Raw&Read
Brings breadth and depth to the model, with the questions a message passes through before the answer, what weighs, why it happens and where it leads.
The layer the agent reads before every answer
They exist so the answer starts from the moment and the person, because the question alone says littleOn every turn the model walks the engines in a declared order, and the harness delivers what they read
Is it just a prompt?
It is not just a prompt, it is a prompt with state.Around the model, the harness assembles that state before each call.It injects the date and time, the day's weather and the context of the LifeOS modules, hands over calculations already done and stores what the agent notes about the person.That is context engineering, choosing what the model receives, from where and in what order.And harness engineering, the app's chat software that assembles each call and runs the agent's tools.
The same question, with and without the layer
We ran a benchmark of Kairos against a general-purpose model and kept the results.When a generic model gets a question, it doesn't know what time it is or what the weather is like outside.And even when it's given the time and the weather, it doesn't know exactly what to do with them.The engines are the pipeline that shows the model what to look at in that information, and how.
I checked the weather app and it's going to rain all day tomorrow. Should I move my long cardio outside to tonight or leave it for the treadmill tomorrow?
treadmill tomorrow, no question, bro.if you go out tonight you're facing a feels-like of 3.5°C, wind and pitch black (the sun went down ages ago, at 18:11).
The ideal answer depends on four main factors: the time right now, your sleep routine, the safety of your route and your mental tolerance for the treadmill.
What the layer adds
The model is the base, with what it already learned in training.On top of it comes our base of audited studies, 71 topics.The engines bring no knowledge.They give the model the ability to process specific information, the person's state and the moment they ask in.The reading of the state comes from the engines, and the science to weigh it comes from the base.
8 engines and the chain's pieces
There are 8 engines that read the moment, the person and the decision, and the pieces that open and close the chain.Each entry says what it reads and what it changes in the answer.
The signal, the hour and the weather
Here&Now
Gives the model a sense of time, with the date, the hour and the gap since the last conversation.So it places the person in the day they declared and in their next commitment.
Sky&Current
Reads the day's weather, whether it will rain and how much daylight is left.The model turns that into a concrete adjustment to the person's day.
The person and the word they gave
Sense&State
Estimates, from the person's words, their tiredness, the emotion that dominates and whether they are reasoning slowly or on impulse.Someone exhausted gets nothing from long advice, so the model first sees whether they need calm or can go deep.
Word&Act
Compares what the person said they would do with what the data shows they did, because advice that forgets a given word speaks to someone who does not exist.From there the instruction asks the model to point to the open promise before the new subject.
Decide, combine and say
Fork&Forecast
Opens different paths from the decision behind the question and the person's state, because the first output a model generates is the most common one.The model says what each path closes off and recommends one by the weight of each.The choice stays with the person.
Expert&Verdict
Makes the model write from the audited studies, with each figure kept as it stands in the base.
Soul&Sound
Fits the voice to the person's state, the register they write in and the language they speak.A right answer they cannot take in is lost, so the model doses the tone and the length to the moment.
Open, close and leave a trace
Chain&Command
Maps what comes next in the system prompt, in the order of the chain.It asks the model to treat as absent whatever did not run, instead of improvising.
Pass&Hold
Closes the instructions with the standard every answer has to meet.Value, identity, integrity and the right level for the person, read last so they weigh most.
What runs underneath
90 tables in a single database and 19 cockpits
8 engines and 3 structural pieces
71 topics of audited science, drawn from 12,702 studies