with the model
- revisit a past case
- reason through uncertainty
- person
- task
- model
- interaction history
- what evidence the person considers sufficient;
- how they express uncertainty;
- how well confidence tracks outcomes;
- what tends to change their mind;
- where they rely quickly on AI;
- which model errors they tend to catch;
- how they respond to challenge, guidance, or disagreement.
Each finding stays tied to context: a particular person, task, model, and interaction history.
Its value is practical: can it choose a better way to work with this person next?
Track what changes over time
AI can take over work where human involvement no longer adds value. That is often exactly what should happen.
The harder case is when the organization still depends on human judgment for rare failures, ambiguous cases, or higher-order decisions.
Then today's performance is not the only question.
Repeated cases
- Case 01
- Case 02
- Case 03
- Case 04
- Case 05
- reliance changes
- calibration shifts
- still catches model misses
- review stops adding useful value
Are we still exercising the human judgment we depend on?
Janus keeps that trajectory in view: how reliance changes, how calibration shifts, where the person still catches things the model misses, and where review has stopped adding useful value.
The distinction is between useful delegation and quietly losing a human backstop the workflow still expects to have.
Change how much of the thinking AI does
Janus uses that history to adapt the next interaction.
Depending on the person, task, recent history, and current conditions, it can:
Not a sequence
- give the answer directly;
- let AI draft and ask the person to verify;
- show the evidence before the conclusion;
- ask the person to form a hypothesis first;
- challenge the reasoning or ask for an alternative;
- use a case as deliberate practice;
- remove scaffolding when it is no longer useful;
- add support when the person is overloaded;
- debrief after the outcome is known;
- recommend full automation where human involvement no longer earns its cost.
Janus can ask for more independent thinking, make the work easier, or remove the human step entirely.
The division of cognitive work is something to adapt, not set once for a role.
Learn from what happens next
Janus keeps a history of what it believed, how it adapted the interaction, and what happened next.
Belief, expectation, outcome, revision
- Hypothesis / v1
- Expectation
- Adapted interaction
- Later case / outcome
- Confirm / contradict / unresolved
- Confirm
- Contradict
- Unresolved
- Revised model / v2
A useful hypothesis should imply something about what Janus expects to see after a different way of working together. Later cases and outcomes can confirm that expectation, contradict it, or leave it unresolved.
As that evidence accumulates, earlier hypotheses can strengthen, weaken, or be replaced.
The result is a continually revised model of where this professional and this AI work well together, where they do not, and what kind of assistance helps next.
Keep the deeper model private
The professional-facing relationship can be richer than the information the organization needs.
Janus separates the private adaptation model from the organization-facing operational layer.
Inside Janus, the response layer can work from bounded behavioral guidance rather than raw internal labels, scores, or model mechanics.
The private side can hold detailed hypotheses about reasoning, calibration, reliance, blind spots, and responses to different kinds of assistance.
The organization-facing side can receive the operational output it needs: whether an independent check is required, another review or escalation is warranted, a review step is no longer useful, more automation is appropriate, and how the workflow is changing overall.
Private adaptation model, bounded guidance, operational output
- reasoning
- calibration
- reliance
- blind spots
- responses to assistance
- whether an independent check is required
- another review or escalation is warranted
- a review step is no longer useful
- more automation is appropriate
- how the workflow is changing overall
The organization does not need access to the internal model Janus uses to adapt to the professional.
That separation also protects the interaction itself. If ordinary reasoning with Janus becomes a performance dossier, people have a reason to stop reasoning openly with it.
The longitudinal idea
A static workflow asks:
Who should do this task?
Janus asks:
Who should do which part of the thinking now — given what has happened between this person and this AI so far?
The answer can change.
So Janus keeps learning.