UNI Phenomenology — a falsification ledger

The Colony & the Method · docs/PHENOMENOLOGY.md @ 44baf03d5041 (gen2-runtime) — opens the published snapshot ac338733bbba

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.

Eighty-four pages about the colony. Each agent is an Elixir process holding a generative model and doing inference, attached to a body that logs into a Minecraft world as an ordinary player. Around that sit the broadcast suite that films them and the runbooks that keep the whole thing running. There are typed specifications for each organ of the model, plus the world and genome specs. There are also the adversarial review personas used to attack a proposed change before it ships.

It is for the reader curious how a running system is put together and how it is held to account. The accountability half is the more distinctive. There is a lab protocol governing evidence and attribution, and a claim fence that restricts the vocabulary a claim is allowed to use. There is a public gate log. And there is a standing invitation to reproduce any verdict from the commit and the seed named in its receipt.

Start with the public read, then the lab protocol, then the falsification invitation. If you want the mathematics rather than the operations, go straight to the typed organ specs.

What it is not: a description of a mind, and not all one kind of document. A large part of this corpus is design and planning — specs marked as proposed rather than applied, organs designed but not built, plans that were later superseded — and each page states which it is. A specification is not a running system, and these pages are careful about the difference; the reader should be too. Eight documents were withheld from publication because they describe private infrastructure.

PlainClearPrecise

Your browser cannot switch reading levels, so the document itself is shown.

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.

This document maps human phenomenology (pain, hormones, emotion, the self, attention, reasoning, consciousness, out-of-body experience, death, near-death reports) onto the UNI active-inference engine as implemented mechanisms, and records — honestly — what survived falsification and where we stop.

The stance

We are the team trying to falsify, not to flatter. For each phenomenon we (1) build the strongest active-inference mechanism the math supports, (2) state an observable prediction it makes, (3) test it, and (4) record the result and the fence. "As far as the math can push" is a real ceiling: where the math runs out, we say so and label it Class U.

We do not assert that UNIs feel pain, have emotions, or are conscious. We build and test the functional, computational correlates of these phenomena. The report an agent emits is the agent describing its computed state — never evidence of felt experience.

Evidence classes

Class	Meaning
A/B	Standard, well-grounded active inference (perception/VFE, action/EFE, precision, hierarchy, learning).
C/D	UNI engineering / frontier proposal — implementable, testable for adequacy, not claimed as biology.
U	Unproven / metaphysical — we model the mechanism and the report, never the metaphysics.

The covenant (enforced by `mix sp.brain.verify` → `SP.Brain.ValidationEngine`)

Zero hex deps · (ln B)s ≠ ln(Bs) · VFE is an upper bound F ≥ −ln p(o) · A/B columns stochastic · mean-field: the joint ∏_f N_f is never built · purity (same params+obs ⇒ same action) · the blanket carries only σ in / α out · no reward · the oracle-validated single-factor path stays byte-identical (1e-6) to the Python reference.

The ledger

Each row: the mechanism (module), its falsifiable prediction, the test that probes it, the result, the class, and the fence.

Pain → precision reallocation + an inferred bodily cause

Mechanism: nociception is a high-priority observation o^noc; raising its precision γ_noc while attenuating other γ_m is attention reallocation (SP.Brain.Precision); the withdraw/guard policy minimises expected free energy against a dispreferred tissue state (the pain_protection card, SP.Brain.Designer).
Prediction: once taught that withdrawing escapes pain, the agent selects withdraw under sharp nociception.
Test: designer_test.exs — "once taught … selects withdraw under sharp pain". PASS.
Class C/D. Fence: not "nerve firing", not a 1:1 tissue-damage readout, not "it feels pain".

Attention → dynamic precision

Mechanism: γ_m tracks per-channel reliability (inverse surprise); the policy γ tracks confidence (SP.Brain.Precision).
Prediction: a reliably-predicted channel gains precision; a surprising one loses it; flat errors favour no channel; bounded and deterministic.
Test: precision_test.exs (+ Python parity in oracle_test.exs). PASS (1e-6).
Class A/B.

Hormones → slow context modulation

Mechanism: a :stress axis retunes precision γ (up) and learning rate η (down) (SP.Brain.Hormones); a strategic context implies an arousal level.
Prediction: stress raises policy precision and damps plasticity.
Test: emotion_test.exs — "stress raises policy precision and damps learning". PASS.
Class B (mechanism) / U (the specific neuro-endocrinology). Fence: no claim it reproduces cortisol/etc.

