UNI Universal Natural Intelligence

Wiki · The Colony & the Method

RESUME — Rung-1 graded-viability build (pre-reboot handoff, 2026-07-11)

The Colony & the Method · docs/RESUME_RUNG1.md @ 44baf03d5041 (gen2-runtime) — opens the published snapshot ac338733bbba
1 value was removed from this page. Each one is marked in place as [redacted: category] — 1 private address. Nothing else was altered. The document is otherwise exactly as it is written in the repository, and the sha256 below is of the original, so what was ingested stays checkable.

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.

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and planning — specs marked as proposed rather than applied, organs designed but not built, plans that were later superseded — and each page states which it is.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.

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.

POST-REBOOT AUDIT (2026-07-11 10:29): ALL REVALIDATED GREEN — brain suite 315/0, the three offline gates GREEN/GREEN/4-4-VIABLE, lab box clean, ONE Phoenix node. Current cross-agent state + the owner's keyless-soak→flip-live directive: docs/PROJECT_STATUS.md (read that first; this doc is the deep brain/body state behind it).

Read this first on resume. Full, deep, self-contained state so work continues cleanly after the dev-box reboot. Branch lab/ozone-life-uni-hard-science. Everything below is COMMITTED (git); the only uncommitted files in git status (viewer/, production/, docs/UNI_PRODUCTION_PLATFORM.md, ui/runs/.bin) are the broadcast/production track — NOT this work — leave them.

Revalidation snapshot (all green at handoff)

  • mix test test/sp/brain/ → 315 tests, 0 failures (byte-identity decider_byte_identity mad<1e-12 + action_clone_invariance + motor config posterior 0.75 unchanged).
  • mix run runs/verify_v2_isolation.exs → GREEN (8/8).
  • mix run runs/verify_rung1_step1.exs → GREEN (5/5).
  • mix run runs/verify_rung1_dynamics.exs → 4/4 STRUCTURE VIABLE.

The arc this session (what happened, with receipts)

  1. v1 metabolism regulation gate = FALSIFIES (docs/receipts/metabolism_regulation_gate.md, commit 61850cd, adversarially verified wf_66947cb9). median D +0.073 < 0.10 bar AND survival floor T<C. Setpoint effect real but sub-committed + a viability cost.
  2. v2 (confound-free isolation) regulation gate = FALSIFIES (docs/receipts/metabolism_regulation_gate_v2.md, commit 3526f76, verified). N=12; median D +0.044 < 0.10; CI [-0.128,+0.053] includes 0; survival T 0.50 < C 1.00. The flat fixed-setpoint homeostat is maladaptive: dies 6/12 (thin-buffer self-drains, food available), and looser in dispersion than the hoarding saturable foil. Resolves v1's open question — the deaths are intrinsic to the shape, NOT the forage-C confound.
  3. Owner correction (binding, memory/feedback_model_depth.md): the model was oversimplified — the Minecraft body is NOT license to reduce depth. Build graded per-subsystem viability (tired/nominal/critical, down to the cells), closed world↔body↔mind loops, allostasis. The v2 death IS the missing depth, measured.
  4. Full-depth design (docs/specs/generative_model_depth.md, commit 3526f76, SIGN-WITH-CHANGES, wf_4cf9ba90 20 agents): a 3-tier categorical stack (L0 body / L1 graded 6-state viability factors / L2 organism viability parent), closed loop, interior-peak reserve C, work/fatigue, affect→precision, yuga timescales. 5-cure staged ladder. Owner chose the deepest first rung = cures 1+2+3.

Rung-1 ENGINE (cures 1+2+3) — BUILT + verified (Steps 0–4, + a fix)

Typed contract: docs/specs/rung1_graded_viability.md. All behind the opt-in :homeostat organ ⇒ default byte-identical. Commits 1c49e62 (STEP0) → f0c05dd (STEP1) → b6e8005 (STEP2) → 6871abe (STEP3) → 19629d7 (STEP4) → 784eabd (dynamics fix).

