neo fluidfix 0.15.0 Install
Six laws as the brain

Repairs the bug.
Refuses the guess.

Every decision is a ruling from a machine-authored integer law, judged by your own test suite. No model in the loop.

pip install -U fluidfix Play intro

This is not a picture of the brain. It is the whole brain. Every situation the tool can ever be in, and what it decides in each. 4 squares out of 256 are where it edits your code; the rest are refusals with a reason attached.

Every situation this tool can ever be in. Gold: it edits your code. Grey: it refuses, and says why.hover any square
Re-derive it yourself fluidfix selfcheck

Rebuilds all six laws from their specifications and checks every state. No network, no model, nothing cached.

5,376states checked
0.10sto check them all
0wrong
What the brain is made of

178 integer operations. Nothing else.

The whole engine law is one expression over one variable: 85 ANDs, 56 subtractions, 23 additions, 14 shifts. No floating point. No branches. No weights. It is not a description of a decision that something else has to interpret — it is the arithmetic, running on the same bits every processor already works in.

A neural network ends up as bit arithmetic on that same silicon too — but with billions of learned floating-point parameters and a sampling step between the input and the answer. That is the layer you cannot read. Here there is no layer.

1,555characters, the entire law
2.01nsper ruling, compiled
0floats · branches · weights
For skeptics

“That is just a lookup table.”

Any function on 256 inputs can be tabulated. This one is a lawful table: a closed form derived every row — nobody wrote 256 of them — and it obeys laws a table never has to. Composition holds on all 4,096 cases, identity on all 256, every mask terminates. Teaching a new class adds no row; its act code is derived from your one example on every call. And the body has no decision of its own: every exit consults the law, and that is tested.

Small enough to prove is the trade. A model’s decision function cannot be checked exhaustively; this one is, in a tenth of a second, on your machine. It does not reason and it does not learn — it picks one of eight acts, only one of which writes code, and your suite decides whether that write was right.

4,096/4,096composition cases hold
0rows written by hand
0exits that skip the law

The brain is bounded and exact. What it knows is what you teach it. What it decides, it can prove.

Intro

Watch it rule.

fluidfix guard .
observationstoggle any bit
byte 1 00000001
SHIP
A candidate passed the suite and nothing blocks it. Write it, byte-exact.
Product

One brain. Four surfaces.

Guard

fluidfix guard .

Runs unattended on every push. Repairs what it can prove, refuses what it cannot, leaves your tree byte-identical either way.

Explore →

Teach

register(kind, name, …)

Write one incident down, once. Every future member of that fault class is decided by the kernel for free.

Explore →

Hotspots

fluidfix hotspots .

Your history already names the files that break. Ranked by defect density, with what testing them buys.

Explore →

Verify

./docs/verify.sh

Six laws re-derive over their whole input space on your machine, in about a second. The brain proves itself.

Explore →
Built for

Teams whose regressions come back.

Live-service studio10.7a year

qualifying fixes in an actively shipped game framework's history.

Engine team71%of fixes

of its real bug fixes sit inside one file and 25 lines — the ceiling a taught class reaches.

Open source$0AGPL

Everything, unlimited, no gate. The laws stay public and verifiable.

Proof

Every number has a file behind it.

Box2D · math_functions.h:21246suite runs
- a.x * b.y + a.y * b.x + a.x * b.y - a.y * b.x

The sign flip that rotates everything the wrong way. Byte-exact in 82.8 s, no class written.

measured on a clean clone
Head to head · same defect0tokens

fluidfix: 13 suite runs, 6.3 s, byte-exact. An unaided agent: 43,006 tokens, 73.8 s, also correct. Both arms preserved.

research/headtohead-2026-09-07
Ambiguity12/ 12 separated

Two programs, both green at one line. Before: shipped 12, wrong 6. Now: refused 12, pinning test written, zero extra suite runs.

tests/test_amb_is_program_identity.py
The audit0/ 63 in a ruling

44 research reports, 20 adversarial attacks. Every wrong outcome ever recorded lived in a measurement, never a ruling.

research/AUDIT-2026-09-07.md
cglm · 1,131 tests100suite runs
- dest[0] = a[0] - b[0]; + dest[0] = a[0] + b[0];

Searched past the callers to reach the callee. 22 broken, 1,131 restored.

measured on a clean clone
Teaching · real history4fixes, 1 example

One class taught from a 2016 commit repaired fixes in 2016, 2018 and 2023, across four files.

research/laws-2026-09-07/44
C# / .NET · xunit14suite runs

Listed unsupported until it was tried. The language cost one file extension and one regex.

measured
Available at no charge

For developers

Everything. All six laws, every language path, unlimited repositories, no feature gate.

Install
Six weeks · credited in full

For teams

We measure the number that decides everything: how many of your regressions are this shape, and how many your suite catches. $5,000, and we tell you if it is too small to license.

Start a pilot