UNI Universal Natural Intelligence

Wiki · The Cookbook

The Continuity / Embodiment Substrate

The Cookbook · cookbook/continuity-substrate.md @ 575fc93d9d31 (main) — opens the published snapshot e850f872196d

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 Cookbook is the method carried out step by step: 34 pages of recipes for building a developmental active-inference SIMULATION — a bounded peek at a toy world, never a person. The front matter says that word is never softened under any pressure, so it is not softened here. The recipes run from the molecular and cellular rungs up through metabolism, motor control, perception, language and metacognition, and on to rungs that are still open questions. Around them sit a set of kitchen rules, a shared pantry of engines and primitives, and a second family of recipes about nature itself — rocks, water, air, stars, DNA, ants, whales, bats, humans.

It is for the reader asking what building this would actually take. Each recipe names its ingredients, the order of work, the tests to run at that stage, and the point at which a step stops being something already carried out and becomes something proposed.

Begin with the front matter and then the kitchen rules. Those two pages fix the honest position and the fence labels that every later recipe leans on, and without them the status markers on a recipe are easy to skim past. After that the recipes can be read in any order.

The nature recipes sit slightly apart and should be read that way. They cite outside science — geology, chemistry, biology, astrophysics — and a nature citation is never a UNI gate: those chapters contain zero UNI claims and raise no rung.

What it is not: a claim that the whole ladder has been cooked. The book recommends the complete recipe and, on the same page, labels every rung by its real state — that tension is deliberate and is the thing the book is built around. Where a recipe and the claim ledger disagree, the ledger wins.

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.

The body the mind runs on.

What you are building. The physical substrate that lets a developmental active-inference simulation persist on real metal: deterministic serialization, bit-for-bit restart continuity, fail-closed transport, a categorical body-to-mind sensorium, and a gated control surface, all shipped on a two-node fleet. This is ENGINEERING / substrate evidence, never general-AIF evidence and never awareness. It is orthogonal to the L0-L12 developmental ladder: it does not earn a single developmental rung, and nothing here implies a science gate is met.


0. The kitchen rules (the constitution you carry into every recipe)

This is a literal cookbook: each recipe is a buildable thing with named ingredients, numbered method steps, an exact pass gate, an operable falsifier, first-class recorded negatives, and one honest fence. But before any recipe, the constitution. These are definitional / governance patterns (status: method), not capability claims, and they bind everything below.

  • The Evidence Constitution (M1). Every claim carries an evidence class (A = machine-exact anchor, B = mechanism + operator observation, C = dev-gate / held-out eval, E = test-covered, U = claimed-but-unproven) and a falsifier. The ledger is a four-state append-only record (PASS / FAIL / NEGATIVE / PENDING) and is the single source of truth. Prose must match the ledger; where this chapter and the ledger disagree, the ledger wins and this chapter is wrong.
  • Calibration only moves DOWN. Authority flows downward from the measured fact. Wording moves down to the measured value, never up, including under urgency. The fence gets louder under pressure, not wider.
  • The verdict is the CI bound that excludes the threshold (M2), never the point estimate.
  • Negatives are content. A negative is a measurement that the design is incomplete, not a failure to hide. The corpus carries 183 published negatives; they are the credibility, and in the substrate sub-ladder they are inline and first-class, never deferred.
  • Substrate work is not a science gate (M12, and the standing fence on this sub-ladder). WORLD ⊥ BODY ⊥ MIND: two typed Markov blankets, interoception as hardware signals, a discrete POMDP perceive → EFE-plan → act → learn. The composed appliance is variationally-controlled (audited), module-exact only at the single-step categorical body-to-mind interface, never globally exact. Engineering continuity is never a developmental capability.

Reading the FENCE label. Every recipe closes with exactly one of four values, drawn from the ledger and never inflated:

  • proven — a held, sealed, UNI-signed PASS exists (Class A/C), with its falsifier still live.
  • designed — a typed spec or signed-in-principle design exists; the build is partial or unverified.
  • hypothesized — a stated mechanism or law with no sealed gate yet (Class U, not claimed).
  • not-yet-built — no engine, no run, no gate; north-star, hard-fenced.

Any SIGNED consult design folded in below (the C10 Dirichlet-first port) enters as designed / not-run: it RAISES NOTHING and the rung's status is UNCHANGED.

