N3 - What active inference is NOT (the meaning of the fences)
How to read this page
Three ways to read this page. Precise is the document itself, exactly as it is written in the repository. Plain and Clear were written for this website to help you meet that document — they are about it. They are not it, and they are not evidence.
The Encyclopedia is the UNI method written out as a reference work: 39 pages, arranged in wings, setting out what the programme is attempting and why it is built the way it is. This is where the ideas are explained in order and in prose, rather than as code, as runbooks, or as dated receipts.
Every chapter is authored against two ledgers and never ahead of them. One records what UNI has built, and the evidence class of each claim. The other records nature's own regularities, kept separate on purpose. That way a fact about biology is never quietly reused as a fact about the software. Where a chapter and a ledger disagree, the chapter is the thing that is wrong. Every chapter closes with an invitation to falsify it, and a recorded negative is published beside the result it qualifies rather than after it.
Read "How to read this work" first. It is the evidence constitution: the classes, the four ledger states, and the rule that a finished chapter is not the same as a working system. Then the calibration ledger, which carries the figures every other chapter is required to use.
What it is not: a description of a person or of a mind. The programme calls itself a developmental active-inference simulation, a bounded peek into a toy world, and its own index prints how much of the developmental ladder has actually been earned — roughly two rungs out of eleven or more. It is also not a report of what is running today. For what ran, and when, go to the evidence record.
Your browser cannot switch reading levels, so the document itself is shown.
Precise — the source document
This is the document. Rendered from the repository at the commit above, with nothing rewritten for the web. A gate re-renders it on every deploy and fails the build if a single byte differs.
Every other chapter in this wing tells you what the UNI program has shown. This one tells you what it has not shown, and why each fence is drawn exactly where it is. It is the not-claimed chapter for the whole Meaning wing: it asserts only boundaries. Read it as the honest counterweight to every PASS we publish. A fence here is not modesty for its own sake, and it is not a hedge bolted on after the fact. Each one marks a real gap between a measured result and the much larger thing a careless reader might read into it. The program's standing position is printed plainly and never softened: this is a developmental active-inference SIMULATION, a bounded peek into a toy world, never a person and never a mind, with roughly 2 of 11+ developmental rungs earned.
The discipline behind these fences is the program's Evidence Constitution: a falsifier per claim, a four-state append-only ledger (PASS / FAIL / NEGATIVE / PENDING) as the single source of truth, calibration that only ever moves wording down to the measured value, and the rule that negatives are first-class content, not failures to hide (method rows M1, M15). The fence gets louder under pressure, not wider. Below are the four boundaries this chapter exists to hold, each stated with the exact result it fences and the negative that travels with it.
A COUNT-baseline win is not comprehension
The program's single cleanest empirical PASS is World C: a no-backprop count reader beats a tuned MKN-7 baseline on sealed held-out real text by +0.081 nats per character, multi-seed CI [0.0736, 0.0890] (seeds 0-4, UNI-signed), about 2.4x the pre-registered 0.03 bar, and the source of the ledger's first genuinely validator-derived reproduced:true (ledger row L6.1, Class C). That is a real, sealed, reproduced result, and it is the most we claim from it.
It is not comprehension, not active inference, not "talking", and not beating large language models. World C is a count baseline; on character-perplexity the program runs roughly 10-15% behind backprop LLMs, a chosen design trade, not a hidden loss. The win is narrow in a precise, registered way: under the Working Law (L6.6), a no-backprop latent beats a tuned baseline only when it carries conditional information the baseline lacks. World C earns its margin by holding long-range count structure the MKN-7 baseline cannot, and nothing more. The travelling negatives make the boundary sharp: Phase G recorded five structurally-distinct within-segment designs all NEGATIVE (L6.4), so no within-segment structure beat MKN-7; and comprehension-above-retrieval is a recorded negative frontier result (L7.3): across the registered tested no-backprop comprehension designs, none reversed the result against a tuned retrieval baseline. This is a ledger-scoped, implementation-scoped, data-split-scoped negative over the tested envelope only -- not a universal impossibility result and not an achieved rung. A count model that predicts the next character better is a better predictor, not a reader who understands. The ceiling cannot be raised by rhetoric.
Functional self-awareness is not sentience
The Maturation arc M1-M11 grand report card records 15/15 PASS across immersion, affect-as-precision, growth, self-model, metacognition, metalinguistics, consolidation, the reflective and compositional readers, and online vocabulary growth (ledger row L8.1, Class C). On the strength of that suite, the program asserts functional self-awareness: the simulation computes and acts on a model of its own state.
