UNI Universal Natural Intelligence

Wiki · The Colony & the Method

Spec — Survival-C (curriculum removal), the deepest/slowest layer

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

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and planning — specs marked as proposed rather than applied, organs designed but not built, plans that were later superseded — and each page states which it is.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

Part I of A4. Reads with generative_model.md (backbone) + sensorium.md (Part II). Design-only; ship gate = formal /lab-team-review + owner go-ahead. Corrections folded from docs/receipts/a4_lab_team_review.md.

Hypothesis (to prove, not assume)

With C grounded ONLY in viability + a real death edge (the metabolism emptying-B) + a natural epistemic drive, UNI keeps its body viable and acts without any curriculum or goal-setting. Phase-2 showed the curriculum was a confound; removing it lets us ask the pivot cleanly. OPEN question.

I.1 StateSpace + the CORRECT setpoint shape (blocker #9)

No new factors. Survival-C reuses the existing action-independent survival factors and changes only their C: status (viability edge), threat, @self_pref, @social, and — via :metabolism — energy/satiety. Correction (the earlier "flat-top F8" claim was wrong): @energy_setpoint/@satiety_setpoint (curriculum.ex:33-34) = %{0=>-8.0, 1=>-2.0, 2=>+3.0, 3=>0.0} is a single-peaked interoceptive setpoint: peak at "ok" (bin 2, +3.0), neutral at "full" (bin 3, 0.0), steep penalty toward "empty" (bin 0, −8.0) — an inverted-U, NOT a flat top. It is non-saturable because there is no monotone-increasing pragmatic value in over-filling (bin 3 < bin 2, so stuffing past "ok" is mildly dispreferred); the only standing gradient is away from depletion. F8 falsifier restated against the true shape: if any survival-C factor's C is monotone-increasing in "more" (a more-is-better reward), it is a preference-hack and is struck. (A genuine flat-top variant would set bin3=bin2; the current shape is stronger non-saturability than a plateau.)

I.2 PreferenceModel — phase-independent survival table

Replace the phase-indexed Curriculum.preference(phase, modality, no) (curriculum.ex:47-51) with a phase-independent survival table for curriculum: :survival_only lineages: return the viability vectors for status/threat/self/social/energy/satiety at all times, and all-zeros (neutral) for every task modality (inventory/vision/sky/scene/depth/...). Keep the allostatic gain (satiety→C attenuation, metabolism.ex:75-90) — positive-appetitive-lobe-only whitelist; it never touches the depletion penalties (the suicidal-when-sated backdoor). That is real interoceptive dynamics, not task-C.

  • Affect→precision (deferred, honestly staged): the colony has NO 8-channel Z vector — only surprise-driven gamma_m/gamma + a one-axis :stress→γ (hormones.ex) + a non-causal Emotion read-out. A later cure may extend :stress→γ into a small interoceptive-affect precision channel — but γ must stay a GLOBAL policy/ factor precision (sharpens the whole softmax uniformly), never a selective gain on any C's positive lobe (that would be reward-in-a-wig, bypassing the qo·C-only rule). Its Z-ablation falsifier is pre-registered before that cure runs. Do NOT fabricate the full Z.

I.3 What is REMOVED / neutralized — BOTH task-C channels, by viability-provenance (blocker #4)

Gated on curriculum: :survival_only:

  • Channel 1 — Curriculum task-C: the phase 1–4 inventory/vision/sky weights (curriculum.ex:37-45) → neutral (the survival table returns zeros for them).
  • The climb: phase_goal_met?/2 + maybe_advance_phase/2 + set_phase/2 C-refresh (mc.ex:219-227,479-500) → no-op.
  • Channel 2 — runtime strategist_config/1 (mc.ex:428-466), easy to miss: it injects absolute per-option C overrides. Whitelist by viability-provenance, not by name: KEEP status(needs_safe) + threat (danger_calm/danger_flee) — viability-derived. DROP inv_forage/inv_build/vis_tree/vis_shelter and light_surface/sky_surface — all are hand-authored spatial/task preference-hacks NOT derived from any viability setpoint (e.g. inv_forage rewards has_wood +2.0, vis_tree rewards seeing-a-tree +3.0). A "curriculum-free" agent that still carries build/anti-bedrock preferences is not curriculum-free.
  • diagnose.shadow_wood (diagnose.ex:86-93) becomes N/A for these lineages.

