How Dual AI works in Interview Signal

Dual AI sends the same follow-up and scoring requests to two AI models at once. In Interview Signal, follow-up suggestions then appear in two columns, A and B, each headed by the model that answered, and the scorecard shows the second model's score beside each area, flagged with "The two models disagree — probe this area" when the two are three or more points apart. The scorecard itself stays one scorecard, from the first model. Dual AI is off by default and can be turned on in Settings or during a call. This describes version 0.9.10, current as of October 2026.

How do I turn on Dual AI?

1. Open Settings and find the "AI during interviews" section.
2. Choose the first model under Model for follow-ups and scoring.
3. Tick Dual AI — run a second model alongside.
4. Choose the Second model.

During a call, the AI button in the live screen's top bar opens the same choices: AI model, Dual AI — a second model alongside and Second model. With Dual AI on, the button reads, for example, "Dual AI: Cerebras + OpenAI". Changes apply from the next follow-up or score and are remembered for later interviews.

Which models can I choose?

These are the options in version 0.9.10; the providers are listed on the security page. The second model cannot be the same as the first. Pick the same one and the app says "Pick a different model for B — the same model twice is one opinion." If you have not chosen one, the second model is OpenAI, or Cerebras when the first is OpenAI.

What changes on screen with Dual AI on?

• Follow-ups appear in two columns, A and B, each headed by the model that actually answered. Each column fills as soon as its own model replies, so a slow second model never holds back the first. Suggestions in B that repeat one in A are removed.
• If B's request fell back to the same model as A, column B reads "Fell back to the same model as A, so there is nothing different to show." A column whose model did not reply says "This model did not answer this time." or "No connection."
• The scorecard shows the second model's score beside each area, as "B" and a number; rest the pointer on it to see which model gave it and its one-line assessment. When the two scores are three or more points apart, the area shows "The two models disagree — probe this area".
• The answer check combines both readings, shows the stronger, and adds "both models" when both flagged the answer. See the answer check.

The second model's quotes are held to the same rule as the first's: any quote not found word for word in what the candidate said is dropped.

Which score goes into the submittal?

One scorecard goes forward: the first model's reading, verified against the transcript, or the second model's when the first did not answer. Your own scores, set with the + on any area, override either. The fit score and the client submittal are built from that scorecard. The second model's scores are saved with the interview on your computer so you can see them during the call, but they are not written into the submittal. See the scorecard with verified quotes and the client submittal.

When is Dual AI worth turning on?

• When a score will be challenged: a final-round shortlist, or a client who questions submittals. A disagreement flag shows you which area to probe before the call ends, not after.
• When you are scoring outside your own field and want a second opinion on the areas you cannot judge yourself.
• When you want more follow-up options: two models reading the same answer often reach different questions.

It is less useful on short logistics screens, where there is little to disagree about. Disagreement between models is a prompt to ask another question, not a vote; our guides to scoring interviews consistently and calibration sessions cover the human side of the same problem.

Does Dual AI cost more?

Not in minutes. Your free minutes or bought hours are counted on transcription time, not on AI requests, so Dual AI does not use them faster (see pricing). The app's own note says it "uses twice the model calls", which is the real cost: two requests for every follow-up round and every score.

What Dual AI does not do

• It does not touch the kit. Building questions from a job description uses the model the app picks for that job, not your Dual AI choice.
• Ask the AI and Suggest 3 use only the first model.
• The accuracy check uses only the first model, since two readings of the same claim would be a debate, not a check.
• It does not average the two scores or let the second model change the first.