Demismatch / Cor

Raising $22,000active

An evidence-graded atlas of what humans need, and the first inter-rater reliability study of whether it holds in anyone else's hands.

Project Details

Cor is the result of more than a decade working with this subject matter in various forms, and one year full-time, unfunded, culminating in an atlas/reference that answers exactly that: what a human being needs. It is a stated method applied to a large body of sourced literature from the width of human sciences, published so it can be re-run. It lists 15 motivational-emotional systems in a human, for each with what activates it, what resolves it - and what fake proxies activate it too but without supplying the actual function. https://cor.demismatch.com/mechanisms.html

Main public page: https://demismatch.com. Embedded here are the public explainer film "You're Not Broken" and the (for this application more important) academic breakdown of "Cor". Just in case, the direct youtube links:
"You're Not Broken": https://www.youtube.com/watch?v=GyNi2ZUSsuk&t=1s
Cor - Academic breakdown: https://www.youtube.com/watch?v=PyomiIbZIQQ

Cor front-page: https://cor.demismatch.com
The derivation: https://cor.demismatch.com/derivation.html
The gap register, logged before anyone asked: https://cor.demismatch.com/reference.html#ref-gaps
A worked case, a lonely 22-year-old and a chatbot: https://cor.demismatch.com/case/ai-companion.html
Cor-engine, a working prototype: every citation resolves to a real record or the report doesn't ship: https://github.com/MaartenRischen/cor-engine
Decode:talk, the first tool built on the atlas: https://demismatch.com/decode-talk/

The strict academic approach is to honor the importance of the subject by removing my own subjective bias and opinion from it as much as possible.
In that same vein, the ask for this grant is to be able to devise coding manuals to make the fifteen systems executable, and then get them in the hands of independent raters, strangers who have not worked on this at all.

Theory of Impact

AI is increasingly the environment humans live in, and it's being optimized against what people want. But wanting is a signal from the human organism that something is missing, and nothing in the alignment stack says what. So an optimized AI can satisfy the signal perfectly while the thing the signal was about goes unmet, and the satisfaction, short and endlessly repeatable, reads as proof it worked.

That's how you get a system that follows instructions, is non-deceptive and perfectly preference satisfying, and still optimizes people into dependency and a narrowing of what they can do for themselves. No treacherous turn to catch, no deception to detect, no capability threshold to gate - none of the machinery built for x-risk sees this one. It's a curtailment pathway with no measurement layer, and the endpoint is a species that got everything it asked for, can no longer do anything for itself, and never once got what the wanting was actually about.

Cor is a candidate for the missing layer. Fifteen systems, and for each what activates it, what actually resolves it, and what only fires the cue. That gives an alignment target something to score against other than the satisfaction it produced. It's the difference between a model that the company and one that points towards the real thing.

People

Maarten Rischen

Team Member

Grants Received

Add grant

SFF Speculation Grant 2026

from Survival and Flourishing Fund

$10,000

Funding Asks

grantmaking.ai Launch Round

Applied

Minimum

$22,000

Ideal

$36,000

What this ask is

Demismatch is my umbrella initiative about evolutionary mismatch; how most modern suffering is an accurate response of the human organism in a world that it is simply not made for. I've been obsessed with how these mismatches are being hijacked through fake - often technological - proxies, but also what does meet the actual deep human urges. None of this is primitivist or anti-tech, and my materials will show that. The goal is demismatching the world so it fits humans. AI matters a lot here, because it increasingly is that environment or at the very least shapes it. And currently it's being optimized against preferences - the field is working on this, but doesn't definie what the wanting is a signal from, let alone what a human really needs. That is why I built Cor.

Cor is the result of more than a decade working with this subject matter in various forms, and one year full-time, unfunded, culminating in an atlas/reference that answers exactly that: what a human being needs. It is a stated method applied to a large body of sourced literature from the width of human sciences, published so it can be re-run. It lists 15 motivational-emotional systems in a human, for each with what activates it, what resolves it - and what fake proxies activate it too but without supplying the actual function. https://cor.demismatch.com/mechanisms.html

Main public page: https://demismatch.com. Embedded here are the public explainer film "You're Not Broken" and the (for this application more important) academic breakdown of "Cor". Just in case, the direct youtube links:
"You're Not Broken": https://www.youtube.com/watch?v=GyNi2ZUSsuk&t=1s
Cor - Academic breakdown: https://www.youtube.com/watch?v=PyomiIbZIQQ

Cor front-page: https://cor.demismatch.com
The derivation: https://cor.demismatch.com/derivation.html
The gap register, logged before anyone asked: https://cor.demismatch.com/reference.html#ref-gaps
A worked case, a lonely 22-year-old and a chatbot: https://cor.demismatch.com/case/ai-companion.html
Cor-engine, a working prototype: every citation resolves to a real record or the report doesn't ship: https://github.com/MaartenRischen/cor-engine
Decode:talk, the first tool built on the atlas: https://demismatch.com/decode-talk/

The strict academic approach is to honor the importance of the subject by removing my own subjective bias and opinion from it as much as possible.
In that same vein, the ask for this grant is to be able to devise coding manuals to make the fifteen systems executable, and then get them in the hands of independent raters, strangers who have not worked on this at all.

How the money will be spent

Ideal: $36,000. Four months, August to December 2026.

Project Director, 4 months at $4,500 — $18,000
Four independent raters at $2,500 — $10,000
Pilot rater (excluded set) — $2,500
Rater recruitment and vetting — $1,500
Compute: three model arms, twenty cases, multiple runs — $4,000
Total — $36,000

Four months of my time produces the instrument. Fifteen coding manuals precise enough that a stranger can't drift. Twenty documented cases, including the ones where Cor loses. A preregistration that freezes the manuals, the cases, the statistic and the failure thresholds before a single score is entered. Three model arms. Then publication, whatever the number says. That is the deliverable, and it has to be right the first time, because a preregistered study can't be revised after the data lands.

$10,000 buys the four people I can't be. Each rater reads twenty cases and makes 300 case-by-system judgments against criteria they've never seen, with written reasoning on every non-trivial call: eight hours of qualification, thirty of scoring, ten of documentation. A working week. $2,500 is the rate for a professional who takes it seriously.

$2,500 buys one revision. A pilot rater scores an excluded set first, so a badly worded manual surfaces before the confirmatory set is frozen rather than after the data is in.

$4,000 runs the models. Three frozen families, three conditions each (names only, definitions, full criteria), twenty cases, multiple runs at fixed settings, with contingency for reruns if a provider version changes mid-study.

Minimum: $22,000. Three months of director time ($13,500), compute ($4,000), the pilot rater ($2,500), case production and publication ($2,000).

That buys the fifteen coding manuals, the twenty-case benchmark, the preregistration, and the model arm. All public, all attackable. It does not buy the raters.

So the minimum doesn't fund the study. It funds everything the study needs, so that anyone can run it. The manuals and the benchmark exist either way. At that point the answer costs $10,000 and four strangers, and it's available to whoever wants it. Fund the ideal and I get it in December.

Discussion

No comments yet. Be the first to share your thoughts.