5 reproducible studies · open data · open code

You Can't Know Someone's Prior

Bayesian models pathologize what they can't predict. The only honest response is safety.

5
Studies
11
Priors tested
70%
FPR (Eklund 2016)
16.2%
Check sensitivity

Schizophrenia has been called "aberrant Bayesian inference." Autism has been called "deficient predictive processing." Both assume you know what someone's brain should expect. You don't. Nobody does. Nobody ever will.

The 5 Studies

01

Prior Sensitivity

Same data, 11 priors, 3 sample sizes. At n=25 (typical clinical study), prior bias reaches 40%. The prior controls the conclusion.

Conjugate posteriors 11 × 3
02

The Dead Salmon Problem

125,000 voxels of pure noise. The stronger your prior belief in activation, the more "activation" you find. In nothing.

11 pipelines 125K voxels
03

Diagnosis Depends on the Prior

500 people, 3 Bayesian classifiers. Same person is "pathological" or "adaptive" depending on what the clinician believed before they walked in.

3 classifiers N=500
04

Computational Intractability

Exact Bayesian inference is NP-hard. The brain has 86 billion neurons. Even approximations disagree with each other.

Variable elimination 5-25 nodes
05

Safety as the Only Honest Prior

200 agents, 500 timesteps. Safety-first: everyone converges. Punitive: everyone diverges. You don't need the prior. Make them safe.

Agent-based N=200

Run It Yourself

You don't need a science degree. You need Python and 5 minutes.

1

Get the code

Download the ZIP from GitHub, or:

git clone https://github.com/webmasterproT/the-prior-problem.git
cd the-prior-problem
2

Set up (one time)

bash setup.sh          # Mac / Linux
python3 setup_windows.py   # Windows
3

Run all studies

bash run_all.sh        # Mac / Linux
python3 run_windows.py     # Windows

Or run just one: bash run_one.sh 3 (diagnosis depends)

4

See the results

Each study's results/ folder has the output. results/figures/ has the charts. WHAT_THIS_PROVES.md explains it in plain language.