Pick one from a list
Ask a question with a fixed set of answers, up to 26 of them. You get back the one it chose and how far ahead it was. Handy for sorting, routing, matching and triage.
See it decide an access request →
Capabilities
What Piko can decide
Every closed question has a shape. Piko supports exactly three, and returns the reading that tells you whether to act on the answer.
Answer true or false
A yes or no question you write yourself, or one with no wording at all. It answers with the chance the statement is true, and stays quiet when the two sides are too close.
See it judge whether a move is fatal →Give a score on a scale
Ask for a grade from 1 to 5 and get a single number that already averages every grade it weighed. Even a hesitant grader still gives you something to act on.
See it grade how urgent a case is →Say nothing when it is unsure
One dial sets how far ahead the winner has to be before it answers. Raise it and it answers less often but is right more often. Staying silent is a valid answer here, not a failure.
See it abstain on a hard call →Show the gap, not just the answer
Every answer comes with the distance between the top two choices. A wide gap means go ahead. A narrow one is your cue to look at it yourself.
See the gap on two similar names →Understands pictures and sound
The same question can be asked about an image or a recording. Show it a photo or a sound and it answers from what it sees or hears, not from a caption someone wrote.
See it read an image →It cannot invent an option
It can only answer with one of the choices you wrote down. It has no way to produce a name, a number or a plan you did not offer, which is what makes the answer safe to automate on.
See it choose where a piece lands →Uses the machine you already have
Point it at the model server you already run, the same one answering your chats. Nothing leaves your machine, and every answer can be re-checked later without it.
See a decision come off your own server →Same answer on another machine
Run one question on two machines and a small share of answers move. Above a high confidence almost none do, so that is the part worth storing, not the exact number.
See a question with a stable answer →Ask many questions at once
Send one situation and up to 26 questions. Each one comes back with its own answer and its own confidence, in a single request.
See three questions on one screen →
How it works
One letter holds the whole answer
Ask for a single letter instead of a paragraph. That is the whole trick — and it is why the answer can be measured instead of believed.
01
You write the choices down
You give it the situation and the options, labelled A. B.
C. It is never asked to invent an answer — only to point at one of yours.
02
It answers with one letter
One short request, one letter back. Alongside it comes the odds the machine gave every option, which is the part that would normally be thrown away.
03
You get the answer and its confidence
The chosen option, the gap to the runner-up, and a clear signal when the machine was not sure enough to answer at all. Nothing is written down on a refusal.
TEXT · IMAGE · SOUND
One model reads all three
The same question can be asked about a paragraph, a photograph or a recording. The model answers from what it sees or hears, not from a caption someone wrote. Here is what each kind of request costs, measured on a single card.
Text
A document, a rendered board, a conversation: anything already written down.
54msper decision
€0.000013 per decision
Image
A photograph, a screenshot, a diagram. Read as raw pixels, never as alt text.
84msper decision
€0.000022 per decision
Sound
A recording of speech, music or an ambient scene, answered from the waveform.
68msper decision
€0.000013 per decision
One model, gemma-4-12b, serves all three modalities on a single rented card at €0.35/hour. A decision reads a single output token, so the cost is the prompt: an image adds about 270 tokens over plain text and roughly 30 ms, a recording about 106 tokens and 14 ms. The price quoted already carries a 30% margin.
Speed
The difference is one letter, not one paragraph
An ordinary model writes its answer out word by word, and you wait for the whole thing before you can trust it. Piko reads a single letter it was going to produce anyway — and the reading is done the moment that letter lands.
Same question · same machine
one real reading, replayed
Piko
—
—
One letter, read once. The gap to the runner-up comes with it.
A writing model
—
The same answer, spelled out. The words keep coming after the
decision was already made.
Use cases
Where the reading earns its keep
Not free-form writing, and nothing whose answers you cannot write down yourself. Where a single letter says nothing, Piko refuses to pretend otherwise.
pacman · recorded run
Gaming
A game it plays without seeing the board
The program works out every legal move and describes each one in a sentence; the machine points at one. When it is not sure enough, the game asks again instead of guessing.
Open the Pacman example →
tetris · recorded run
Gaming
A piece placed by one question
The code lists every placement the piece could take and what each one would cost. The machine points at one, and the reading says whether the choice was close.
Open the Tetris example →
robot arm · recorded run
Agents
An agent that says how sure it is at every step
One question picks the next action from a list; two more track whether it is still on goal and whether it is stuck. The same machine also writes the words to type.
Open the agent example →
drone · recorded run
Robotics
A drone that flies itself through a course
The whole vehicle and the whole map go in as one question, and the next move comes back from a short menu. A safety rule written in code always outranks the model.
Open the drone example →
browser agent · real run
Agents
A browser that picks its next move from the page
One observation becomes a table of what is on the screen; the machine chooses the operation and the element to act on, and only the typing is left to words. The same request carries the question and the reading.
Open the browser example →
jevpilot · live drive
Autopilot
An autopilot that steers and says how sure it is
JevPilot is a real drive, deciding several times a second: the road, the traffic and the lane it is in go in as one question, and a steering and a speed come back from a short menu — each with the gap that says whether the call was close.
See a tick-by-tick decision →
Piko API
One endpoint. Typed answers. Nothing to install on the model.
Send one situation and as many questions as you like, and get an answer, a confidence and a refusal back for each — or point the client you already use straight at it.