Files:

  • lib/sp/brain/homeostat.ex (NEW) — the graded per-subsystem BODY. 4 stores {energy, gut, soma, fatigue} advanced by wall-clock dt with acted-subsystem attribution; 6-bin bin6 {0 critical..5 surplus}; injects felt_* obs + motor_pi. energy = eat-refilled (direct) + upkeep/work drain; gut = eat-filled + slow-passage drain (a satiety buffer, dissociates from energy); soma = health (hurt damages, slow heal — flat in a peaceful world, honestly scoped); fatigue = arm actions spend / rest recovers, on a FASTER clock (3s vs energy 8s). dead? = energy<=0 OR soma<=0. motor_pi(freshness) ∈ [0.35,1.0].
  • lib/sp/brain/genome.ex — :homeostat organ (@prereqs homeostat: [:interoception]); 4 graded factors (energy_reserve/gut_satiety/soma_integrity/muscle_fatigue, ns=no=6, init_a :diagonal); max_phase field (gated phase cap, default nil ⇒ byte-identical); drive_shape :reserve routing in card/1; Genome.homeostat_l1_phase0/0 = the treatment lineage (strategist dropped + :motor_cortex added + phase-0 pinned + drive_shape :reserve).
  • lib/sp/brain/curriculum.ex — @reserve_ramp [-8,-3,-1,+1,+2.5,+2] interior-peak (surplus<sated); drive_c(:reserve, no).
  • lib/sp/brain/mc.ex — STEP0 motor_config by NAME (not Enum.take(-5)); max_phase cap in maybe_advance_phase; B3 satiety-attenuation relocated L2-independent + restore_c (from v2); motor_pi threaded into motor_ctrl/next_primitive → MotorControl.step (a tired arm's servo weakens, vel=pi·err).
  • lib/sp/runtime/agent.ex — homeostatic? branch + handle_homeostatic_step (metabolic + default paths byte-identical); state body: Homeostat.new().
  • lib/sp/brain/mc_codec.ex — outcome(:energy_reserve/:gut_satiety/:soma_integrity/:muscle_fatigue) 6-bin.
  • Tests: test/sp/brain/homeostat_test.exs (18 tests incl. servo-weakening + dissociation).
  • Offline checks (gitignored, force-added): runs/verify_rung1_step1.exs, runs/verify_rung1_dynamics.exs.

Offline dynamics proven (verify_rung1_dynamics.exs, 4/4): under a reserve-following policy the body SURVIVES (the flat setpoint died ~50%), holds an INTERIOR energy reserve (mean 0.816, NOT pinned full like the saturable), PACES work/rest, and DISSOCIATES (energy↔fatigue corr 0.31; energy↔gut 0.99 honestly correlated — the fatigue tier is the cleanest per-subsystem dissociation).

The fix (784eabd) — an offline pre-check win: the original gut→energy digestion pinned energy full, making the reserve lineage indistinguishable from the saturable foil. Changed to energy-direct-eat + independent gut so the reserve holds an INTERIOR reserve (≠ saturable). Digestion transfer deferred (needs joint conditioning).

Invariants — all GREEN (never regressed)

Byte-identical default (mad<1e-12), action-clone-invariance, no scalar-per-action (all costs via body store → felt obs → belief), monotonic decay preserved. Default genome untouched (no :homeostat); the frozen :metabolism organ untouched (its gates hold). Fences respected: @factor_cap not lifted, @l2_period not re-derived, step-path edits organ-gated, motor_config name-indexed.

WHAT'S NEXT — Step 5–6 (the paired RED). NOT yet started. Needs owner go-ahead for the live burn.