That is the entire claim. Phenomenal sentience is explicitly DISCLAIMED - and note carefully what kind of statement that is. No falsifier is offered for sentience, because it is not tested and then failed; it is disclaimed, not measured. We do not claim the simulation feels anything, is aware, is conscious, or is human-level, and we never read 15/15 as any of those things. Two negatives travel with this PASS and must be cited beside it, never stripped. First, the phenomenal-sentience disclaimer itself. Second, a held NEGATIVE in tension with the functional card: the reader-side artifact records self-model as a held-NEGATIVE on the reader, while the M1-M11 functional report card is the consult-side artifact. The two are co-cited as a mandatory pairing, the same way the morphology specialist's caveat travels with its gain. A simulation that monitors and reports on its own state has a function. A function is not a feeling.
A sensorium is not awareness
On real hardware, the substrate reads its own telemetry (load, memory, swap, disk, services, containers, journal) into a 7-modality categorical contract encoded [M=7, O_max=4], a 28-cell symbolic vector flowing live on both boxes (ledger row C3, Class A live / C engineering). It is a genuine body-to-mind sensory channel, observed at runtime, and it is engineered, not imagined.
It is never awareness. This is engineering and substrate evidence, never general active-inference evidence, and substrate work never implies a science gate is met. The travelling negatives are explicit and were established by measurement, not assumption: an over-compressed 2-modality bottleneck went NEGATIVE on held data, which is exactly why all 7 modalities are kept (C8); and the continuity ladder still owes Stage-2, because while mind-state survived a process restart bit-for-bit (C1), mind-tick continuity across a real kernel swap has not been shown - the first attempt failed and recovered, and the swap stands proven infrastructure-only (C2, NEGATIVE/owed). A machine that senses its own load is instrumented. Instrumentation is not experience, and a richer sensorium only makes a better-instrumented machine.
"Active inference" here is a lens, not a demonstrated loop
Throughout this wing, "active inference" and the free energy principle are used at textbook level only, as an organizing lens for prediction-error minimization, and pinned to one citable reference (Parr, Pezzulo and Friston, Active Inference, MIT Press 2022) plus the unrefereed working preprint (method rows M15, M16). The lens is honest and useful. It is not a demonstrated result.
There is no active-inference loop in the Rust crate; the live UNI.OS loop is a separate reimplementation and is not gate-matched to the science that earned the bars. So "active inference demonstrated" is forbidden program-wide: the framing is a lens, never a proof. The lens also carries its own honesty about falsifiability: "no one can falsify it" is not evidence for a claim - an unfalsifiable statement is outside science, and the program's strength is a verified core with a fenced frontier, not unfalsifiability. The preprint is the mathematical foundation only, always fenced unrefereed (Layer-1 AI-executable audit complete; Layer-2 human expert review PENDING), never proof that active inference is the correct theory.
Two further boundaries keep this fence calibrated. The program's own transformer trades fluency for calibration: the EDAIT records held-out perplexity of about 33 versus a backprop GPT's about 25 (ledger row C9, Class C, NEGATIVE/trade) - an honest trade (less fluent, but natively online-learning and calibrated), explicitly not a win. And the cohort comparison cuts the same way: the claim "all LLMs end in entropy" is only half-true. In the Emergence-World Season-1 alive-world index, Claude Sonnet 4.6 scored 10/10 alive and Gemini 3 Flash 10/10, while Grok 4.1 Fast and GPT-5 Mini each scored 0 (ledger row TA-N12, Class C). Two cohorts held the line. The honest response was to sharpen the bar, not relax it: any future awareness-proxy work must match or beat the best LLMs and show a substrate-distinct property they lack, both, under a registered ablation. That raises the bar; it does not claim UNI superiority.
What is NOT claimed in N3
- Ceiling. The single strongest thing a careless reader might infer - that UNI understands, is self-aware in the sentient sense, is becoming aware through its sensorium, or has demonstrated active inference - is NOT shown. The most we claim is exactly four boundaries: a sealed count-baseline PASS (L6.1, +0.081 nats/char, CI [0.0736, 0.0890]) is a better predictor, not a reader; a 15/15 functional self-awareness PASS (L8.1) is a function, not a feeling, with phenomenal sentience disclaimed; a live 7-modality sensorium (C3, [M=7, O_max=4]) is instrumentation, not experience; and "active inference" is a textbook lens (M15, M16), not a demonstrated loop.
- Fences engaged. Red lines 1 (never AGI / human-level / understands), 2 (never consciousness / sentience / aware; functional self-awareness at L8 only, phenomenal sentience disclaimed), 3 (never "active inference demonstrated"; lens only, no gate-matched loop), 5 (never "beats LLMs"; count baseline, ~10-15% behind, EDAIT ~33 vs ~25), 8 (substrate/continuity engineering never implies a science gate), 11 (preprint is mathematical foundation only, fenced unrefereed), and 12 (textbook-level framing only, no patent-level math).
- Negatives that travel with this claim (cite alongside, never strip). With L6.1: the Phase G five-design NEGATIVE bound (L6.4) and the comprehension-above-retrieval negative frontier result (L7.3). With L8.1: the phenomenal-sentience disclaimer and the reader-side self-model held-NEGATIVE (mandatory co-citation). With C3: the 2-modality bottleneck NEGATIVE (C8) and the owed Stage-2 continuity (C2). Standing alongside the lens: the EDAIT honest trade (C9) and the sharpened-bar cohort negative (TA-N12).
