# Part III. Decodable is not wearable | The Intent Layer Source: https://setu-labs.co/manifesto-3.html > The incumbent is not the keyboard. It is the human body, which is fast, reliable, free and already installed. [Download PDF ](https://setu-labs.co/assets/setu-the-intent-layer.pdf) ## Decodable is not wearable The incumbent is not the keyboard. It is the human body, which is fast, reliable, free and already installed. A new interface competes with the hand, the voice and the glance. They need no charging, no calibration and no terms of service. The average person types about 50 words a minute[[1](https://setu-labs.co/manifesto-sources.html#s1)] and speaks about 160.[[8](https://setu-labs.co/manifesto-sources.html#s8)] Across seventeen languages, fast talkers and slow ones converge on the same rate of information.[[2](https://setu-labs.co/manifesto-sources.html#s2)] No language is the default, and an interface tuned to one is unfinished. ### Four gates Between a signal and a product stand four tests, passed in order. 1. **Information.** Is the content physically present in what the sensor measures? 2. **Reliability.** Does decoding survive movement, sweat, hair, fatigue and a sensor that sits two millimetres to the left of yesterday? 3. **Utility.** Do accuracy, delay, setup, comfort and price add up to something better than the alternative at hand? 4. **Agency.** Can the person switch it on and off, see what was inferred, and reject it? Part II was a tour of the first gate. Products are made or lost at the other three. #### Figure 5 (interactive): The journey of a signal Decoding is a chain, and the weakest link is usually the first. Degrade the signal and watch what an honest decoder should do, then let a language model paper over the gap. Pattern the body produced · illustrative What the sensor recorded After artifact removal and filtering Signal quality at the sensor Clean (close to the source) Noisy (far from it) Let a language model fill the gaps Decoder output Move Thursday’s review to Friday Supported by signal The cleaned trace still resembles the source. The decoder proposes one reading; the user confirms. 1. Sensor and synchronisation 2. Artifact removal 3. Learned representation 4. Task decoder 5. Language prior 6. Uncertainty estimate 7. Personal calibration 8. Deliberate intent gate 9. Editable output and feedback ***Synthetic demonstration.** All traces are generated in your browser from sine bursts and pseudo-random noise. They are not recordings and the slider is not calibrated to any device or measured signal-to-noise ratio. The three output states are scripted to illustrate a principle: when the signal no longer supports a reading, a fluent sentence produced by the language model is *supplied*, not *measured*. * ### Two clocks A software agent can read a thousand pages while you read this sentence. Then it stops and asks a question, and the clock that matters changes. Its clock runs in milliseconds. Yours runs in the time it takes to understand, decide, find words and get them out. The pressure this creates is easy to predict. When the human turn is the slowest step, there is a standing incentive to remove it. People are rarely pushed out of the loop by decree. They are pushed out by impatience: the approval that becomes a rubber stamp, the review that is skipped because the queue is long. > Whatever makes a human turn cheaper makes it likelier to survive. Better prompts, better dashboards and more autonomous agents all work around the bottleneck. Only an interface that picks up intent closer to its origin attacks it. Most of what is proposed for keeping AI in check works by constraining the machine. BCI is one of the very few levers that works the other way, by raising what the human can do. #### Figure 6 (interactive): Two clocks, one task An agent finishes its part quickly and waits. The human turn is four activities, and a faster input channel shrinks only one of them. **AgentHuman** - **Agent working - **Comprehend the question - **Decide - **Formulate a reply - **Transmit it - **Waiting With today’s channel, the agent is idle for most of the task. Transmission is a visible share of the human turn, but not the largest. ***Illustrative proportions, not measurements.** Segment widths are chosen to show structure and are not drawn from data; real ratios vary enormously by task and person. The point is structural: comprehension and decision remain when transmission shrinks. The one published throughput estimate cited in the text[[3](https://setu-labs.co/manifesto-sources.html#s3)] is not plotted because it does not decompose into these stages.* ### Different thresholds, different lives For someone who communicates through an eye-tracker at a few words a minute, thirty words a minute with errors changes what a day contains. The people in the trials above accepted brain surgery for that chance. They are who this science exists for, and they are owed finished technology, not prototypes. For someone with working hands, thirty words a minute is a downgrade. A consumer interface will not win on speed first. It will win on silence, privacy, free hands and access. And the bar is a cliff, not a slope. Speech recognition that is right nineteen times in twenty is a toy. Right nine hundred and ninety-nine times in a thousand, it is everywhere. A field can look stalled for years and then tip. #### Figure 7 (interactive): The same error rate, four verdicts Pick a hypothetical decoder. Its value changes completely with who is using it and what a mistake costs. Hypothetical: about 1 word in 4 wrong ##### Restoring communication A person who cannot speak or type Already meaningful If the alternative is a few words a minute, an imperfect fast channel with easy correction is a large gain. ##### Silent everyday text Anyone, in public or with hands busy Not worth wearing Fixing every fourth word costs more than thumb-typing. Novelty fades in days. ##### Directing an agent or robot Operator giving short commands Only with a small command set Works if the vocabulary is constrained and each command is confirmed before it runs. ##### High-stakes control Payments, vehicles, medical devices Unacceptable An error is an action. No amount of convenience offsets it. Hypothetical: about 1 word in 10 wrong ##### Restoring communication A person who cannot speak or type Conversational Listeners repair one error in ten without effort, as they do with any accent or noisy line. ##### Silent everyday text Anyone, in public or with hands busy Useful in niches Worth it when speaking is impossible or unwise and the message is short. Not a keyboard replacement. ##### Directing an agent or robot Operator giving short commands Useful with confirmation A capable agent can ask about the ambiguous word instead of guessing. ##### High-stakes control Payments, vehicles, medical devices Still unacceptable alone Usable only as a proposal that a second, independent step must confirm. Hypothetical: about 1 word in 50 wrong ##### Restoring communication A person who cannot speak or type Close to natural The remaining limits are speed, fatigue, expressiveness and upkeep, not accuracy. ##### Silent everyday text Anyone, in public or with hands busy Depends on everything else Accuracy is no longer the obstacle. Comfort, setup time, battery, price and trust decide. ##### Directing an agent or robot Operator giving short commands Strong Low-effort approvals and corrections become practical at the pace agents work. ##### High-stakes control Payments, vehicles, medical devices With confirmation, perhaps The error rate of the whole loop, including the confirm step, is what must be certified. ***Author’s judgment, not data.** The three error levels are hypothetical and are not claims about any device or about “where the field is”. Verdicts are qualitative assessments that assume a workable correction mechanism; they are offered to show that usefulness is a surface over need, burden and consequence, with no single threshold.*