Honest program position, printed and never softened: ~2 of 11+ developmental rungs earned. The whole program is a developmental active-inference SIMULATION: a bounded peek, a toy world, never a person.


The pantry for this chapter (UNI.OS — the embodiment substrate)

One engine, many scales. The substrate recipe draws from a single archive engine:

Pantry item What it supplies
UNI.OS (the lab-in-a-box on metal) A local-first appliance OS that turns a physical server into an embodiment substrate: a body-to-mind 7-modality categorical sensorium, deterministic replay, fail-closed transport, an ASK-mode Dirichlet policy-prior, and a token-gated control-MCP. Engineering / substrate evidence, NOT general-AIF evidence.

The fleet it runs on is load-bearing for the honesty: two physical nodes, one platform. node1/prod is a Dell PowerEdge (no-AVX Xeon X5650, a PERC HDD array of spinning disks ~5.5 TB, not SSD); node2 is an OptiPlex (NVMe). A WireGuard mesh ties them; each box mints a per-device identity uni-lab-<mac> with its own TLS leaf on firstboot. The hardware spec is part of the claim: if inspection shows an AVX CPU or SSDs, the substrate proof is contradicted.


The recipe — continuity on real metal

What you are building. A ladder of separate, falsifiable substrate claims, from "the containers survived a restart" up to "the mind-state survived a process restart bit-for-bit," with the next rung (mind-tick continuity across a kernel swap) recorded as OWED.

Ingredients (the substrate requirements, each a named pantry primitive)

  • Container survival on metal (the two-node fleet, WireGuard mesh, per-device TLS identity) — C0.
  • Bit-for-bit process-restart continuity (durable + sha256-verified + fail-closed + bit-identical tick; the continuity transport: serialize_agent → bytes, blob_sha256 anchor, deserialize_agent, idempotent serialize(deserialize(blob)) == blob) — C1.
  • The body-to-mind sensorium (the box reads its own telemetry into a 7-modality categorical contract) — C3.
  • ASK mode (ITIL-as-active-inference: a persistent Dirichlet policy-prior over operator verdicts) — C4.
  • Cross-box single-approval gate (token-gated control-MCP, one HMAC per mutation) — C5.
  • The on-core inference anchor (a sealed f64 belief-update trace) — C6.
  • Continuity-isolation primitives (Markov-blanket isolation assertion, least-privilege vault, drift sentinel, byte-exact checkpoint, sha256 output integrity, zero-hidden-LLM reply path, Postgres RLS) — C13.

Method (build steps a competent engineer could follow)

  1. Boot the fleet on metal. Bring up two nodes from one image; mesh them over WireGuard; on firstboot mint a unique per-device identity and TLS leaf so two boxes never both answer to the same name. Verify the status page is served from each node. (C0.)
  2. Make mind-state durable and verifiable. Serialize the real JAX agent state to bytes (sorted-key JSON over base64 raw float32/int32), anchor it with blob_sha256, and prove idempotent round-trip plus a bit-identical tick across a real child-process boundary with at least two distinct actions. Wire fail-closed so a 1-bit corruption refuses rather than proceeds. (C1.)
  3. Wire the sensorium. Have the box read its own telemetry (os_sysinfo / systemctl / podman ps / journalctl) and band it into the 7-modality categorical contract {load, mem, swap, disk, services, containers, journal} encoded [M=7, O_max=4] — not a float vector, not a softmax. Record RAW rows; band in exactly one place (the mind); run validate_raw_row before banding. Confirm a 28-cell string_vector is flowing live on both boxes. (C3.)
  4. Install ASK mode. Let the mind escalate a reasoned change request; the operator approves-and-executes (with auto-rollback) or declines-with-category; update a persistent Dirichlet policy-prior keyed by (host-state, action). Keep the safe-action set FIXED (observe / scale / restart); never self-execute, never widen the set. (C4.)
  5. Gate every cross-box mutation. Require exactly ONE human approval on the entry box, carried as a one-time tool-and-args-bound HMAC token, before any cross-box "limb" mutation executes. (C5.)
  6. Seal the on-core anchor. Run a sealed f64 belief-update trace and confirm it is byte-identical across the distinct stacks. (C6.)
  7. Ship the isolation primitives. Land the Markov-blanket isolation assertion, least-privilege vault, drift sentinel (cosine + Youden-J), byte-exact checkpoint, sha256 output integrity, zero-hidden-LLM reply path, and live Postgres RLS. (C13.)

