# Feasibility tests

> Decided designs in this repository rest on assumptions nobody has run. Each test below answers one question that gates one decision. They are research, not product: the code is throwaway and does not belong in `conver...

Canonical: https://chattic.us/wiki/feasibility-tests/

Decided designs in this repository rest on assumptions nobody has run.
Each test below answers one question that gates one decision. They are
research, not product: the code is throwaway and does not belong in
`conversation/src`.

Run these before building on the decisions they gate.

## How to use this file

- Each test states its question, the assumption under test, what it
  gates, a method, and what to do when it fails. A test without a written
  failure branch is not ready to run.
- When a test has run, record the result in the document it gates and
  delete the test from this file. This file holds open questions only.
- Spike code is disposable. Do not productionize it, do not add it to the
  quality gates, do not let it become the implementation.

## What is not a feasibility test

The thin end-to-end turn -- human message, DynamoDB, model call, streamed
chunks, one committed message, no computer anywhere -- is the first build
step, not research. It runs on the computerless-worker path from
challenge 5 and is the smallest thing that is actually Chatticus.

Do it using challenge 1's transport result in
[Design challenges](/wiki/design-challenges). Do not let it absorb the
spikes: a measurement that becomes a feature is a measurement nobody trusts.

## Voice: on-device listening in the browser

The spike (`chatticus-a2000f`) ran tests 1 and 2 from
[Voice](/wiki/voice#feasibility-tests) headless, and test 4 only partially
(12 synthetic utterances). The results are in
[Voice](/wiki/voice#spike-results). These parts are still open, each with the
failure branch written in that document:

- A live microphone in a real browser for 60 minutes, foreground and
  background, with battery drain measured.
- iOS Safari on a device, where the question is memory: about 650 MB above a
  blank page.
- Cross-origin isolation (COOP `same-origin` plus COEP `require-corp`)
  against the Google and Cognito sign-in and the Vultus avatar.
- Test 4 as specified: about 50 real recordings, false-addressing rates,
  and phonetic name matching.