Two threads, one engine
AITIA is where two Causis research programs converge.
Causis built two systems from first principles. CR‑Causal learns how a system actually works — how it changes when you act on it — so it can tell cause from coincidence, imagine the what‑if, and design its own experiments. CR‑Stream is the substrate it runs on: it learns from an endless stream of events, keeping only compact summaries, so its memory stays flat and it never has to retrain. AITIA is the two as one.
The cause‑and‑effect scenes below are CR‑Causal; the memory‑and‑merge scenes are CR‑Stream. Everything you see is one engine.
The law crystallizes from noise
Give it messy measurements; it hands back the equation — and grows surer the more it sees.
No equations are given — only noisy measurements over time. Watch a rough first guess sharpen into the exact law that generated them.
After eight short runs the error against the true law is ~0.001 — the engine has read dx/dt = +1.50·x − 1.00·xy (a predator–prey rule) straight out of the noise.
Correlation, or cause? Only acting can tell
It separates real causes from coincidences — what correlation alone can never do.
A hidden common cause makes X and Y rise and fall together. In one world X truly drives Y; in the other it does not. Watching can't tell them apart — both invent a spurious +x. So the engine acts: it forces X and watches what Y does.
A link is kept only when forcing X actually moves Y far beyond chance. The engine drops the fake link (barely a flicker, 0.4σ) and keeps the real one (a strong response, 8.7σ) — a test you can't fake.
Two labs, one law — neither alone
Two sites combine what they learned — without sharing any raw data — and find a rule neither could alone.
One lab only watches; it can't pin the shape. Another only experiments; it has nothing to fit a shape to. Each fails alone. They swap compact summaries — never raw data — and together recover a law neither had.
The two summaries simply add up to exactly what all the data together would give. Pooled, they recover dz/dt = −z + 3·hill(x; K=1.50, n=4) — an S-shaped saturating response — that neither lab could find alone.
It imagines what never happened
It answers “what would have happened if…” — and we can prove it isn't bluffing.
Take one real, noisy run. The engine infers the exact hidden randomness behind it, then replays that same randomness under a different action — a true “what-if” for that specific case.
The what-if matches the real same-randomness outcome almost perfectly, and clearly departs from what actually happened. Deliberately distort the law by 25% and the match falls apart — proof the engine is genuinely reconstructing, not curve-fitting.
It designs its own experiments
When watching isn't enough, it designs the experiment that gets the answer.
The key number lies beyond anything the system naturally does. Passive watching — and fixed, blind pokes — only ever see the gentle start of the curve. The engine spots that its own model is uncertain out there, and deliberately pushes the input further than it has ever seen.
Starting from a range of [0.35, 1.5], it pushes its own experiments out past 8, surrounds the turning point, and pins it exactly (K = 4.00) — where watching alone cannot.
Exact memory, exact federation
Pause, resume, or combine across machines — the knowledge stays exact, in a tiny footprint.
Its knowledge lives in compact summaries. Stop, save, reload, and keep learning — bit-for-bit identical. Train two machines on different data and combine them — identical to training on everything at once.
Resume reproduces the law exactly; combining two machines matches training-on-all to the last digit a computer can represent. And the memory it needs stays flat — no matter how much data flows in.
Run it yourself
Feed it data and watch the law sharpen — while the memory it needs stays flat.
This is the real engine, computing in your browser. Each click adds fresh noisy runs of a predator–prey world; the engine re-reads the law, its error against the truth keeps falling, and its memory holds steady — no matter how much you pour in.
The footprint stays bounded because the engine keeps only compact summaries, not the raw history. That is what lets it learn from an endless stream in constant memory.
Where this goes
The machinery behind these toy worlds unlocks a few things black-box AI simply can’t.
Each builds directly on what you just watched — reading a law, acting to find cause, designing an experiment, imagining a what-if, and merging knowledge without moving data.
