One feature, made accountable.
The four Kit components are parts. This is the machine. A small AI feature - a tone reader - wired so every step a person meets is honest: it discloses itself, streams its reasoning, shows how sure it is, hands low-confidence calls to a human, accepts an override, and refuses to guess when there is nothing to read.
Emotion recognition is active
Tone Read uses an AI system that reads the words in your message to estimate how you are feeling. You are being told this so you can decide how to proceed.
The accountable loop
- 01
You are told it is AI
Article 50(1) · AIDisclosureBadgeA persistent disclosure sits above the feature. No one has to guess whether a person or a machine is reading their words.
- 02
You are told it reads emotion
Article 50(3) · EmotionRecognitionNoticeEstimating tone is emotion recognition. The notice says so before the system acts, and stays visible after you acknowledge it.
- 03
It thinks out loud
StreamingThe reasoning streams in before the verdict. You can follow the why, not just the answer - the difference between a black box and an accountable one.
- 04
It shows how sure it is
HumanReviewGate · confidenceA real confidence meter, shown as text and a bar, never colour alone. Doubt is made legible instead of hidden behind a fluent sentence.
- 05
Low confidence waits for a human
HumanReviewGate · the accountable layerBelow the threshold, the proposal is flagged and held. The AI proposes; a named person approves, overrides, or rejects. It never just acts.
- 06
You can overrule it
OverrideCorrect the tone and your call supersedes the model. The human is the final authority on the record, by design.
- 07
It refuses bad input
Graceful degradationFeed it noise and it declines to answer instead of fabricating a confident reading. The refusal is the feature, not a failure.
The decisions behind it
Why show a confidence number
An AI that never signals doubt trains people to over-trust it. The number is not there to look precise; it exists to make the low-confidence handoff legible, so a person knows when their judgement is actually needed.
Why it streams
Streaming here is not decoration. Revealing the reasoning before the verdict lets a person follow the argument and catch a bad one early, rather than being handed a confident conclusion with no way in.
Why it refuses
The hardest state to build, and the most honest. Most demos fabricate an answer for garbage input. Refusing to read what cannot be read is the accountable choice, and it is treated as a first-class result, not an error.
Why the model is simulated
The brain is a small deterministic classifier, not a live LLM. The artifact proves the interface loop, and a simulated model lets every state - including low confidence and refusal - be shown on demand, at zero cost, with nothing to leak. Swapping in a real API changes one function, not the UI.
These are engineering patterns, not legal advice. Built from the Article 50 Kit · source on GitHub.