Drag to turn it. Scroll to walk through it.
Small. Explicit. Inspectable. Adaptive. Portable.
We began with a small problem: could a child’s toy have a useful brain of its own, offline? Solving it forced us to work out how intelligence is compressed, represented, moved, inspected — and finally changed by experience. Scroll, and watch the brain assemble in the order we built it.
00Day 1
Can a child's toy have its own small, offline brain?
Intelligence that lives inside the product — not rented from a cloud for as long as the product exists. Big models as teachers and tools, never as the permanent brain.
01Feb–Mar 2026
Small, specialised brains
The first body was a communication app for children who cannot rely on speech (AAC). Its brain was a store of conversation trees, running offline on cheap tablets.
On the record Peel began with augmentative and alternative communication.
02Mar–May 2026
Teacher → verifier → student
A large teacher writes candidate data; a verifier admits it; a small student is trained from scratch only on what was admitted. It is not distillation: nothing reaches the student that the verifier did not admit.
On the record TinkyBrain v6: 25.5 million parameters, 24 MB, about 2.9 s per answer on a 2017 Kindle Fire.
03Mar 2026 →
Knowledge stops being trapped inside weights
A typed store of records and relations: every answer is a row you can read, cite and delete. The model is demoted to generator or renderer; the store is the only author of facts.
On the record About 372,764 structured records. Conditioning on the structure raised answer diversity from 0.589 to 0.966 with the same model.
04Mar 2026 →
The system remembers what happened, and why
Every fact carries where it came from. Decisions become inspectable state transitions with provenance, instead of output that is gone once it is read.
On the record Every fact is traceable to the source it was extracted from.
05Sep 2026
Failure becomes the mechanism that changes the brain
Observed failure → evidence → hypothesis → a candidate change → a controlled test → reject, not proven, or promote — inside an authority the learner cannot widen.
On the record 4,280 cards became 30 rules and 203 chained evidence tiles. Space Invaders +32%; Freeway −12%, published.
06Sep 2026
It changes while it runs, without gradient descent
Operational knowledge changes between attempts while the system is running: an evolver proposes, a doctor diagnoses, pruning removes — each only with proof. No neural weight is updated.
On the record Doom: 37 changes proposed, 1 promoted. BattleZone: pruning with proof, +66% over the native rules.
07Sep 2026
Stop building a new brain for every environment
The same architecture runs Doom, GoldenEye, Fallout, tanks, a drone course, photoreal driving and legacy software — and a brain trained in one game can be sealed, moved and run in another.
On the record A tank brain trained in BZFlag scored 17% above BattleZone's native rules over 80 seeds.
08Next
The world stops being software
Hardware in the loop, low-cost physical platforms, a humanoid testbed, then sanctioned robot-vs-robot competition — each gated behind the one before.
On the record The robot we plan to put in a ring is a descendant of the toy brain.
That work became SML. SML led to explicit knowledge structures. Those structures led to Trace. Trace made failure observable. Failure led to Fail-First. Fail-First led to Peel. And Peel brought us back to where we started: a factory for brains.
A neural model can have billions of weights. Peel does not need to change them to learn something new: its persistent learned state is explicit — rows you can read, cite and delete — and the floor bounds what any of it may do.
Learning without weight updates is not new: symbolic learning, inductive logic programming, program synthesis, evolutionary methods and shielding all do it. What we claim is narrower — this architecture, the method that produced it, and the evidence behind every change it makes. The Fail-First Model and its prior art →
Doom is not the company. The fighting robot is not the company. The children’s communication app is not the company. They are bodies and worlds a Perslis brain has been placed in.