The range, and its edge

What makes work fit this

Janus needs particular conditions to work at all. They are worth stating as a test you can run on your own work rather than as a market anyone claims.

  1. 1

    The decision repeats.

    A hypothesis about one person and one model needs more than one case to form, and more than one to be tested.

  2. 2

    A person and a model both touch it.

    Not work the model cannot reach, and not work it already finishes alone.

  3. 3

    The decision carries a cost.

    Getting the split wrong has to be expensive enough to be worth measuring.

  4. 4

    Expert attention is scarce.

    Otherwise nobody minds that a review adds nothing.

  5. 5

    The outcome is eventually adjudicated.

    Someone, or something, establishes what was actually true.

The fifth is the strict one, and it rules out more work than the other four together. Without ground truth a prediction cannot be tested, a hypothesis cannot lose confidence, and the loop is decoration.

It also has a quality, not just a yes or no. In security an incident is often closed asbenign, probably, and the label arrives weeks after the decision. Partial and slow adjudication does not disqualify the work; it sets how fast anything can be learned from it, which is the constraint that matters most here and the one I would want measured first.

Where the conditions hold

Classes of work where all five hold. This is a test being applied, not a list of deployments — there are none.

  • Security incident investigationadjudicated by what the incident turned out to be
  • Fraud and AML alert reviewadjudicated by the investigation outcome
  • Clinical review and triageadjudicated by follow-up, and what the patient turned out to have
  • Credit and underwritingadjudicated by repayment
  • Legal review, relevance and privilegeadjudicated by quality control, and what survives challenge
  • Trust and safety escalationadjudicated by appeal and audit
  • Quality inspectionadjudicated by field failure

The adjudication is the load-bearing half of every line. Where it is missing, the other four conditions do not save the case.

In several of these — clinical, credit, anti-money-laundering — a human checkpoint is a legal requirement rather than an open question. That does not remove the problem: it fixes one of the eight settings in advance and leaves the rest live. It does mean AUTOMATE is not available there, and a page implying otherwise would be selling something that cannot be bought.

A worked instance: security investigation

Security operations are where this is being built first, and the reason is the conditions rather than the category. The decisions repeat, many times a day for one analyst. They matter. Expert attention is scarce and expensive. The agents are improving quickly. And the failures that matter most are rare enough that it is easy to stop watching for them.

Inside an investigation, Janus can

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

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

Being first does not make it the category. The conditions above are what define where Janus applies. Security is where they are met most sharply, and where the first workflow is being built.

Excluded / does not fit

Where this does not fit

Naming the exclusions is cheaper than being found at them.

Work whose outcome is never adjudicated.
Most strategic and creative decisions: you rarely learn what the road not taken would have done. Janus has nothing to test a prediction against there, and should not be sold into the gap.
One-off decisions.
A merger, a reorganisation, a single hire. There is no second case for a hypothesis to be tested on.
Work where the model’s share is negligible, or total.
If the model is not in the decision there is no split to manage. If it finishes alone and finishes correctly, the useful answer is AUTOMATE, and nobody needs a layer to arrive at that.
Work where the expert is not scarce.
A redundant review is then cheap, and cheap problems do not get solved by buying something.

Next

If your work meets the five, the thing worth reading iswhat a first engagement would be.