I.4 Seams — additive+gated, :metabolism BINDING, heritable-field discipline (blockers #2, #13)

  • Genome field curriculum: :survival_only | :phased (default :phased). Gates I.2 (survival table vs Curriculum.preference), the climb no-op, and I.3 channel-2 neutralization.
  • :metabolism is a BINDING prerequisite of the survival-C treatment (blocker #2 / embodiment). The ONLY non-identity emptying/filling B in the whole model is the metabolism organ (genome.ex:109-111, b_init: :emptying, pb_seed: 50.0). status/threat/self/social carry setpoint C but have identity "states-persist" B (passive read-outs = preferences, not homeostats). A survival_only genome WITHOUT :metabolism is "setpoint C with no emptying B" = the Phase-1-insufficient preference-relabel — RED-A's FALSIFIES would then be pre-ordained, not earned. So the survival-C lineage MUST carry :metabolism.
  • Heritable-field discipline (blocker #13): back-fill slow_defaults Map.put_new(:curriculum, :phased) (genome.ex:360-366 precedent); read Map.get(dna, :curriculum, :phased) in card/1 (mirroring novelty_gain, genome.ex:240); if heritable, APPEND the Det draw LAST in mutate/2 (genome.ex:311) to preserve draw order (else every existing lineage's mutation stream shifts).
  • Byte-identity: default/0 stays :phased and untouched. Because C never enters A/B/D/policies-tensors (backbone) and phase-0 default is already survival-only, express(default()) is bit-identical. Test: extend decider_byte_identity_test.exs to express BOTH :survival_only and :phased and assert the default golden stays mad<1e-12 over the depth-5 Plan path. Run action_clone_invariance A1/A2/A3 on the survival_only lineage (its informative-A survival factors are exactly what the guard needs; strategist_config injects action-INDEPENDENT per-outcome C, so A2's no-action_cost guard must be exercised here).

I.5 RED-A — paired, pre-registered (single-variable; corrections #2, #15)

  • Arms (differ in EXACTLY the curriculum: field): treatment :survival_only vs control :phased (the current climbing curriculum — that IS the cure under test). Both arms carry :metabolism + energy/satiety factors + identical novelty_gain, world/seed/body/kin, and are pinned to the SAME start phase (live default is phase:1, colony.ex:107/lineage.ex:132 — pin it explicitly). A probe asserts all-else-equal.
  • Replication unit = ≥5 distinct world-seeds (backbone RED-discipline), not N UNIs in one seed.
  • Activation gate (numeric, FIRST): energy-posterior depletes/refills (pre-registered depletion slope) AND G5b energy-severed twin dies (ticks-in-V(acting) − ticks-in-V(noop-twin) > 0, p<0.05 over the seed set). Miss ⇒ WITHHELD.
  • World-ceiling: the reference controller (one role, pinned before T0) shows the natural-behaviour target IS reachable in these worlds; else WITHHELD + re-scope.
  • PASS = non-inferiority + activation: treatment live-fraction ≥ control − 0.15 (G5a) AND treatment ≥ the pre-registered natural-behaviour floor (distinct resources touched / placed_used ≥ 1, RCON) — i.e. removing the curriculum does NOT collapse the agent. G6 plateau-break is SECONDARY + expected-FAIL for survival-C-alone (do not spin a G6 non-move as pass or as the cure failing).
  • FALSIFIES: the survival-C agent goes inert (no viable self-maintenance, no natural foraging) while the curriculum control sustains ⇒ the curriculum was load-bearing for any competent behaviour (the pivot needs the epistemic drive first). — Valid ONLY because :metabolism is bound in (else this branch is pre-ordained).