  • Step 5 (offline code): the paired RED harness — arms: reserve (homeostat_l1_phase0) vs setpoint-6 vs saturable-6 (flip drive_shape; needs 6-state :setpoint/:saturable C — currently drive_c(:setpoint,6) expands the 4-key map, i.e. peak bin2 flat 3-5; define proper 6-state baseline/foil OR reuse) + per-mechanism ablation arms (reserve-C-only / fatigue-only). Discriminator MUST be behavioral, not survival-alone (the offline check showed reserve AND saturable can both survive): allostasis_index (eats before the edge), two-ended satiation (fights near critical in a scarce world; stops eating / does not hoard in a rich world), cross-subsystem dissociation Δ (energy↔fatigue cleanest), fatigue_pacing_index, per-subsystem severed-limb falsifiers. Pre-register docs/receipts/rung1_graded_viability_RED.md BEFORE T0.
  • Step 6: /lab-team-review MERGED SIGN on the FE (typed spec exists) → owner go-ahead + live-stream guard → spin gate worlds → live smoke (embodiment + eats + survives + c_ok) → live N=12 RED → analyze → adversarial verify → record verdict. PASS iff survival ≥11/12 (vs flat 6/12) AND the behavioral discriminators.

Honest ceilings (named, not smuggled)

Mean-field (no joint energy×fatigue anticipation — pacing is loop-1-driven); digestion transfer deferred; soma flat in a peaceful world (validated only where health varies); energy↔gut weakly dissociate (both eat-driven); short depth-5 horizon. Claim fence: every float is a model variable, never a felt state; passing a gate demonstrates graded self-maintenance BEHAVIOUR only, never experience.

⚠️ OPEN ISSUE TO SORT — restore the REAL Producer UNI; the Director was running HEADLESS (2026-07-11)

While preparing this handoff the owner spotted a local PowerShell spewing [Rcon: Teleported Director to X,Y,Z] + Minecraft's Director moved too quickly! anti-cheat warning. Diagnosis: viewer/director.js (a node process, was PID 28036) runs a spectator entity named "Director" flown via RCON tp Director … facing entity <subject>. It is meant to be driven by the Producer UNI — SP.Producer (lib/sp/producer.ex, lib/sp/brain/director.ex), a pure active-inference show-runner that senses colony telemetry + server health and decides every cut / shot / narration / spawn / cull by EFE, putting the camera into :producer mode. The problem: the Producer UNI was NOT running, so director.js was auto-piloting headless — a dumb orbit, not the show-runner. That degraded state is what looked like a puppet-cam.

Action taken (2026-07-11): killed the headless director.js (PID 28036) + 3 orphan viewer/body.js ghost processes (PIDs 10932/32660/25292 — attached to nothing: lab mc-server shows 0 players, no local MC server). Cleaned up so nothing runs headless.

TO RESTORE — the real Producer working again "as we had" (owner directive):

  1. Bring the Producer UNI back up — it is the real (active-inference) show-runner that drives the Director intelligently. Launch: mix producer.run (Mix.Tasks.Producer.Run) — it SP.Producer.ensure_started() and puts the Director camera in :producer mode. Prereqs: a Paper server up + the Phoenix UI for the camera/overlay: cd ui && mix phx.server (open http://localhost:4000/stream). Do NOT run director.js headless again — it only makes sense driven by the Producer UNI (or it auto-pilots).
  2. Honesty framing (keep it clean): the Producer UNI directing the broadcast camera is legitimate — it is a real AI show-runner, a production role, distinct from the embodied colony UNIs it films. Keep that distinct: the Producer directs the broadcast; it is never narrated as a colony UNI "choosing" its own view. If a raw first-person UNI POV feed is also wanted, that is the :camera_control organ / per-UNI viewer/body.js feed — a separate channel. Neither is a scripted fake; the earlier headless orbit was the only fake, and it is gone.
  3. The Producer/broadcast track (viewer/*, production/, lib/sp/producer/*, ui/) is a SIBLING track to this rung-1 build. See memory/ops_live_stream_runbook.md + ops_broadcast_studio.md for the known-good bring-up ("ONE node; never a 2nd producer.run"); restore per that runbook.