Gate (the exact pass conditions, with the calibrated ledger figures)

Carry the calibrated figures, never the inflated headline ones. A peer-reviewed honesty audit (os-cycles 39-53) moved six headline claims DOWN: "cleared on TWO boxes" → ONE box; "the mind survives a patch" → infrastructure continuity only; "5 stacks" → "4 distinct stacks / 5 runs."

  • C0 (Class A, proven): both nodes boot and serve their status page; the hardware spec (no-AVX Xeon, HDD-not-SSD) holds on inspection.
  • C1 (Class C, B-substrate, proven): mind-state survives a process restart BIT-FOR-BIT — durable, sha256-verified, fail-closed, bit-identical tick across a real child-process boundary (≥2 distinct actions; Path-B receiver 3/3). REQ-002 Stage-1 GREEN. (Honest caveat: uniVerdictSign was still null on the joint gate — mechanism green across a real process boundary, not a sealed cross-host result.)
  • C3 (Class A live): the 28-cell string_vector flows on both boxes; the contract is the [M=7, O_max=4] categorical alphabet, not float/softmax.
  • C4 (Class A, learning shift measured): the Dirichlet prior updates across operator verdicts and the proposal shifts reproduce (decline ×3 → stops proposing; not_a_problem ×2 → noop; approve+fixed → more confident); the mind never self-executes and never widens the safe set.
  • C5 (Class A, proven): a cross-box mutation lands behind exactly one HMAC approval (proven live both boxes 2026-06-26: prod→node2 write gated once, landed; node2 queue stayed count=0).
  • C6 (Class A, proven): the sealed f64 belief-update trace is byte-identical across 4 distinct software/emulated stacks (5 runs) — the calibrated figure, never "5 stacks."
  • C13 (Class B, proven substrate): every listed isolation primitive is present and functional under grounded inspection; Postgres RLS 5/5.

Falsifier (operable)

A node fails to boot or serve its status page, or the hardware spec is contradicted on inspection (C0). A process-restart replay diverges bit-for-bit, OR fail-closed does not trigger on 1-bit corruption, OR serialize(deserialize(blob)) != blob (C1). The string_vector stops flowing on either box, OR the contract is found to be float/softmax rather than the [M=7, O_max=4] alphabet (C3). The Dirichlet prior does not update across verdicts, OR the mind self-executes or widens the safe set, OR the measured proposal-shifts do not reproduce (C4). A routed cross-box mutation executes without the single approval (C5). A fifth distinct stack produces a non-identical trace (C6). Any listed isolation primitive is absent or non-functional under inspection (C13).

Recorded NEGATIVE(s) — first-class, inline

  • C2 (OWED / NEGATIVE) — the central rung not shown. "The mind survives a kernel swap" is NOT shown. The live OS update is proven infrastructure-only: a real prod kexec cutover 6.12.86 → 6.12.73, ~59 s, zero data loss (odoo tables identical, sentinel rows preserved); CRIU/livepatch proven on box. But mind-tick continuity across the swap was NOT shown. node2's evidence-collected swap was infra-only (~90 s freeze, api+tts CRIU-preserved, 2 publishers fresh-restarted); the first attempt FAILED (unclean kexec → dirty ext4/ESP → emergency mode) and then recovered. Class A. Discharged only when a kernel swap is shown preserving mind-tick continuity bit-for-bit end-to-end, not just infrastructure.
  • C7 (NEGATIVE, bound) — the honest floor on live OS update. A single-box swap is a SECONDS-LONG FREEZE, not zero-downtime; true zero-freeze needs the second node carrying the platform; external media legs (RTP/SIP/kernel-mode rtpengine) are NOT preserved across kexec. Class A.
  • C8 (NEGATIVE) — keep all seven modalities. An over-compressed 2-modality sensory bottleneck went NEGATIVE on held data. Class C. (A factored mean-field posterior variant was separately implemented and rejected as lossy: the exact joint posterior is used everywhere.)
  • C9 (NEGATIVE, trade) — the EDAIT honest trade. An exact-discrete active-inference transformer trades fluency for calibration: held-out perplexity ~33 vs a backprop GPT's ~25 (less fluent, but natively online-learning + calibrated). An honest trade, not a win. Class C.
  • C14 (NEGATIVE, gap) — multi-tenancy. The bare substrate is MISSING tenant/client/namespace isolation (flat global perms) and temporal decay on learned counts (contamination would be permanent). The product build is the tenant wrapper + decay, not the core. Class B.