Target code (gated, additive; ONLY after formal MERGED VERDICT + owner go-ahead)

curriculum.ex (survival table), genome.ex (curriculum: field + back-fill + :metabolism prereq for the survival-C lineage constructor), mc.ex (climb no-op + strategist_config viability-whitelist, both gated), tests as in I.4. Launcher runs/survival_c_lineage.exs + probe + the world-ceiling reference.

sha256 8a5a5567c248625b — 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 design specification for removing a taught curriculum from an agent, so that what it does comes only from staying viable rather than from a goal handed to it. It is design only, and it opens by calling the idea a hypothesis to test rather than something to assume.

The interesting parts are the corrections. An earlier claim about the shape of a preference is described as wrong, and the true shape is given. There is a single peak at a comfortable level, with being over-supplied mildly less preferred than being comfortable, so there is no standing pull toward more.

The removal is done by provenance rather than by name. Preferences that come from staying viable are kept; hand-authored spatial and task preferences are dropped, on the grounds that an agent still carrying those is not curriculum-free.

One prerequisite is binding: without a real emptying-and-filling internal state, the whole test would be decided before it ran.

Plain · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is 8a5a5567c248625b

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 part one of a paired design, read alongside a backbone document and its sibling. It is design-only, with a ship gate of formal review plus the owner's go-ahead, and it carries corrections folded in from a review.

The hypothesis is stated as something to test rather than assume. With preference grounded only in viability, a real death edge, and a natural drive to find things out, an agent keeps its body viable and acts without any curriculum or goal-setting. An earlier stage is cited as having shown the curriculum was a confound, so removing it lets the question be asked cleanly. The page calls this an open question.

No new factors are added. The change is to the preferences on existing factors, and here the document corrects itself. An earlier description of the preference shape is called wrong, and the true shape is given: a single peak at a comfortable level, a steep penalty toward depletion, and a neutral value when over-supplied. Because being over-supplied is slightly less preferred than being comfortable, there is no monotone pull toward more, so the only standing gradient is away from depletion. The refuting condition is restated against that true shape: if any preference is monotone increasing in more, it is a preference hack and is struck.

The replacement preference table is phase-independent, returning viability values at all times and neutral values for every task-related channel. One existing attenuation is deliberately kept, because it acts only on the appetitive side and never touches the depletion penalty, which the document names as a backdoor it is avoiding. A further extension is deferred honestly, with a note that a precision term must stay global and must never become a selective gain on the positive side of a preference, because that would be reward in disguise.

The removal itself is done by provenance rather than by name, which is the passage worth reading. Two channels carry task preference. The first is the obvious curriculum weighting, neutralised. The second is a runtime configuration that injects absolute overrides and is described as easy to miss. Preferences traceable to viability are kept; hand-authored spatial and task preferences are dropped, with examples named, and the reason is given in one line: an agent that still carries build and avoidance preferences is not curriculum-free.

A seams section covers how the change is gated behind an inheritable field defaulting to the existing behaviour, and how that field is back-filled and read defensively so existing lineages keep their sequence. One prerequisite is binding. The treatment must carry the organ that provides a real emptying and filling internal state. Without it the arrangement would be preference with no dynamics, and the refuting branch would be decided before the run rather than earned.

The experiment is paired, with arms differing in exactly one field, everything else pinned and asserted by a probe. The replication unit is a distinct world seed. A numeric activation gate comes first, and missing it produces a withheld result. A reference controller pins reachability. Success is defined as non-inferiority plus activation rather than as an improvement, and one broader measure is named as secondary and expected to fail, with a warning against spinning it either way. The refuting condition is that the agent goes inert while the control sustains, which would mean the curriculum was load-bearing.

A final list names the code the change would touch, gated and additive, only after a formal verdict.

Clear · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is 8a5a5567c248625b