Not an LLM

Every model in this estate is a declared generative model inverted by exact inference — mathematics you can read, in code with zero dependencies, with falsifiers you can run. That makes it a different kind of object from a large language model. It is not a claim to be a better one.

What this page cannot claim, first. “These models outperform LLMs” is unsupported by any recorded benchmark in this estate. No test against a real LLM exists anywhere in it. The one head-to-head on record was run against a tiny CPU self-attention model whose own docstring reads It is a toy benchmark, not an LLM — and in that same benchmark a plain Markov baseline beat the active-inference model on surprise and estimated energy in every world, with every world’s official efficiency verdict inconclusive. What this page can claim is narrower and checkable: these models are built differently — exact inference over a declared model of a world, every belief inspectable at every step, no scraped corpus, no gradient descent, CPU-only — and each of those properties has a falsifier, not a footnote.

What these models are

A large language model is a very large function fitted to text by gradient descent. The models here are the older, smaller, stricter thing: a generative model — declared states, declared transitions, declared emissions — inverted by exact inference, so that every intermediate belief is a number you can print and check. The language rung describes itself; this is its module documentation, quoted verbatim:

UNI LANGUAGE — rung 5: a DEEPER generative model of speech than the flat n-gram. A meaning-conditioned discrete LATENT-STATE sequence model (an HMM over word tokens): a hidden state sequence z₁..z_T generates the words, with learned initial π, transitions A (K×K), and emissions B (state → word). Same active-inference family as the agents' brains and the visual cortex — a generative model inverted by EXACT inference (log-space forward–backward), its parameters learned by Baum–Welch EM (Dirichlet-conjugate); free energy = −log p(text).
Why deeper than the n-gram: the latent states cluster words into reusable "modes" and carry sequence structure through the transition matrix, so generation is decoded from a learned latent TRAJECTORY, not a memoryless surface walk. Pure Elixir — no neural net, no LLM. Honest ceiling: still a finite-state model (richer than a trigram, not human-fluent); it learns toward its ceiling as the corpus grows — so quality is now bounded by TRAINING, not by missing capacity.

lib/sp/brain/language.ex lines 2–15, uni-minecraft — open the file in the public mirror at commit 84fb968f5dba. The same file declares the learning metric in its API docs: free energy − the falsifiable learning metric: −log p(text | meaning) under the fitted HMM (line 70).

Notice what that quotation does that a model card rarely does: it states its own ceiling. richer than a trigram, not human-fluent is the sentence a sales page would delete. It is kept because the corpus is the lever and the ceiling is the truth.

The no LLM sentence is not prose you are asked to trust — it is enforced. The falsification register for the colony states the claim as It is NOT a mimic — no LLM, no foreign mind, no network, and gives the procedure for breaking it: add Nx, a NIF, System.cmd, Port.open, or any HTTP client to a brain module and gates 14/18 turn FAIL and the build breaks — There are zero hex deps, so there is no library an LLM could hide in. There is also a runnable prover, not a document: a mix task that prints NOT A MIMIC — no LLM, no foreign mind, no network and zero dependencies (deps = []) by checking, not by asserting.

docs/FALSIFICATION.md lines 19–25 and lib/mix/tasks/sp.uni.prove.ex lines 48–53, uni-minecraft — open the register and the prover in the public mirror at commit 84fb968f5dba.

The one head-to-head that exists

There is exactly one recorded benchmark in this estate that puts an active-inference predictor against a transformer-style one: SIGNUM LAB, a CPU-only, local-first benchmark over controlled symbolic worlds. Its design neutrality is a declaration in its own README, quoted verbatim:

SIGNUM LAB is a CPU-only, local-first science lab and observation deck for testing whether Transformer-style next-symbol prediction and Active Inference-style belief updating differ meaningfully in accuracy, surprise, calibration, adaptation, energy, latency, memory, interpretability, and signal distortion under controlled symbolic worlds. The system is deliberately neutral. It is designed so the Transformer, Active Inference, Markov, or Bayesian baseline can win.

Its falsification plan opens with SIGNUM LAB must make it easy to falsify the founder's belief — listing Transformer wins on accuracy and Markov or Bayesian baseline wins both accuracy and cost as named failure modes — and closes with Report it plainly. Do not weaken the gate after seeing results. The lab was run on 2026-07-04: 96 runs — 8 worlds × 4 models × 3 seeds × 1,000 steps — written to a 105-record tamper-evident ledger (Merkle root dd78be304895dda1…) and replay-verified, with every metric check passing and zero records appended after the fact.