A self-driving laboratory
Wire it to a robotic lab or a simulator and it runs its own campaign: choose the most informative experiment, run it, update its causal model, repeat — reaching a mechanism in far fewer experiments than brute-force screening of drugs, materials, or reactions.
Learn together, share nothing
Hospitals, banks, and labs keep their raw data in-house, exchange only compact summaries, and build one shared causal model — provably identical to training on all of it pooled. Collaboration without a data breach.
True what-ifs, one case at a time
Reconstruct the hidden history behind a single patient, trade, or machine, then replay it under a different decision — a unit-level counterfactual, not a population average. “Would this one have turned out differently?”
Models that survive a changing world
It learns the cause, not the correlation, so it keeps holding when the surface patterns break — the moment a market regime flips, a process drifts, or a living system surprises you, and pattern-matching AI fails.
A white box you can audit
It returns the actual equation governing a system — readable and checkable — so a scientist, an engineer, or a regulator can see why, line by line. Not a score and a shrug.
Learns forever, runs anywhere
It folds an endless data stream into bounded memory and never stops improving — no expanding data lake, no nightly retraining. Light enough to run in the field — on a sensor, a vehicle, or an instrument — not only in a data centre.
The pattern is always the same: read the law, tell cause from coincidence, imagine the alternative, and keep learning — on the systems where those four are what matter.
Two of these, on a real case — right now. ↓
Rewind one life, change one thing
It replays a single real history under a different choice — holding that one person’s exact luck fixed.
Here is one patient’s recovery: a symptom that spiked and slowly settled, with every good day and bad night that were hers alone. The engine never sees the biology — only this one wobbly line. It learns how recovery works, backs out her exact luck, then asks: with that same luck, what if the infection had been brought under control from the start?
The what-if lands on the true same-luck recovery — she’d never have spiked — and departs sharply from what actually happened. Feed the engine a wrong theory of recovery and the replay drifts off. So when it matches, it has found the real cause, not a guess.
Invent a rule. It reads your mind.
Hide a rule in noise and the engine makes seven independent reads — they scatter on little evidence, then snap into agreement as you feed it more.
You decide how fast a trend catches on and how big it gets, then bury it in noise. The engine never sees your dials — only the mess. Give it little evidence and its seven reads disagree; give it more and they lock together onto the exact rule you chose.
The engine, on one page
What AITIA is, what it does, and how it’s proven.
- What it is
- A causal world model — it learns the explicit law behind a system and reasons through cause and effect.
- Input
- Noisy measurements of a system over time. No equations, no labels, no derivatives.
- Output
- An explicit, readable law · a cause-and-effect map · true what-if trajectories · a compact state that resumes and merges exactly.
- Lineage
- The convergence of two Causis research threads — CR‑Causal and CR‑Stream.
The model
AITIA is a causal world model, not a neural network. Where a transformer learns statistical patterns from oceans of data and predicts the next token, AITIA learns the mechanism of a system — the equations that actually govern it — and reasons through cause and effect. Its knowledge is a compact library of laws, not billions of opaque weights, so every answer is legible and the model stays small, exact, and fast.
Three levels of reasoning
Most AI lives on the first rung — spotting patterns. AITIA climbs all three: it sees the law in the data, acts to tell cause from coincidence, and imagines what would have happened under a different choice. Association, intervention, and counterfactual — the full ladder of causal reasoning.
What it does
- Even from a cold start, with no prior knowledge of a system, it reads the governing law out of noisy data — and grows surer the more it sees.
- Tells cause from coincidence by acting — not by correlation.
- Designs its own experiments when watching isn’t enough.
- Imagines true what-ifs for a single case — and a wrong law breaks the match.
- Recognizes an unfamiliar kind of structure and extends its own grammar to capture it.
- Learns continually in constant memory, and merges exactly across sites — sharing only summaries, never raw data.
At a glance
How it’s proven
Every result is answer-blind: the engine never sees the truth, and the bar is set by the data’s own noise, never by hand. Validated across chaotic, oscillatory, gene-regulatory, stochastic, and hidden-confounder systems, including held-out interventions.