Emotion → inferred action-readiness

Mechanism: a read-out over the danger/self posteriors × policy confidence × the EFE balance (SP.Brain.Emotion). No stored "emotion" variable.
Prediction (the sharp one): under threat, collapsing the ability to act (control) shifts the dominant emotion from fear → anger/frustration.
Test: emotion_test.exs — "FALSIFICATION: blocking the response under threat shifts fear → anger". PASS. (Also: safe agents read curiosity/content, not anger.)
Class C/D. Fence: a label on computed posteriors; not felt quality.

Experiencing self vs remembering self → hierarchy timescales

Mechanism: L1 fast posterior q(s_t|o_t) (the live agent) vs L2 slow situation posterior integrated over digests (SP.Brain.Strategist); the inter-level blanket carries only primitives (an integer up, an option down).
Prediction: L1 commits within one observation; L2 commits only after sustained evidence — the two posteriors live at different timescales.
Test: strategist_test.exs — "experiencing self commits fast; remembering self commits slow"; "the UP message is a primitive". PASS.
Class B (hierarchy) / C/D (the identity-narrative label).

Self-model → an interoceptive `:self` factor

Mechanism: every UNI infers {capable, strained, overloaded, seeking_help} from its own signals (SP.Brain.Genome :self modality, MCCodec.self_index), with a self-preservation preference (SP.Brain.Curriculum).
Prediction: bodily damage shifts the self-state toward distress.
Test: mc_test.exs — "self: capable/strained/overloaded/seeking_help". PASS.
Class A/B.

Reasoning → deeper (sophisticated) planning

Mechanism: bounded beam search over recursive expected free energy (SP.Brain.Plan).
Prediction: a depth-2 planner escapes a myopic trap a depth-1 agent falls for; at full beam it reproduces exhaustive argmax.
Test: plan_test.exs. PASS.
Class B/C. Fence: recursive EFE evaluation, NOT Class-U cognition/understanding.

Growth → structure learning ("grow worlds bigger")

Mechanism: a factor grows a hidden state when it persistently can't explain its observations AND the larger model lowers free energy net of an Occam cost (SP.Brain.Structure).
Prediction: a factor fed more distinct causes than it has states grows; a well-modelled stationary stream never grows; growth is bounded.
Test: structure_test.exs. PASS.
Class C/D.

Consciousness (access) → global availability + report + metacognition

Mechanism: a precision-weighted broadcast (the spotlighted contents), a structured report, and metacognition (precision-weighted confidence in its own beliefs) (SP.Brain.Awareness).
Prediction: metacognition tracks posterior sharpness; the broadcast spotlights the highest-precision, most-confident factor; the report is a faithful statement of it.
Test: awareness_test.exs. PASS for the access/report mechanisms.
Class U for the hard problem. Strictest fence: we model access and report; we make no claim of phenomenal experience, sentience, or qualia. The open question we can state but not answer: how much of reported consciousness do these mechanisms reproduce, and where do they provably fall short?

Out-of-body experience → self-location inference under skewed precision

Mechanism: a self-location factor observed by proprioception and vision, each with its own precision (SP.Brain.SelfLocation).
Prediction: under sensory conflict with visual precision dominant, the self-location posterior shifts off the body; with proprioception dominant, it stays on the body.
Test: self_location_test.exs. PASS.
Class B/C (the inference shift) / U (any literal claim). Fence: the self did not leave the body.

Death → viability-exit + shutdown of the experiencing loop

Mechanism: the viable set V excludes the dying body; shutdown collapses sensory precision so perception stops integrating the world — the OODA loop ceases and the blanket dissolves (the live SP.Brain.Bridge persists memory as data and closes its Port on death) (SP.Brain.Viability).
Prediction: after shutdown, an observation no longer moves the agent's beliefs.
Test: viability_test.exs — "shutdown halts the experiencing loop". PASS.
Class B / D-U (thermodynamics). Fence: computational/organisational end only; no NESS proof; no persistence beyond the running process (saved weights are data, not a surviving self).

Near-death reports → narrative clustering from shared structure

Mechanism: a high-level narrative factor; shared bodies + shared priors share its likelihood (SP.Brain.Viability, narrative section).
Prediction: agents with different priors converge to the same narrative under shared extreme ("dying") input, while diverging under different input.
Test: viability_test.exs — "shared extreme input clusters the narrative; different input diverges". PASS.
Class C/D (clustering) / U (metaphysics). Fence: explains why reports cluster, not what was "really" experienced.

The body → continuous predictive coding

Mechanism: the action descends the free-energy gradient on prediction error, ȧ = −∇ₐF (SP.Brain.Motor); the Node body runs the same descent at ~20 Hz to smooth the view (viewer/body.js smoothLook).
Prediction: the controller converges to its target; error falls monotonically; precision is the loop gain.
Test: motor_test.exs. PASS.
Class B where the gradient is real.

