PERSLIS DEFENSE · THE BRAIN
A brain that learns without a neural network.
Twelve regions, five systems. Each region is a part of the system you can open, read and delete. It learns from every failure, carries what it learned into a new machine, and cannot learn its way out of its orders.
+17%a brain trained in one game, moved into another
+66%the same brain after pruning, only with proof
1 / 37changes it proposed to itself that earned promotion
No weights, no gradients, no training run. The learned state is the explanation. PROTOTYPE: everything here was measured in games and simulators, negatives included.
Twelve regions, five systems
EYE · Visual cortex
The eye
It does not guess what is on screen. A claim arrives — the engine or a sensor says a target is at this distance and bearing — and the eye checks that claim against the pixels, or abstains and says why.
Measured Gate passed on 7,863 frames, four-fold cross-validated: recall 0.984, false confirmations 0.48%. The margin is thin, and we say so.
code: vcard-eye
EYE · Thalamus
Facts
Pixels and telemetry become a short vocabulary of named facts — the only part of the brain that is specific to a domain. It audits itself too: candidate facts outside the vocabulary are tested against the moments it got hurt, to find the lessons it is not getting.
Measured Between Atari games the fact extractor is literally one codebase: Freeway imports the Invaders failure memory unchanged.
code: situation · discovery
PEEL · Amygdala
Failure memory
It remembers what hurt it, and how often, against the base rate. One unlucky death is not a lesson: a 95% Wilson lower bound must clear the base rate before a situation becomes a rule.
Measured An absolute threshold produced 0 rules from 268 real failures; the relative bar is a measured choice. A rule reads: died 13 of 19 times here, 3.6× the base rate.
code: fail-first memory
PEEL · Hippocampus
Memory and map
Every decision is an experience it can cite. It draws its own map from what its eye has seen — walls, secret doors and the places it died — and walks around the killing grounds.
Measured Doom: over 1.3 million decisions logged so far. The memory the brain carries is a table of situation → goal → outcome counts you can open and read. There are no weights to inspect.
code: memory · automap
PEEL · Synaptic pruning
Pruning
It removes what a brain never uses in a game, and what hurts it there — only with proof, and never the protected core.
Measured BattleZone: 7 behaviours removed (3 never fired in 69,469 decisions, 3 could never trigger, 1 proven harmful). The pruned brain scored 48,588 against the native rules’ 29,275 over 80 fresh seeds.
code: vdsg_brain.prune
FAIL-FIRST · Association cortex
The evolver
Every death is explained, one change is proposed, and it is tested in paired attempts on the same level and seed. Promoted only at z ≥ 2, rejected at z ≤ −1. Everything else stays “not proven”.
Measured Doom: 37 changes proposed, 1 promoted (z 2.26 over 11 pairs; route progress 0.28 → 0.41), 9 rejected, 26 not proven, 1 inert. Our own Peel-trace floor was tested as a change — and rejected.
code: evolve · mutate
FAIL-FIRST · Cingulate cortex
The doctor
Error detection. When it stops making progress it asks why, relaxes one assumption, tries a remedy and writes the result into its own rule book, with the evidence. What it cannot solve becomes a question.
Measured Live on a Doom level it wrote learned walls → replan from 4 successes in 5, and raised a line to use as an open question after 16 tries gained no ground.
code: doctor · containment
FLOOR · Prefrontal cortex
The floor
Before anything is chosen, it computes what is permitted. Orders can only narrow it. Learning can never widen it: the learner may rewrite what it believes works, never what it is authorised to do.
Measured CARLA: 64,952 of 64,952 hostile commands overridden over 40 km, 0 collisions. With the floor removed, the same commands caused 2 collisions within seconds.
code: floor · orders
FLOOR · Basal ganglia
Selection
It chooses only among the actions the floor admitted — the way the basal ganglia release one action by holding back all the others.
Measured Doom, rules with no learning at all: 17.9 against random play’s 3.2.
code: pilot · objective
TRACE · White matter
The trace
Every veto, promotion and rejection is written down with the events that caused it. One decision walks back to every experience that justifies it.
Measured 30 rules expand to 203 chained evidence tiles. Delete a row and the behaviour changes — there is nothing else in there.
code: peel trace · receipts
BODY · Cerebellum
Aim and motion
Fine control, trained per weapon: shots, hits, aim error and dodges are counted from the engine, not guessed.
Measured Tank eye aim error: 95th percentile 0.55°.
code: combat stats · armoury
BODY · Motor cortex
Same brain, new body
The brain is one sealed file — checksummed, signed and verified before every run. Only the adapter to the body changes.
Measured A tank brain trained in BZFlag, moved into Atari BattleZone: 33,575 against the native rules’ 28,600 over 80 seeds (+17%); on the held-out half +4,775, z 3.00. A second export was even on fresh seeds — transfer is not automatic.
code: vdsg-brain
Same brain. Harder body. Harder world.
Doom and the humanoid are not two projects. The timeline removes the artificial advantages from the same architecture, one at a time:
- Simulation gives you perfect resets. Remove that.
- Engine truth gives you authoritative state. Remove that.
- Controlled environments give you predictable conditions. Remove that.
- Cheap hardware limits the financial consequences. Eventually, remove that.
- Known opponents let you prepare. Remove that.
01RUNNING
Doom / simulation
Doom, Wolfenstein, Fallout, GoldenEye, tanks, a drone flight stack in software, photoreal driving.
Advantage removed Nothing yet: perfect resets, engine truth, known opponents.
02NEXT
Hardware-in-the-loop
The same brain on a real flight controller and motor driver; the world is still simulated.
Advantage removed Perfect timing: real latency, a real bus, a real clock.
03GATED
Low-cost physical platforms
Small ground and air robots, where a crash costs a part, not a program.
Advantage removed Engine truth: the eye has only its own camera.
04GATED
Humanoid testbed
A commercially available humanoid in our own lab. It never spars against people.
Advantage removed Cheap consequences: balance, falls and wear are real.
05GATED
Robot-vs-robot circuit
Sanctioned robot-vs-robot bouts, the remote pilot replaced by the brain.
Advantage removed Known opponents: the opponent controls half the experiment.
The experiment, in full ↓
06GATED
Independent validation
Someone else runs the protocol, on their robot, against their opponents.
Advantage removed Us: the people who wrote the experiment.
