PERSLIS DEFENSE · AUTONOMY THAT ACCUMULATES OPERATIONAL EXPERIENCE
Adaptive autonomy. Fixed authority.
Imagine a machine that learns without changing a single neural weight. No retraining every time the world changes. No learned strategy buried inside billions of parameters. Perslis uses weightless symbolic learning: experience becomes explicit knowledge the system can test, reject, retain, and carry forward. Meet the Perslis Brain—a self-evolving architecture built for autonomous machines operating in real missions, real environments, and everyday robotics.
Read the white paper ↗ The Fail-First Model →
+17%a tank brain moved into a different tank game, against that game’s own rules — the brain moved was hand-written by us, so this is transfer, not learning
+66%the same brain after cutting what didn’t work, only with proof — its own tests chose the cuts; we wrote the procedure
1 / 37fixes it tried on itself that proved out — the other 36 didn’t make the cut
Twelve parts. What each one does for you.
EYE · Visual cortex
It doesn’t guess.
A sensor reports something at a range and bearing. The eye checks that report against what the camera actually sees — or it says can’t confirm and why. Bad data gets flagged, not acted on.
Why not a neural model? A neural detector hands back a label and a confidence score. It can be confidently wrong, and it cannot show you which part of the check failed. Ours confirms the claim against the pixels, or abstains and says why.
EYE · Thalamus
Plain facts, not a black box.
Everything the sensors pick up is boiled down to a short list of named facts you can read. That list is the only part rebuilt for a new mission. It also checks itself for the lessons it keeps missing.
Why not a neural model? A neural network’s internal features are millions of numbers nobody can name. Ours is a short list of named facts you can read, and a new mission means rewriting that list, not retraining a network.
PEEL · Amygdala
It remembers what got it hurt.
One bad break isn’t a lesson. When a spot keeps getting it hurt far more than anywhere else, that becomes a rule: stay out. The same ambush stops working on it.
Why not a neural model? A deployed neural model is frozen. It cannot turn today’s ambush into a lasting lesson mid-mission and change strategy; that takes collecting the data and retraining its weights, back at the depot. A language model can adjust inside one conversation, then forgets it. Peel changes strategy during the mission, the moment a failure is proven, and keeps the lesson.
PEEL · Hippocampus
It draws its own map.
Walls, hidden passages, and the places it got killed. It routes around the kill zones, and every move it makes points back to the experience that justified it.
Why not a neural model? A neural policy’s memory is smeared across its weights; you cannot point to the experience behind a choice. Every choice here cites the experiences that justify it, in a table you can open.
PEEL · Synaptic pruning
It cuts what doesn’t work.
Behaviours it never uses, or that get it hurt, are stripped out — only with proof, and never the core safety rules. Less on board means less to go wrong.
Why not a neural model? You cannot delete one behaviour from a neural network without retraining it and hoping nothing else moves. Here a behaviour is one entry: removed only with proof, and the core safety rules are protected.
FAIL-FIRST · Association cortex
It only keeps what’s proven.
Every loss gets explained. It tries one fix at a time, runs it head-to-head against the old way on the same ground, and keeps it only if it wins clearly over repeated runs. Most ideas fail. That’s the point.
Why not a neural model? Retraining a model changes millions of weights at once; nobody can say which change made it better or worse. Here exactly one change is tested at a time, head-to-head, and kept only if it wins clearly.
FAIL-FIRST · Cingulate cortex
When it’s stuck, it works out why.
When it stops making progress it questions one assumption, tries a fix and writes down what worked, with the evidence. What it can’t solve, it hands to a human as a question.
Why not a neural model? A stuck neural policy keeps doing the same thing until people collect new data and retrain it. This one diagnoses the stall itself, writes the fix into its own rule book, and flags what it cannot solve.
FLOOR · Prefrontal cortex
It learns what works. It doesn’t decide what’s authorised.
You and your chain of command set the rules. The floor holds the machine to the version of those rules it was given — constraints written from orders, mission rules, safety limits and the applicable ROE. Before it picks a move it works out what those constraints allow. Orders can narrow that list; nothing it learns can widen it.
Why not a neural model? A neural model’s limits are learned, so more training can quietly move them, and a clever prompt can talk a language model past them. Here the limits are computed before every move, and learning has no access to them.
FLOOR · Basal ganglia
Forbidden moves are off the table.
It chooses only from the moves the floor allowed. No acting first and getting vetoed after — the forbidden options are gone before it decides. One pass, no arguing with the safety layer.
Why not a neural model? Safety filters on neural models usually check an action after the model has picked it. Here forbidden actions never reach the choice at all.
TRACE · White matter
Every decision has a paper trail.
Every block, every change, every rejection is written down with what caused it. At the after-action review you can walk any decision back to why it was made. When something goes wrong, someone can answer for it.
Why not a neural model? A neural model can produce an explanation, but not the actual cause: its decision is billions of multiplications. Here the reason is the record itself — delete a row and the behaviour changes.
BODY · Cerebellum
Aim and movement: counted, not claimed.
Shots, hits, aim error and dodges are counted per weapon from the simulator, so you know how good it actually is — not how good it says it is.
Why not a neural model? Neural control is usually judged by one aggregate reward number. Here every shot, hit, miss and dodge is counted per weapon, so you know exactly where it is good and where it is not.
BODY · Motor cortex
Same brain, different machine.
The brain is one sealed, signed file, checked before every run. Move it to a new vehicle and only the connector to that body changes; what it learned comes along. Not every move works, and we say so.
Why not a neural model? Moving a neural policy to a new vehicle usually means new data and retraining. This brain is one sealed file; only the adapter to the body is written. Proven so far between two tank games.
CONNECTION
It runs your legacy systems too — every system you run, under one brain.
Connect all of your endpoints right into the brain. Old computers with no API, no problem: it drives legacy software through its own buttons, fields and screens, reaches mainframes through the ways in they already have, and pulls old data out with a receipt for every step. Nothing gets ripped out.
You would not let a language model run your nuclear safety computer. The GAO found the system coordinating the nation’s nuclear forces running on a 1970s IBM Series/1 with 8-inch floppies. Systems like that are exactly where a model that guesses has no place — so don’t let one run anything where you need to know exactly what it will do, and why. Peel does the same thing for the same facts, every time, and writes down why.
How the connection works →
GO DEEPER
Pick your question.
- Connection Every system you run — even the ones with no API — under one brain.
- Defense How it stays alive and inside its orders, compared game by game.
- Offense It fights in DOOM, and how it did against the big language models.
- Evolution How it gets better by losing, and the scoreboard for every game.
- Timeline From a toy’s brain to a combat humanoid under human command.
- Robotics From a drone course to the robot boxing ring.
- Alignment Aligned to its commander, not to a model. Completed or failed, and every failure becomes a rule.
Talk to us ↗