Janus

Human×AI decision layer

The more work you hand to AI, the less you know what your people are still adding.

Janus sits between your experts and the AI tools they work with. It learns where a particular person’s judgment changes the outcome and where it does not — then changes how the work is done: what the AI handles alone, who decides first, what gets shown before a verdict, when a second review earns its cost.

The point is a better joint decision on the next case — not a payoff after a training period. The model of each expert makes it sharper over time.

Read the technical brief Contact

First product — Adaptive Security Investigation

The Janus horizon A drawn instrument circle divided by a tilted line: the machine field above, the human field below, and a point on the line where the decision sits.

AI arrived faster than the work could be redrawn around it.

A copilot here, an agent there, human review on top of both. The reasonable assumption is that this gets tidied up the way every rollout gets tidied up: ship it, watch where it jams, fix it next quarter.

That assumption is where this goes wrong.

Where the line between human and machine judgment sits, decision after decision A horizontal band spanning a sequence of decisions. The upper part of the band is the share of the decision carried by the AI, the lower part the share carried by the person. The dividing line between them is not level and not fixed: it rises and falls across the sequence, and the points where it changes are marked with the interaction policy that moved it — human first, evidence only, challenge, automate. Which side of the line a given decision falls on is not reported by ordinary tooling. The line rises, drops almost to nothing where a step is automated, and comes back up again: it moves in both directions. Decisions over time The line moves. Almost nothing reports where it is. carried by the model carried by the person HUMAN_FIRST AUTOMATE EVIDENCE_ONLY Each mark is a change to how the next decision is run. Automating a step is one of the moves, not the destination.
Fig. 1 The line between what a person decides and what the model decides moves constantly. Almost nothing in the stack reports where it currently is.

Deploying the AI changes the thing you are trying to optimise.

With ordinary software you can ship a mediocre process and improve it against a stable target. Here the target moves as you work on it.

The model arrives. People work differently around it. Their decisions change, and so do their habits. The value of a review changes with them. The right division of work shifts. By the time you have measured it, it has shifted again. Then the model is upgraded and some of what you learned no longer applies.

This is why “we will optimise it iteratively” fails here specifically. Iteration assumes the thing you are measuring holds still between measurements.

Two ways to be wrong.

Under-automate

You keep paying scarce experts to re-check work the model already had right. The productivity gain the AI was bought for never fully arrives.

Over-automate

You remove judgment that was doing real work. The failures that follow are rare, which is what makes them expensive: you stop watching for them, so they surface late — and they surface where you were counting on someone to catch them.

The expensive part is not choosing wrong. It is not knowing which of the two you are currently doing. The bigger the AI’s share of the work, the harder that is to see.

A decision layer between people and the AI they work with.

Not another assistant, and not a dashboard. Janus belongs in the path of the work — an investigation console, an analyst’s workspace, a review queue — not in a separate tab. It governs how the person and the model reach a decision together.

It decides what the person sees, in what order, and whether they are asked to commit to a view before the model shows its own. It is above any single model: swap the underlying agent and the pairing changes, without the history being thrown away.

The eight it chooses between

  • HUMAN_FIRST
  • AI_FIRST
  • EVIDENCE_ONLY
  • HIDE_AI_VERDICT
  • CHALLENGE
  • SCAFFOLD
  • SECOND_REVIEW
  • AUTOMATE

Each gauge shows who carries the decision — mostly the person at one end, the model alone at the other. What each one does is set out in the brief.

“A person agreed with the AI” is four different events.

That agreement can mean:

  1. the person had independently reached the same conclusion;
  2. the AI corrected a mistake they were about to make;
  3. the AI pulled them off an answer they already had right;
  4. they did not think about it at all and clicked approve.

From outside — from a log, from a screen recording, from an agent watching the session — all four look identical. Telling them apart is not an observation problem. It is an instrumentation problem, and it is where Janus does its work.

  1. 01 — Instrument the decision

    When the answer matters, capture the person’s own judgment and confidence before the model speaks. Sometimes show the evidence and withhold the verdict. Sometimes let the model lead. Conditions vary deliberately, so their effect can be read.

  2. 02 — Hold hypotheses, not notes

    Not “this analyst over-trusts the model”. Something testable: on this task class, with this model, shown in this order, this person stops catching a specific kind of error. Each carries a version number, its evidence, and its uncertainty.

  3. 03 — Predict, then check against the outcome

    Every hypothesis implies a prediction about what a different configuration would do. The adjudicated outcome tests it. A hypothesis that fails loses confidence or dies — Janus is structurally able to be wrong about a person, and to be shown so.

  4. 04 — Change the next interaction

    The hypotheses that survive choose how the next decision is run.

What is modelled, and what is not

The model is of professional decision-making, not of a person: how someone forms a hypothesis, how well their confidence tracks their accuracy, which of the model’s errors they catch, what kind of help makes them better and what turns them into an approval step. Not their personality, and not a score. What the organization sees is a separate, operational projection — the split is set out in the brief.

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

Janus does one thing: it re-decides what the human role should be, case by case.

Automation, independent judgment and capability are not three products. They are what follows from getting that one decision right.

Where the person adds nothing
The step is automated, and the automation case is made from evidence rather than from nerve.
Where independent judgment is genuinely needed
Janus records it before the model speaks, so what arrives is judgment rather than approval.
Where the organization will still depend on that judgment later
Janus can keep that judgment exercised instead of letting the workflow retire it — putting the person back in the reasoning rather than at the end of it.

Only the first two are the reason to buy Janus. The third is what keeps the first two true a year from now.

The longitudinal core exists. The enterprise result does not yet.

Built today

A deployed longitudinal inference system: state that persists across sessions, versioned hypotheses and predictions, the full revision history, provenance and audit, and state transitions you can inspect rather than take on faith. TypeScript and Node on PostgreSQL with pgvector.

Building now

A task-bounded model of Human×AI complementarity, and the first enterprise workflow built on it: Adaptive Security Investigation.

Not claimed

No enterprise outcome has been validated with external users. And we do not yet know whether rich longitudinal Janus beats strong, simple personalization.

That is the next test. It is not a result we are pretending to have already.

Know what your people are still adding.

Janus is the layer that learns where human judgment improves an AI-assisted decision — and where it should get out of the way.

Read the technical brief Get in touch