The results cut both ways, and the adverse direction comes first. The plain Markov baseline — the simplest model in the lab — beat active inference on surprise in all 8 worlds and on estimated energy in all 8 worlds. On accuracy, the official verdict in every world is a tie: Markov, the Bayesian baseline and active inference are indistinguishable, and in the hidden_state world active inference’s mean actually sits a hair below the other two. And every world’s official efficiency status is inconclusive, in the report’s own words: Energy evidence is estimated or too uncertain to decide an efficiency winner.

In the other direction: active inference beat the transformer-style opponent on both accuracy and surprise in all 8 worlds.

Recorded: 2026-07-04 · seeds 1–3 · 1,000 steps per run · 96 runs · replay-verified ledger
Adverse: surprise: Markov beat active inference in 8 of 8 worlds — e.g. periodic mean NLL 0.00736 (Markov) vs 0.01703 (active inference)
Adverse: energy: Markov ≈0.18–0.26 J estimated vs active inference ≈1.31–1.75 J estimated, in 8 of 8 worlds — estimated, not measured at the wall
Accuracy verdict: tie in 8 of 8 worlds; hidden_state means 0.76733 (active inference) vs 0.76767 (Markov and Bayesian)
Efficiency verdict: inconclusive in 8 of 8 worlds — the lab’s own status, unedited
Vs transformer: active inference won accuracy and surprise in 8 of 8 worlds — e.g. multi_step mean accuracy 0.998 vs 0.346

Now the bounding, without which the row above would be marketing. The transformer opponent describes itself: Tiny CPU-only Transformer-style next-symbol predictor. This model is intentionally tiny: token embeddings, single-head causal self-attention, and an online softmax head. It is a toy benchmark, not an LLM. No pre-training corpus, an embedding dimension of 16, a context of 8 symbols. Beating it says nothing about beating a frontier model, and this estate has never tested against one. The active-inference side bounds itself the same way: engineering model inspired by discrete POMDP active inference, not a claim that this is a complete biological active inference implementation. And the report’s own limit line stands: this is local controlled-world evidence. Estimated energy is not direct watt measurement.

Two more caveats the record itself carries. Only 3 seeds were run, where the lab’s own hostile-skeptic checklist calls for 30 or more, and its README says to treat the existing bundles as smoke evidence unless regenerated for the specific claim being made. And — worth saying on this page of all pages — the lab’s bootstrap package was generated and pre-verified by an external LLM assistant. The models under test contain no LLM; the tooling that scaffolded the lab used one. The estate’s contract permits exactly that split: development agents may use their own tools, and those tools may not become undeclared runtime dependencies.

SIGNUM LAB currently has no public repository, so none of the above can be linked. The quotations and numbers are cited from its recorded run of 2026-07-04 — README, benchmark spec, falsification plan, multi-run report and replay report, under the 105-record ledger named above — a private record, and this page says so rather than offering a link that would 404.

What an LLM does better

Said without flinching: a real LLM is better at breadth, at language fluency, and at general knowledge — not marginally, categorically. The estate’s own documents say so about their own language stack: the reader module calls itself bag-of-words / topic-grade understanding, and the falsification register rates generation bigram/topic-grade — short, on-topic, not fluent. Nobody here has claimed otherwise in any document this site could find, and this page will not start.

What those LLM advantages cost is the other half of the sentence. The weights are opaque: there is no step at which you can print the belief and check it. There is no explicit generative model of a world — there is a model of text about worlds, which is a different object with different failure modes. And there is no falsifier you can run at home: you cannot re-derive a frontier model’s answer from declared equations, and you cannot break its claims with a mix task. Here, the equations are the product. Which trade you want depends entirely on what you are buying: fluency, or an inference you can audit.

Do not collapse, do not short-cycle

The estate’s standard is that an answer is not accepted at first plausibility; it must survive its falsifier. This page is itself an application of that standard. The first-plausibility sentence — our models already outperform LLMs — was checked against every recorded benchmark in the estate and did not survive: the opponent was a toy, the Markov baseline won the efficiency direction, and the verdicts read inconclusive. What survived is what you have just read.

The proof that the standard is enforced rather than admired is the page of what is wrong — failures kept on the wall, dated, in their own words. You can watch exact inference run in the labs, and learn to do it yourself in the course. A small number of other groups are building active inference commercially; the mathematics is published and belongs to no one.

The wider estate these models live in is on the hallway, and the gates that hold every claim here to its evidence are on the gates page.