None of the above: an escape option for LLM classifiers

A "none of the above" option gives an LLM classifier an honest way to say that no label fits. Without it, a model shown an input that fits none of your options must still pick one, and it often picks with high confidence. Curva adds this option to every Choice question by default, under the reserved key none_of_these. You route it to a person the same way you route an unsure answer. This post covers why forced choice fails, how the escape option works, when to turn it off, and the one route where it is off by default.

#Why an LLM needs none of the above: a ticket that fits no team still gets a team

A Choice question asks the model to pick exactly one option. If your options are billing, technical and sales, and a ticket asks about a partnership, every answer is wrong. The model doesn't know that "none" is allowed, so it picks the nearest option, and the probabilities it returns are spread over three wrong answers. Often one of them still comes out confident.

That is the dangerous part. Every downstream rule trusts the confidence. A partnership enquiry routed to sales with 0.9 confidence passes an "automate above 0.8" check and lands in the wrong queue. Nothing errors.

Calibration does not fix this on its own, because the right answer is not on the list. A calibrator can make the model less sure of its wrong pick, but it can't produce an answer that doesn't exist. The fix has to be in the question.

#none_of_these is on by default for Choice

Curva's docs describe the escape option plainly: a Choice question gets an extra option, none_of_these, by default. Without it, a model shown a ticket that fits no team has to pick one anyway, often with high confidence. With it, the honest answer is available, and you can route it to a human.

In practice:

• It is added to every Choice unless you turn it off. In Python, escape defaults to on when options is given.
• It appears in probabilities like any other option, and choice may be none_of_these.
• It counts toward the probabilities the model splits, so a ticket that partly fits shows that doubt across your options and the escape option.
• In TypeScript, the answer type includes it: d.answers.team.choice is a union of your option keys plus "none_of_these", so code that forgets the case is easier to spot.

In the n8n node, the Route operation gives none_of_these its own output branch when **None of These** is on, which is the default.

#escape: false, and why the key is reserved

Turn it off when "none" is already one of your real answers, or when an answer outside the list is impossible by construction. Over HTTP, set "escape": false on the question. In Python, the escape argument of Choice does the same.

The built-in recipes show both cases. The content-moderation recipe turns the escape option off because allowed content has its own category, none, defined as ordinary content. Every post gets one of the eight categories, and "nothing wrong" is a real answer rather than a fallback. Other recipes do the same for questions whose options already cover every case: lead-qualification for its timeline, which has an unknown option for "not stated", llm-output-qa for its verdict, and rag-check for its next step.

The key none_of_these is reserved. You can't use it as one of your own option keys, because Curva needs it to mean exactly one thing: the model said no option fits.

#Route none_of_these and abstain to the same queue

There are two different kinds of "not sure", and both belong with a person:

• **none_of_these:** the model says none of your options fits.
• **abstain: true:** the model picked an option, but its confidence is below the question's min_confidence.

Route both the same way:

if answer.abstain or answer.choice == "none_of_these":
    send_to_human(ticket, decision_id=d.id)

The routing order matters in tools that branch step by step. The docs' order is: abstain first, then the chosen option, where none_of_these also goes to a human. Checking abstain first stops an unsure answer from being routed as if it were sure.

Look at what lands in the escape branch over time. A steady stream of none_of_these for the same kind of input is a sign that your option list is missing a category. Add it, and settle the new wording before you collect labels, because a reworded question starts its calibration over.

#On the /v1/systemone alias it is off unless escape: true

Curva also serves POST /v1/systemone, a compatible alias of /v1/decide for clients that send questions in the criteria shape. On that route a Choice has no escape option unless the question sets "escape": true. The reason is compatibility: those clients expect answers to stay within the criteria they sent.

The Python SDK follows the same rule for the compatible style. Choice(..., criteria={...}) has no escape option unless you pass escape=True, while Choice(..., options) gets one by default.

If you migrate a criteria-style client, decide deliberately. Keep the old behaviour for a like-for-like comparison, then turn the escape option on once your code can route none_of_these.

#Escape vs Multi: no label fits vs several labels fit

The escape option answers one specific problem: no label fits. A different problem is that several labels fit at once, like a ticket that is both a refund request and a bug report. That needs a Multi question, not an escape option.

A Multi scores each option as its own yes/no in the same call, with independent probabilities that don't sum to 1. If no option reaches the threshold (0.5 by default), selected is empty, which is the Multi's own way of saying none apply.

#Next steps

The docs cover the escape option in debiasing, escape and abstain. For the other "not sure" exit, read LLM confidence threshold and abstain. For why forced choices tend to come out overconfident, see LLM overconfidence, and for the full picture, LLM classification confidence scores you can act on. Install with pip install curva-ai.