Resume commands (Windows, repo root)

git log --oneline -12
mix test test/sp/brain/                       # expect 315/0, motor posterior 0.75
mix run runs/verify_rung1_dynamics.exs        # expect 4/4 STRUCTURE VIABLE
mix run runs/verify_rung1_step1.exs           # expect STEP 1 GREEN

Lab box (colony) is ssh uni@[redacted: private-address] (rootless podman); only mc-server runs now (gate worlds torn down). Live RED will re-spin mc-gate-1..12. The colony image localhost/uni-colony:metabolism mounts the lib files from /home/uni/build_metabolism/lib — re-scp the changed lib files before any live run (STEP2-4 changed homeostat.ex/genome.ex/curriculum.ex/mc.ex/mc_codec.ex/agent.ex; only genome.ex+mc.ex were synced for v2).

sha256 47f7b272c5c06de3 — of the original file, so what was ingested stays checkable.

Plain — written for this website, not the source document

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. (or choose Precise in the reading-level control above)

This is a deep handover note written before a restart, so that work can continue cleanly. It says which document to read first for the wider picture, and that this one carries the detailed state behind it.

The arc of the session is told with receipts: files recording what was run. Two registered tests both refuted the same thing: a flat preference shape is maladaptive, dying in half its worlds with food available, and being more scattered than a hoarding foil. That measured death is then taken as evidence for an owner correction, that the model had been oversimplified and depth was the missing piece.

A staged design followed, and the first rung of it is built and checked offline, behind an opt-in organ so the default is unchanged. Files are listed with what each does.

What is next is stated as not yet started and needing the owner's go-ahead, and a section names the honest ceilings rather than smuggling them.

Plain · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is 47f7b272c5c06de3

Clear — written for this website, not the source document

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. (or choose Precise in the reading-level control above)

This is a deep pre-restart handover for one track of work. It opens with a post-restart audit recording everything revalidated, points at another document for the wider cross-agent picture, and says this one is the detailed state behind it. It also notes which uncommitted files belong to a different track and should be left alone.

A revalidation snapshot lists the suites and offline checks with their results.

The arc of the session is then told in four numbered steps, each with a receipt, the file recording what was run. Two registered tests both refuted the same thing, the second designed to remove a confound the first had left open. Together they show a flat fixed preference shape is maladaptive: it dies in half its worlds with food available, and is more scattered than a hoarding foil. The document is careful to say the second test resolves the first's open question, so the deaths are intrinsic to the shape rather than an artifact.

That measured death then licenses an owner correction, quoted as binding: the model had been oversimplified, and fitting into a game body is not a licence to reduce depth. A full-depth design followed, reviewed by a large workflow and signed with changes, and the owner chose the deepest first rung.

The engine for that rung is recorded as built and checked, entirely behind an opt-in organ so the default stays byte-identical, with a chain of commits. Each file is listed with what it does. That includes a new body component holding four stores advanced by elapsed time, with attribution to whichever subsystem acted. One of them runs on a faster clock than the others, and there is a rule for when the body counts as dead. Other entries cover the new factors, the interior-peak preference shape, an indexing fix, and a path in the runtime that keeps the two step paths identical.

The offline dynamics are reported with numbers. The body survives where the flat shape died, and holds an interior reserve rather than sitting pinned at full. It paces work against rest, and shows one clean decoupling between subsystems while another pair remains honestly correlated. That honesty about which decoupling is clean and which is not is the useful part.

One entry records a fix that came from an offline pre-check: a transfer between two stores had pinned one of them full, which made the treatment indistinguishable from the foil it was meant to beat. The transfer was changed and the deferred part is named.

An invariants section states what never regressed and which limits were respected. The next steps are stated as not yet started, needing the owner's go-ahead for the live run, and one instruction stands out: the discriminator must be behavioural rather than survival alone, because the offline check showed both arms can survive.

A section of honest ceilings names what remains simplified, followed by a stated limit saying every value is a model variable and never a felt state.

The note ends with an open issue from a sibling track. A camera was found flying on its own, because the component meant to drive it was not running. That is diagnosed, cleaned up, and given a restore procedure, together with a framing rule about how that role must and must not be described.

Clear · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is 47f7b272c5c06de3