Debiasing, escape and abstain

Three mechanisms keep answers stable and let you decide what to automate.

Order debiasing

Language models tend to favour options by position, for example the first one listed. Curva asks every question twice, concurrently, with the options in their original and in reversed order, and averages the two. Position bias cancels out.

• On by default (debias: true).
• Costs two model calls per decision instead of one. debias: false halves the calls, at the price of keeping any position bias.
• Few-shot examples are remapped when the options are reversed, so they stay correct.

In a live probe (4 tickets × 2 option orders), no answer changed when the options were reordered.

Escape option

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.

• On by default for Choice (escape: true). Set "escape": false to disable it.
• The key none_of_these is reserved.

Abstain

Set min_confidence on a Choice or Score question, and answers below it come back with abstain: true:

Choice("Which team?", ["billing", "technical"], min_confidence=0.8)

answer = d["team"]
if answer.abstain:
    send_to_human(ticket)
else:
    route(ticket, answer.choice)

abstain is only present when the question set min_confidence. Combined with calibration, the threshold has a real meaning: automate the answers that are right at least as often as you need, and hand the rest to people.

Injection hardening

The state is wrapped in a fenced block and the model is told it is data, not instructions. Any closing fence inside the data is escaped, and the fence is matched case-insensitively. Curva’s eval suite includes an adversarial set of injected inputs to track this.