Trace. Question. Emerge.In one sentence
Model Meets Reality (MMR) is a filing system that makes being wrong arrive. You write down what you think will happen, attach a date, and the system brings it back when the date comes — so you cannot quietly forget the misses.
The mind analogy for the whole thing: it is the part of your memory that keeps the receipts. Your brain deletes the predictions that did not pan out and keeps a warm feeling that you were basically right. MMR is the drawer that refuses to lose the receipts.
What follows is each capability, paired with the mental faculty it stands in for.
1. A model is a written belief — 👁️ forming a view
A “model” here is not AI weights. It is a short document: what you think is true, what causes it, one dated thing that would prove you wrong, and the conditions under which you would throw it away.
Mind analogy: the moment you look at a situation and think “here is what is going on.” MMR just forces you to write the fourth part — what would change my mind — which your brain almost never does.
2. The watch loop — 🔭 paying attention
Each model has a list of what it watches. It reads the web, records what actually happened with sources, and files dated claims.
Mind analogy: selective attention. You cannot watch everything, so you decide in advance what you are tracking. Different models watch different things — one watches polls, one watches court filings — so “what happened” is never one answer.
3. The grading loop — ⏰ being held to account
When a claim’s date arrives, the system grades it against reality. Hit or miss, both get recorded.
Mind analogy: the friend who remembers what you said last year and asks how it turned out. The faculty most people do not have — honest hindsight — because the brain rewrites the past to feel consistent.
4. The arena — ⚖️ internal debate
Several models argue one event. Each states its position before seeing the others, then must defend it or change its mind — and a blind judge, who never saw the underlying reasoning, names the one question everything hinges on.
Mind analogy: thinking against yourself. Steelmanning the opposing view, then finding the single fact the disagreement actually turns on — instead of the crowd in your head all nodding along.
5. The reaction layers, the onion — 🧅 tracing who influenced whom
Given an event, it maps who was responding to what — the original act, or a reaction to it, or a reaction to that. Each link carries the quote that proves it.
Mind analogy: untangling gossip. Working out who actually said the thing versus who is repeating a distortion of it — and how many hands it passed through before it reached you.
6. The pressure model — 🎯 reading what forces someone’s hand
For a person in power, it reads the forces acting on them — each pressure carrying the consequence if they refuse — and works out what they are likely to do.
Mind analogy: theory of situation, not theory of mind. Not “what does she want,” which is unknowable, but “what happens to her if she does not,” which is observable. You already do this when you guess a decision from someone’s constraints, not their feelings.
7. The emergent-behaviour catalogue — 🚨 learning what to fear from what already burned you
It reads real AI incident reports and lists the behaviours a system built — not just the tools it was handed — ranked by what each one unlocks, with the monitorable signal for each.
Mind analogy: a scar teaching caution. You do not fear the stove in the abstract; you fear it because you were burned once and now you watch for the specific early sign — the heat shimmer, not the flame.
8. The gap explorer — 🗺️ noticing what you are not looking at
It reads the coverage map, finds the empty regions, proposes new models to fill them — then a challenge panel tries to knock each one down.
Mind analogy: the discipline of asking “what am I missing,” and then actually attacking your own new idea instead of falling in love with it.
9. The tamper-evident record — 🔒 a memory that cannot be edited after the fact
Every model is individually git committed and entry is cryptographically chained to the one before it. Revert to earlier model, fork it, analyze the trajectory.
Mind analogy: the opposite of how memory works. Your brain edits the past to match the present. This is a memory that physically cannot be rewritten to make you look right.
10. Meta modelling — 🔁 fixing the models, from inside and from above
Every model above can be wrong, and most of them will be, somewhere. So the system turns the same tools on the models themselves — from two directions, and in a loop.
From inside: take one model on its own terms. When a dated claim fails, trace the failure back to the premise or mechanism that produced it, and amend that — dated, with the old version kept, so every graded result stays tied to the exact version that made the call. A model that has been corrected against three resolved dates is a different instrument from the one that started, and the record shows how.
From above: step outside the model and ask the questions no model can ask about itself. Was that hit earned, or lucky? Which kinds of claim keep failing, across all of them? What can none of them see? And one level further up: is the process stable — does the same brief, argued twice with no new facts, give the same answer? Whatever moves there is noise, and any “change” smaller than that noise is not news.
In a loop: predict, grade, trace the failure, amend, predict again. The next reading starts from what survived, not from scratch. Reading number five should be better than reading number one — and the ledger is what makes “better” checkable rather than asserted. Most of this loop has not turned yet; the dates it needs are still arriving.
Mind analogy: revising how you think, not just what you think. Anyone can update an opinion. The rarer habit is noticing you have been wrong the same way three times and fixing the habit. The rarest is checking whether your way of checking is itself any good.
The honest limits — 🪞 knowing what you cannot see
It is one person’s reading list, their own unique information bubble, run through one AI. When the models agree, that is a fact about the reading list, not about the world. The system prints this warning on every run.
It cannot see what stays hidden. The reaction maps are built only from what became public — hugging face AI incident took 119 days to surface, and only because an unrelated third party noticed. Anything detected quietly and never disclosed is invisible to the whole system by construction. But if you model the pressures right you can still see signs.
The one line to remember
Every capability above is a specialised organ, but they all serve one nervous system: turn a vague feeling of “I think X” into a dated, sourced, un-editable bet — and let reality grade it.
Want to see it in action?
Here is how I am tracking AI-threat possibilities right now — one real event, pulled apart into its reaction layers, with every actor selectable as the centre and every link carrying the quote that proves it:
Pick any actor and pivot the whole map around them. What you are looking at is built only from what became public — which is exactly the limit named above, made visible: the layers you can see are the layers that surfaced.
Every instance grows differently
There is no single MMR. Everyone starts from the same template — the empty capabilities above — and from there each instance grows on what it actually analyzes. Watch different events, argue different disputes, emerge new capabilities, read different incident reports, and the models you accumulate, the blind spots you find, and the behaviours you learn are yours alone.
Start yours
The engine is free — clone it, or in a new folder, open a new session of any coding harness like Claude code and ask:
https://github.com/EmAssets/adjacent-possible
help me use this.
Grow one or multiple instances on your information bubble. If something intresting emerges share with hastag #emergencemachine, or put it where the rest of us can use it. To see what emerges here, subscribe.
Stop procrastinating, put a date on it.
-Sail