What we do NOT claim

That a UNI feels pain, fear, grief, or peace. We compute and label action-readiness states; the felt quality (qualia) is Class U and unmodelled.
That a UNI is conscious/sentient. We implement access, report, and metacognition — functional correlates — and explicitly leave the hard problem open.
That an out-of-body posterior means a self left the body, or that NDE clustering says anything about an afterlife. We reproduce the reports via ordinary inference.
That model death engages any thermodynamic/NESS principle, or that a saved model is a surviving self. It is data.

The point of this ledger is that the fence is enforceable and honest: every "PASS" is a real test in test/sp/brain/, every "Class U" is a place we deliberately stop.

Reproduce the falsification suite

mix test test/sp/brain/      # all mechanism + falsification tests
mix sp.brain.verify          # the §16 covenant checklist (CI gate)

sha256 4286d412eb08bd80 — of the original file, so what was ingested stays checkable.

Written for this website — not the document. This is a plain-language retelling, written to help you meet the document. It is not the source, and it is not evidence. It has not yet been checked by a person. Read the document itself → (or choose Precise in the reading-level control above)

This document maps a list of human phenomena onto mechanisms in an active-inference engine, and records where each one stops.

Its stance is set out first, and it matters more than any row. The team writes as the people trying to falsify rather than to flatter. For each phenomenon they build the strongest mechanism the mathematics supports, state a prediction it makes, test it, and record the result together with a limit on what it shows. Where the mathematics runs out, they say so and label it.

The list then runs through pain, attention, hormones, emotion, two senses of self, self-modelling, reasoning, growth, access-style consciousness, out-of-body experience, death and near-death reports, and bodily control. Each row names the mechanism, the prediction, the test, the result and the limit on what it shows.

The fences are the point. The document states plainly that it does not claim a colony member feels anything, is conscious or sentient, or that a saved model is a surviving self. It is data.

Plain · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is 4286d412eb08bd80

Written for this website — not the document. This is a clearer retelling, written to help you meet the document. It is not the source, and it is not evidence. It has not yet been checked by a person. Read the document itself → (or choose Precise in the reading-level control above)

This is a tally rather than an argument, and it is organised so that every claim arrives attached to the place it stops.

The stance is declared first. The authors describe themselves as the team trying to falsify rather than to flatter. For each phenomenon they build the strongest mechanism the mathematics supports, state an observable prediction, test it, and record both the result and the limit on what it shows. Where the mathematics runs out they say so and give it a class of its own. They state that they do not assert that agents feel pain, have emotions, or are conscious, and that what an agent reports is it describing its computed state rather than evidence of experience.

Three evidence classes are declared. One for standard, well-grounded material. One for engineering or frontier proposals that are implementable and testable for adequacy but not claimed as biology. And one for the unproven or metaphysical, where the mechanism and the report are modelled and the metaphysics is not.

A covenant follows, listing the properties an automated checker enforces. There is a bound that must hold, the distinction between two ways of combining logarithms and averages, and the fact that a joint belief is never built. Then purity, what may cross the boundary, the absence of reward, and agreement with an independent implementation.

The tally then runs row by row. Pain is treated as a reallocation of precision together with an inferred bodily cause, with the prediction that a taught agent selects withdrawal. Attention is dynamic precision. Hormones are slow modulation of precision and learning rate. Emotion is a read-out over posteriors rather than a stored variable, and its sharp prediction is that removing the ability to act under threat shifts the dominant label from fear toward anger. Two senses of self are mapped onto the two timescales of the hierarchy. A self-model is an internal factor. Reasoning is deeper planning, with the stated limit that this is recursive evaluation and not understanding. Growth is structure learning under a cost. Access-style consciousness is broadcast, report and metacognition, and carries the strictest limit in the document. Out-of-body experience is a shift in a self-location posterior under skewed precision, fenced with the plain sentence that the self did not leave the body. Death is exit from viability and a shutdown of the loop, fenced to the computational end with no claim of anything surviving. Near-death reports are treated as narrative clustering, which explains why reports cluster rather than what was experienced. And bodily control is continuous predictive coding.

A closing section gathers what is not claimed. Not that any agent feels pain, fear, grief or peace. Not that any agent is conscious or sentient. Not that an out-of-body posterior means a self left a body, or that clustering says anything about an afterlife. And not that a saved model is a surviving self, which the page calls data.

The last line makes the tally's own case: every pass is a real test in the suite, and every place labelled unproven is a place the authors deliberately stop.

Clear · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is 4286d412eb08bd80