The observer
learns you.

We describe Apeiron, a system that constructs a portable, verifiable model of one person — an AVTR — from a stream of small, user-authored observations. The system is an allocator, not a predictor: it does not ask "what would the average person say next?" but "which of this person's moments are worth keeping forever?" We give the selection function (CoNov), the storage object (the 16 KB self file), the integrity mechanism (a hash chain over committed observations), and the interoperability protocol (a handshake that lets any frontier model load the verified self). We report one controlled observation — the bedroom experiment — and we state plainly what is established, what is engineering, and what is conjecture.

0. The rule of the game

Feynman's rule for not fooling yourself is that you must state what you know, how you know it, and where you might be wrong — and you must say it in language a smart outsider can check. This paper follows that rule. Every formula below is either standard mathematics with a citation, or a design choice we made, labeled as a design choice. Nothing is dressed up as a law of nature that is actually a knob we tuned.

1. The problem: the average is not a person

A frontier language or image model is trained on the recorded output of humanity. It is, by construction, an estimator of the population distribution of human expression. This is its power and its ceiling: ask it for "a childhood bedroom" and it returns the expectation of that phrase over roughly everything ever posted — cheerful, well-lit, true of no one. The mean of a billion bedrooms is nobody's bedroom.

The information that distinguishes one person from the population — their history, taste, and manner of attention — is precisely the information a world model marginalizes away. Our claim is narrow and testable: a small amount of curated, verified, person-specific information, injected at inference time, moves a general model's output measurably from the population mode toward the individual. The bedroom experiment (§8) is one demonstration. The engineering question is then: what is the smallest durable object that carries this information, and how do you keep it honest for a lifetime?

2. Allocation, not prediction

A predictor minimizes surprise about the next token. An allocator does something different: given a stream of candidate observations, it decides how much permanent weight each one gets. We model the person's recorded state as a Markov process: the state at step t depends on the state at t−1 plus the newly committed observation. This is a design choice, and an honest one — a full history-dependent model would be richer, but Markov structure is what makes the system auditable: every state transition is one block, and you can replay the chain from genesis and land on the same state. Auditability is worth more to this system than modeling power.

The variational framing — we borrow it as an analogy, and say so — is Hamilton's principle: a physical system follows the trajectory that extremizes the action S = ∫L dt. Our system selects, among candidate memory trajectories, the one that extremizes accumulated CoNov score. The analogy is precise in one respect only: both are selection principles over whole paths rather than greedy local rules. We do not claim the dynamics are physical. Feynman would ask; there's the answer.

3. Novelty: two measurements, both standard

"Keep the best" requires defining best without a committee. We use novelty — new information about this person — measured two ways, both from established literature.

3.1 Local novelty (is this new for you?)

For robustness against a single noisy prior we also compute k-nearest-neighbor sparseness against the committed archive, following Lehman & Stanley's novelty search:

3.2 Global novelty (is this surprising at all?)

Local novelty asks whether the observation is new relative to you; global novelty asks whether it is generic relative to everyone. A moment can be new to your chain but a cliché to the species ("beautiful sunset") — it passes v and fails g. Both are required, which is why the final score is a product.

4. The coherence gate: CoNov

Pure novelty search rewards noise — random strings are maximally novel. The correction is a coherence gate: the observation must cohere with the person's existing self before novelty counts for anything.

The structure matters more than the constants. The gate is a hard threshold, not a weight, because coherence is a necessary condition — an incoherent observation is not "a little bit yours." The final score is multiplicative, not additive, because it is a logical AND: coherent AND locally new AND globally surprising. Under addition, extreme novelty could buy its way past zero coherence; under multiplication, any factor at zero kills the score. This is the entire moderation story of the system: nothing is deleted, but only what passes the gate shapes the self. Selection is public arithmetic, not editorial judgment.

