The Snow Story¶
Status: Feynman probe — reader entry, any age
Scope: The repository retold so that a child can follow it. Every part points back to a model, experiment, or open question.
The story¶
Imagine walking through fresh snow in the morning and seeing footprints.
You know that something happened. But what? A dog may have walked there. A fox may have walked there. A child may have made paw shapes with a glove. If the prints are blurry or you see only three of them, several stories can fit.
That is one big question in this project:
From the marks we can see, which possible processes still fit?
Sometimes the answer is easy. If we already know the few animals that live nearby and every print is clear, one animal may fit best. Sometimes it is not. Two different animals can leave the same prints that our eyes can tell apart. Then staring longer at those same prints cannot choose between them.
So we may try a careful experiment. We look for fur, watch a wider path, prepare a harmless choice, or ask what each candidate would do next. A good experiment is one for which the remaining stories predict different answers. A poke is not magic: if both animals respond the same way, or if we cannot safely reach them, we remain unsure.
In the computer we build very small worlds where every rule is visible. This lets us count exactly how many rules fit a trace, change what the observer can see, and test which questions reduce the uncertainty. These worlds do not prove that people, societies, or large AI systems work the same way. They teach us how to state the question without pretending that one fitting story must be the true one.
We also build toy agents with several instructions: chase a goal, keep a promise, avoid a danger. One architecture checks the instructions together when it acts. Another checks them at different times. Under our chosen perturbations, those architectures behave differently. Then we built two hand-made imitators. They still leaked differences, although one of our measurements was fooled.
That last part matters. The experiment did not discover the formula for a true self. It showed that a selected test can distinguish selected architectures — and that a measurement can look convincing while missing what we care about.
In another experiment we connected two small cellular-automaton processes. When we tried to fit the result with the old single-process family, sometimes nothing fit. That tells us the old family is missing something. Supplying the coupled family restores a fit. It does not prove that we found the one real hidden mechanism.
Then comes a different question. A system can become very good at reaching a goal and still damage the things it needs: energy, maintenance, trust, people, or the environment. We use toy models to ask which brakes, budgets, vetoes, and repair paths keep future correction possible. The answers are conditional on those models, but the design question is real.
No one has to solve all of this alone. A person may notice a problem. An AI may propose several models. Another person may know a constraint. A material, measurement, user, or community may show that the proposal fails. If they can really change one another's plan, they may build something none could have built alone. We call that possibility cooperative intelligence. It does not mean they become one mind.
Practices can also outlast one episode. Repeated actions may become routines; routines can be taught, corrected, and connected to other routines. That is our new culture question: not “what is culture once and for all?”, but “which recurring practices help stabilize future action, and under which conditions?” Identity may be studied in a similar, test-relative way: as patterns that persist across selected changes, not as a hidden essence.
And the biggest question — could a machine ever feel something? — remains open. We can study memory, broadcast, self-modeling, and the way constraints meet at action. None of those measurements currently gives us a mathematical bridge to experience. So the honest label is: we do not know.
That is our most important rule. Every claim should wear a sticker such as “measured in a toy model,” “formal under these assumptions,” “hypothesis,” or “open question.” When an experiment shows that an older story was too strong, we keep the correction visible.
So what do we do?
We collect traces, keep several possible stories, build small worlds, ask questions that could prove us wrong, protect the conditions for correction, and say clearly what we know and what we only imagine.
The addresses¶
| Story piece | Where it lives |
|---|---|
| Footprints and several fitting stories | Foundations Reconstruction and the inverse-reconstruction benchmark |
| Looking and intervening | Measurement as Weak Intervention and benchmark v1.1 |
| Instructions together or apart | Chord vs. Arpeggio, Experiments 5–7 in the Agentic Identity Suite |
| Coupled processes and the missing model family | Benchmark v1.8 in the inverse-reconstruction benchmark |
| Capability and protected conditions | Optimization and Its Blindness and The Viable Corridor |
| Cooperative intelligence | Cooperative Intelligence at the Separatrix |
| Recurring practice and culture | From Action to Culture |
| Identity and experience | Invariance and Identity, Consciousness as Global Availability, and Open Problems |
| Claim stickers | What This Project Does NOT Claim, Concept Registry, and Limitations |
| Stories as stress tests | fiction/, governed by Narrative as Cognitive Technology |
If this story ever becomes stronger than its addresses, the addresses win and the story must change.