Federal Reserve holds interest rates steady, citing stable inflation outlook.
Predicted activity
Loading brain mesh
loading
Monarch
Is this media built
to make you feel,
or to make you think?
Predicting how cortex responds to written media.
Built on TRIBE v2 / fsaverage5 cortical surface / Open source
Predicted activity
Loading brain mesh
loading
What it does
It measures how far a piece of text leans on emotion rather than reasoning, and reports the number it gets.
The text is spoken, transcribed, embedded, and passed to an encoder trained to predict cortical responses. Two networks are averaged, one associated with affective salience and one with deliberative control, and the index is the difference between them. It rates the content, never a person, and it predicts a response rather than recording one.
01 / The idea
Most tools check what
media says. Monarch
checks how it lands.
Sentiment tools, fact-checkers, and credibility scores all read the words. None of them tell you whether a piece of media is wired to trigger a gut reaction before you get a chance to think it through.
Monarch uses Meta’s TRIBE v2, an AI trained to predict how the brain responds to media, and turns that into one plain score: how much this content leans on feeling versus reasoning.
FED DESTROYS AMERICA. Your savings are GONE. The collapse they hid from you!
illustrative, not measured
Two sets of stories, scored
Straight news coverage clusters at the calm end. The same facts re-written to grab attention push the whole set to the charged end.
Loading chart...
synthetic, illustrative
02 / How it works
From a piece of media to a
score, in three steps.
Step 01
Read the content
The item is spoken, transcribed for word timings, and embedded, then passed to TRIBE v2, an encoder trained to predict cortical responses to naturalistic media.
TRIBE v2 brain model
Step 02
Predict the brain response
The encoder returns a predicted response for 20,484 points on the cortical surface. Two networks are averaged from it, one affective, one deliberative.
Whole-brain prediction
Step 03
Ask what it would take to matter
A mean-field model states how strong the coupling between this index and a population would have to be before content of this spread could shift a consensus at all.
Mean-field bound
03 / Capabilities
What Monarch does
One number per item
The affective-salience mean minus the deliberative-control mean, in units of the encoder's standardised output. Positive means affective leads.
A measured corpus, not a demo
400 items across four categories, each scanned twice on a GPU. Every value on this site came from those runs.
Physics that is swept, not fitted
A mean-field model of coupled opinion states, evaluated across a range of couplings. No coupling value is fitted or quoted.
Open source
Built on Meta's TRIBE v2. The pipeline, the analysis scripts and the corpus builder are public.
04 / The physics
The maths behind the score
For the curious. You don’t need any of this to use Monarch - this is the physics that turns a brain prediction into a score and a ripple estimate.
The score compares how active the brain’s emotion-related regions are against its reasoning-related regions. It is a difference, because a ratio of the two was undefined for 69 of the 400 items:
The Landau free energy of an Ising-style population under the content field h = α · X is:
Equilibrium polarisation m* solves the self-consistency equation:
The coupling α cannot be measured from this corpus, so the work states the minimum it would need. For content of measured spread ΔX to tip a consensus, the field must reach the critical field hc:
This is a minimum any claim must meet within the mean-field model, not a claim that media flips opinion.
05 / Research status
What was measured, and what is still open
The dissertation was submitted in September 2026 and is held by the CUEA Library. All 400 items were scanned twice. Three papers, written with Dr. Songa Mutambi, are in review before journal submission.
10.7%
of the variation in the score is explained by the kind of content (η² = 0.107)
d = 0.94
fear-driven content, the largest gap from neutral
0.87
agreement between the two runs (ICC)
α ≥ 4.29
minimum media strength to tip opinion at βJ = 2
Paper 1 / theory
For any content observable used as a field through h = αX, no media-driven transition is possible unless α ≥ hc(βJ) / ΔX. The bound states what the coupling would have to be for the mechanism to work at all, and can be checked before any data are collected. In review.
Paper 2 / the corpus
The four categories separate at η² = 0.107 (p ≈ 10-9), carried by fear-driven content. Outrage raises emotion and reasoning equally, so its score stays flat. The measured spread ΔX = 0.124 gives α ≥ 4.29 at βJ = 2. In review.
Paper 3 / validation
Against public brain recordings of people watching Friends, the brain model reaches r = +0.028 on a 2-minute clip, 30% of the best possible (0.096), with a positive sign. The full-episode test is next. In review.
Stated limits
- The observable is a cortical proxy. The checkpoint is cortical-only and cannot speak to subcortical structures.
- Monarch predicts an average brain's response to content. It never scans a person, and predicted activation is not measured activation.
- The coupling α is not measured. The site quotes only its minimum, α ≥ 4.29 at βJ = 2, within the mean-field model.
- Each content category came from one source collection, so category and source are mixed. A word-counting baseline separates the labels better (AUC 0.98 against 0.63), which is why the score is reported as a measurement, not a detector.
- Group averages are reliable; single-item scores are not, since 51 of 400 items changed sign between runs, as measurement noise predicts.
06 / Attribution
Who is building this
Research project
B.Sc. Physics dissertation, 2026
Brian Mwai, The Catholic University of Eastern Africa, Nairobi.
Supervised by Dr. Songa Mutambi. Preview the dissertation
Apparatus
Open-source measurement pipeline
Built on Meta's TRIBE v2, released under the project's own open-source licence.
The 400-item corpus was scanned twice on a Tesla P100.