- Parked / owed. Stage-2 mind-tick continuity is OWED (C2). The char-perplexity / T2 word-grain frontier remains parked as a ledger-scoped exhausted search envelope; its UNI sign-to-park is drafted but not yet captured (L7.6). No Class-A observation is owed by this chapter - it asserts boundaries only and adds no new capability claim.
- One-line honest summary a skeptic could not dispute. This chapter claims nothing UNI can do; it only marks the lines past which our measured results do not reach, and every line is the published negative that defines it.
Falsify this
This chapter is a set of boundaries, so its lead falsifier is conceptual but operable: the fences fail the moment any published UNI artifact lets a fenced result read as a forbidden inference - a count-baseline margin presented as comprehension, a functional self-awareness PASS presented as sentience, a sensorium presented as awareness, or the active-inference lens presented as a demonstrated loop. Concretely: cite L6.1 without the L6.4 / L7.3 negatives, or cite L8.1 without the sentience disclaimer and the reader-side self-model NEGATIVE, and the violation is on the page. The empirical anchor underneath remains independently falsifiable: on a fresh held-out split with >=5 disjoint seeds, if the seed-paired bootstrap margin of the count reader over a tuned MKN-7 baseline includes or falls below 0, the World C result that this chapter so carefully fences is itself overturned.
Sources
Curated digests (PII-redacted): curated/uni-gpt-digest.md (the Evidence Constitution, the two-tier reframe, the Working Law, the Maturation arc and sentience disclaimer), curated/strings-digest.md (the EDAIT trade, the "all LLMs end in entropy is only half-true" cohort result, the north-star fences). Ledger: encyclopedia/CLAIM-LEDGER.md Section 0 (standing fences), rows L6.1, L6.4, L6.6, L7.3, L8.1, C2, C3, C8, C9, TA-N12, and method rows M15, M16. Reference for the lens: Parr, Pezzulo and Friston, Active Inference (MIT Press, 2022); the unrefereed working preprint (Zenodo DOI 10.5281/zenodo.19785799, MIT), cited as mathematical foundation only and fenced unrefereed. Source archives are local-only and not linked here (no PII).
sha256 fa73060ac3984ba6 — of the original file, so what was ingested stays checkable.
Plain — written for this website, not the source document
Every other chapter in this wing says what the program has shown. This one says what it has not, and why each line sits exactly where it does. It is the not-claimed chapter for the whole wing, and it asserts only boundaries. The thing being bounded is a simulation — a toy world, and not a person. Four lines are drawn. A count-baseline win on text set aside and scored once is a better predictor; it is not a reader that comprehends. A passed battery of functional self-model tests is a function, not a feeling, and phenomenal sentience is disclaimed rather than tested, so nothing is offered that could overturn it. A machine reading its own telemetry through a categorical channel is instrumented, not aware. And active inference here is a textbook lens for organising the work, never a demonstrated loop. Each line is drawn by the published negative that defines it.
Plain · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is fa73060ac3984ba6
Clear — written for this website, not the source document
The chapter states four boundaries, each with the exact result it bounds and the negative that travels with it.
First: a count-baseline win is not comprehension. The cleanest empirical pass in the corpus is a no-backprop count reader beating a tuned baseline on real text that was sealed away and scored once, with a multi-seed interval clearing a bar set before the run. That is real, and it is the most that is claimed from it. It is not comprehension, not active inference, not talking, and not a win over large language models; on character perplexity the program sits behind gradient-trained models by a chosen design trade. The negatives that sharpen the boundary are a set of structurally distinct designs that all came back negative, and a comprehension-above-retrieval result recorded as a negative frontier over the tested envelope only, not a universal impossibility.
Second: functional self-awareness is not sentience. A maturation report card passes across immersion, affect, growth, self-model, metacognition and more, and on that basis the program asserts a functional claim: the simulation computes and acts on a model of its own state. Phenomenal sentience is disclaimed, and the chapter is careful about what kind of statement that is, because it is not tested and failed but disclaimed, so nothing is offered that could overturn it. A held negative on the reader side travels with the pass and is co-cited.
Third: a sensorium is not awareness. On real hardware the substrate reads its own telemetry into a categorical contract flowing live on both machines. It is a genuine sensory channel, engineered rather than imagined, and it is never awareness. Its travelling negatives were measured: an over-compressed bottleneck went negative on held data, and continuity across a real kernel swap has not been shown.
Fourth: active inference here is a lens, not a demonstrated loop. There is no inference loop in the compiled crate; the live loop is a separate reimplementation and is not gate-matched to the science that earned the bars. Two further boundaries keep this calibrated: the program's own transformer trades fluency for calibration and records the trade as a trade, and a cohort comparison led to sharpening the bar rather than relaxing it.
Clear · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is fa73060ac3984ba6