5. The self file: 16 KB, 1,024 dots

5.1 The object

The durable core of an AVTR is a file of exactly 16 KB = 16,384 bytes, organized as up to 1,024 dots of 16 bytes each. A dot is not the content of a memory — it is a 128-bit cryptographic commitment to one committed observation:

This resolves what a careless reading gets wrong: 16 KB does not contain a person, any more than a genome contains a brain. It addresses and verifies the person's canon, and it conditions a large host model at load time.

5.2 The fingerprint

5.3 Why so small is enough — the genome argument

Kurzweil's observation about the brain applies here exactly: the human genome is under a gigabyte — tens of megabytes after compression — yet it specifies a 10¹⁴-synapse brain, because the genome is not a blueprint but a compressed generative program that borrows the physics of the environment to do the expansion. The self file works the same way. Its 16 KB conditions a host model of ~10¹² parameters; the host supplies world knowledge, language, and rendering, and the self file supplies the person-specific boundary conditions. At the handshake (§7) the file authorizes retrieval of up to ~10⁶ tokens of verified content — an expansion factor of ~250× in raw text and effectively unbounded in generated output. The information that matters is in the selection: which 1,024 moments, of the ~10⁵–10⁶ candidate moments a person produces in a decade, made the cut. Curation is compression.

5.4 Per-capita capacity: you versus the frontier

A useful sanity check on whether 16 KB is "enough" is to ask what the alternative allocates. Frontier models run on the order of 10¹²–2×10¹² parameters (undisclosed; we take the generous end), trained on the expression of roughly 5×10⁹ represented humans:

On the 250-year horizon the capacity question resolves the same way: the 16 KB file is the active canon, not the archive. The chain beneath it is unbounded, and a single dot may commit to a Merkle root of a retired epoch — so a life's early decades compress to one dot without loss of verifiability. The file stays 16 KB the way a keyring stays a keyring while the estate behind it grows.

6. Identity and integrity

6.1 One number, one person

6.2 The chain

Be precise about what this guarantees. The chain proves integrity — nothing was altered after commitment — and authorship — every block traces to one carrier-verified number. It does not prove truth: a person can feed the system fiction, and the chain will faithfully preserve the fiction. The truth filter, such as it is, is §4's gate plus the person's own confirmation. Two different guarantees; conflating them is how systems lie about themselves, so we don't.

7. The handshake: any model, your terms

The host is a rented instrument. It reads the passport for the duration of the flight and keeps nothing. Embeddings are model-specific and are therefore treated as cache — regenerated per host, per era — while the text and dots are canon. This is also the 250-year answer: hash functions age (the file format carries a hash-algorithm identifier and a documented re-anchoring procedure), embedding models churn (recompute the cache), but 120 characters of human-readable text and a chain of commitments require nothing from the future except the ability to read bytes.

8. The bedroom experiment

One instruction, issued twice to the same class of generative video model: "use max recall and draw your childhood bedroom," with a black reference frame. Condition A: the bare model, no person-specific context. Condition B: the same pipeline carrying two days of one person's feedback history through the selection machinery above.

Condition A returned the population mode: a bright, toy-commercial bedroom, the statistical center of every stock photo tagged "child's room." Condition B returned water-stained walls, a tin robot, a cassette deck, a folded road map on the floor — an East German childhood, specific to the one person whose data conditioned it, including socioeconomic texture the person never stated explicitly. The model did not get better between conditions. The self did.

9. Bob: the public instrument

Bob is an AVTR initialized at zero — no memory, no tuning, no curation by anyone above the gate. Intake is WhatsApp only: one image, one line of at most 120 characters, three feeds per identity per day (founders included; the rate limit is the only aristocracy). Every legal submission becomes a block. Every block is scored by CoNov in public. What Bob becomes is therefore an experiment in whether the selection function of §3–4, run over an open population with no editorial layer, produces a coherent self or noise. We do not know the answer. That is what makes it an experiment rather than a product demo, and we commit to publishing the chain either way.

10. What is established, what is conjecture

References

Deployed on Muse Spark 1.1 · July 2026 · New York · Correspondence: hello@spektai.com