Janus

Technical brief

From human-in-the-loop to longitudinal human assurance

Human review is valuable when it adds independent judgment. Janus learns when it does.

01 The double bind

Three ways to put AI into consequential work, and what each one costs.

Review everything
Scarce experts stay the bottleneck. The verification tax survives the automation.
Automate everything
No independent backstop. Rare model failures pass unchecked.
Nominal human-in-the-loop
The review cost stays. The checkpoint may add no new information.

02 The loop

  1. 01 — Observe

    Pre-AI judgment and confidence. The AI recommendation. Whether the decision moved. How the professional responded to the intervention. The adjudicated outcome.

  2. 02 — Hypothesize

    Versioned, task-bounded hypotheses about this human × this model × this task class × this history, each with provenance and an uncertainty.

  3. 03 — Predict

    Which configuration should work better here?

  4. 04 — Intervene

    One of eight control states is selected for the next decision.

  5. 05 — Measure and revise

    Decision quality. Appropriate reliance. AI-error detection. Review time. The hypothesis is updated before the next case.

Janus treats its model of the human as hypotheses to test, not a profile to believe.

03 The eight control states

Selected per decision, revised after the outcome.

  • 01 HUMAN_FIRST Human an independent judgment is recorded before the AI conclusion is shown
  • 02 AI_FIRST AI the model leads and the person adjudicates
  • 03 EVIDENCE_ONLY Human the evidence is shown and the verdict is not
  • 04 HIDE_AI_VERDICT Human the conclusion is withheld until a judgment is committed
  • 05 CHALLENGE Both a counterargument is required before the case closes
  • 06 SCAFFOLD Both the reasoning is supported step by step rather than replaced
  • 07 SECOND_REVIEW Human ×2 a second pair of eyes is worth what it costs on this case
  • 08 AUTOMATE AI alone the human step is not changing the outcome and is removed

Each gauge shows who carries the decision. State 08 is not a fallback: a system that can only add human steps has an interest in keeping them, and would be the wrong system to ask.

04 Private reasoning, public assurance

For Janus to learn anything true, a professional has to be able to say I am not sure without the hesitation turning into a performance record. So the depth and the exposure are separated by design, not by policy.

Stays with the professional

  • reasoning history
  • confidence calibration
  • where they disagree with the AI
  • recurring error classes
  • how they respond to challenge or scaffolding
  • what happened after past interventions

Goes to the organization

  • independent judgment obtained
  • required review completed
  • material disagreement remains
  • second review required
  • escalation required
  • aggregate workflow and outcome measures

The organization never receives a dossier on anyone. It receives evidence that a human checkpoint did real work at the moment the system depended on it.

05 Adaptive Security Investigation

The first product.

Security operations are where this gets tested first. The decisions repeat. They matter. Expert attention is scarce and expensive, the AI agents are improving quickly, and the failures that matter most are rare enough to be easy to stop watching for.

Inside an investigation, Janus can

  • hold the AI’s conclusion back until the analyst has committed to one of their own;
  • show the evidence without the verdict;
  • ask for the counterargument, or route the case to a second pair of eyes;
  • let the agent finish alone when human review is not changing the outcome.

Same AI. Same incident. Different analyst — and possibly a different right way to run it.

06 What is already built

The longitudinal machinery is not a design document. It runs today.

  • state that persists per user across sessions;
  • evidence → versioned hypothesis → prediction → revision;
  • a full revision history with provenance and audit;
  • current state that demonstrably changes later system behaviour;
  • evidence, state transitions and hypothesis history open to inspection;
  • a deployed TypeScript and Node runtime on PostgreSQL with pgvector.

What that proves is bounded, and the boundary matters: it proves the inference architecture exists. It does not prove enterprise outcomes.

07 What is being built now

  • task-bounded professional signals in place of the prototype’s domain-specific inference;
  • capture of pre-AI judgment, confidence, AI influence, task and model context, intervention response, and adjudicated outcome;
  • longitudinal state connected to workflow policy;
  • repeated cases instrumented for outcome-linked evaluation;
  • a pilot-ready Adaptive Security Investigation workflow.

08 Where Janus sits

An operational AI platform supplies the world: context, agents, actions, permissions, decision and outcome lineage, evaluation harnesses.

Janus supplies something those systems do not hold — persistent Human×AI state, versioned hypotheses, longitudinal prediction and revision, and the policy selection that follows from them.

Operational platform

world state → agents → actions → permissions → outcomes

Janus

Human×AI history → versioned hypotheses → prediction / revision → interaction policy

The two compose. Janus is built to sit alongside that infrastructure and travel with the workflow from one system to the next.

09 How we intend to falsify it

Four arms, in increasing order of what they assume.

  1. AGeneric AI assistance.
  2. BStrong universal interaction design.
  3. CStrong shallow personalization.
  4. DJanus longitudinal state.

The comparison that decides the product is D against C. If rich longitudinal state does not materially beat strong shallow personalization, the honest response is to simplify the product, not to defend the architecture.

10 Claim boundary

What we can show now

The engineering reality of longitudinal state and inference. Persistence, versioning and provenance. A mechanism you can open and inspect.

What we still have to prove

That these signals are learnable fast enough in real professional work. That rich longitudinal state earns its complexity over shallow personalization. That the result is a measurable decision and economic benefit. That repeated adaptation preserves human capability over time.

We are not asking anyone to believe our model of the human is correct. Janus makes explicit predictions about a Human×AI pair, changes the workflow on the strength of them, and uses the outcome to test the prediction and revise the model.

Where this problem comes from

Automation did not remove the pilot. It created a discipline.

When autopilots arrived it stopped being enough to build a good aircraft and train a good pilot separately. The system between them had to be engineered: what the automation does alone, what the pilot sees, when it hands back, what counts as a meaningful intervention, and whether someone kept in reserve is still able to act when the reserve is called.

Intellectual work is arriving at the same problem, one layer up: what is being automated now is parts of thinking. That layer is what Janus makes adaptive and specific to a person, instead of a fixed policy written once for everyone.

Whether the people a system keeps in reserve are still able to act is not currently reported by anything. Making that visible is what Janus is for. Whether Janus improves it is the claim boundary above, not a result we have.

Who is building it

Egor Chirkunov — Founder & Builder

I came to this from directing and interactive narrative, by way of product. At G5 Entertainment I built and led a narrative function of around fifty people. For several years after that I worked out the methodology behind Janus, and since early 2026 I have been building the system itself, hands-on, with AI-assisted development tools.

Based in Valencia, Spain.