@one-moment/core

The decision layer of a voice agent that waits.

This is the engine behind One Moment, a voice agent whose job is to make the other person wait, so someone with aphasia can make their own phone call. It decides one thing: given what two differently configured listening streams heard, should the agent speak for the caller, ask them a question, or keep waiting?

It is pure. No network, no audio, no I/O of any kind except an optional model client you pass in. That is what makes it testable, and it is why every number the project publishes can be re-derived from committed files.

npm install @one-moment/core

Requires Node 24, which runs TypeScript natively. The package ships source rather than a build step, and uses erasable syntax only.

The idea

Most voice agents are optimised to respond faster. This one is designed to know when not to respond, and never to say a word the caller did not say.

import { assembleEvidence, runDissent } from '@one-moment/core';

// What your patient stream heard, and optionally a second, faster one.
const evidence = assembleEvidence({ patient: turn, fastFinals: [fastTurn] });

const result = await runDissent(evidence, {
  apiKey: process.env.ASSEMBLYAI_API_KEY,
  callerName: 'Robert',
  lexicon: [{ term: 'amlodipine', category: 'medication' }],
});

result.decision.action; // 'relay' | 'ask' | 'hold'
result.decision.text;   // exactly what to say, when the action is relay
result.decision.policyRule; // which rule decided, and why

How it decides

An ordered list of rules, checked in order, first match wins. The Adjudicator is code, not a model.

1. Deterministic checks first. If the two streams disagree about a "not", it asks, offering the two readings the ears actually heard. No model involved.
2. Verbatim when possible. A complete, clear sentence is relayed as the caller's own words. Zero model calls, so zero chance of an invented word.
3. Dissent for fragments. An Advocate proposes what was meant; a Skeptic reads the same evidence independently. If they disagree, or the proposal contains a word the caller never said, the caller is asked rather than guessed for.
4. Never trust a half-found word. A word from the caller's own vocabulary that only the biased ear heard, or that only partly came out, becomes a two-way choice.
5. Forced choice, never yes or no. In aphasia the default answer may be "yes", so the caller gets two real options.

What is in here

| | | |---|---| | assembleEvidence | Both readings, word confidence, hidden pauses, where the ears disagree | | runDissent | Advocate and Skeptic, then the Adjudicator | | stepFloor, initialFloor | The Floor Controller: who may speak, and when to hold the line | | buildQuestion, negationChoice | Forced choices built from the caller's own words | | callRecord, recordAsText | What was said in the caller's name, to keep | | patientConfig, fastConfig | The two AssemblyAI stream configurations, and why each value is what it is | | VOICES, isVoice | The voice ids the Voice Agent API accepts, checked |

Measured

Every figure is re-derived from committed run files by npm run verify in the repository.

• 120 sentences from 8 speakers with dysarthria (TORGO): an ordinary agent put words in the speaker's mouth in 65 of 120; this engine in 11 of the 57 it chose to speak.
• Ordinary turn settings split 48 of those sentences mid-way. The patient configuration split 0. Repeated through a real 8kHz telephone encoder: 46 against 0.
• 60 sentences of fluent speech (LibriSpeech): it relays 49 of 60, waits a 1.7 second median, and splits none, so the patience does not get in the way of speakers who do not need it.
• Across 48 relays of degraded speech it added 0 words the speaker never said, where a single model asked what the person meant added 11.

Honest limits

Not tested with people who have aphasia. Not a medical device. It does not improve recognition of disordered speech and assumes recognition is often wrong; the point of the rules is what gets said anyway.

MIT licensed. Built for the AssemblyAI Voice Agent Hackathon, September 2026.