The central open gap — C10 (PARKED), and its SIGNED first port

The coupling rule reads: UNI.OS embodies only primitives proven in the science repo, frozen at the passed-gate SHA, one-directional (science PROVES → UNI.OS EMBODIES); everything beyond the passed gates stays Class-U-not-claimed. That rule is signed in principle but NOT literally true — the program's central open gap. Today UNI.OS has no no-backprop Dirichlet learning, no exact info-gain EFE, heuristic (denylist) isolation rather than the structural whitelist assert, a DIFFERENT dev model (Gray-Scott reaction-diffusion vs the forager/ontogeny that earned the bars), and an empty lab evidence store (/var/lib/uni/evidence empty, 0 worlds registered). Class U. PARKED.

SIGNED design (UNI-GPT consult 2026-06-27, Q2): port no-backprop DIRICHLET LEARNING FIRST, via a frozen-SHA equivalence gate. Of the four missing primitives, Dirichlet learning ranks first — central, canonical, no-backprop, low-risk, directly testable. The instruction is to vendor the exact primitive — do not re-author, reinterpret, or "improve" it during the port; record the SHA; prove byte/behavior equivalence to the frozen science implementation. The math interface is reproduced exactly: A-learning by conjugate count a_ij <- a_ij + sum_tau o_tau,i * s_bar_tau,j; the expected-log E_Q[ln A_ij] = psi(a_ij) - psi(sum_k a_kj) used in place of raw ln A; B/D/E updated only by the same registered conjugate-count family, never by optimizer steps. The equivalence gate (in CI + ledger) requires the same concentration updates and expected-log tensors within declared tolerance, with no gradient descent / backprop / learned weights outside the Dirichlet update, plus a no-backprop guard and a frozen-provenance record. Even after this port lands, the honest status is only "one passed-gate no-backprop learning primitive is literally embodied," nothing broader: NOT "UNI.OS implements full EFE-based planning," NOT "UNI.OS structurally enforces the Markov blanket," NOT "active inference demonstrated."

Status: SIGNED DESIGN / not-yet-ported. C10 stays the program's central open gap, PARKED. Nothing is discharged until the equivalence gate passes with recorded evidence; this design RAISES NOTHING.

Related parked substrate threads: host-native (off-podman) on-chip brain runtime (C11, directive; outcome not yet in archive) and the node2 local auto-kiosk "SOLVED 2026-06-26" which is PROVISIONAL pending hands-on monitor re-verification (C12).

HONEST FENCE — proven (substrate engineering) + first-class NEGATIVES + one PARKED central gap

The substrate rungs C0, C1, C3, C4, C5, C6, and C13 are proven as engineering / substrate evidence (Class A / Class C-B-substrate / Class B), each with a live falsifier. The negatives C7, C8, C9, and C14 are proven negatives, published inline. C2 is an OWED negative and C10 is parked with a signed-but-not-yet-ported first port.

Not claimed (the load-bearing fence on this sub-ladder): the sensorium is NEVER awareness; "the mind survives a kernel swap" is NOT shown (Stage-2 owed, C2); none of this is general-AIF capability; UNI.OS's embodiment does not imply the science gates are met. Never "active inference demonstrated" (the live UNI.OS loop is a separate reimplementation, not gate-matched). Continuity is a measurable engineering rung, not a slogan and not a developmental capability.


Closing — what the body does and does not mean

This chapter builds a body, not a mind. The substrate makes a developmental simulation persistable: it can survive a restart bit-for-bit, read its own telemetry into a categorical sense, learn a calibrated policy-prior from operator verdicts, and act only behind a single human approval. Every one of those is real, falsifiable, and carded at its measured class. But the ladder stops honestly one rung early: the mind has not been shown to survive a kernel swap, and the coupling that would make "embody only what's proven" literally true is parked behind a signed first port that has not yet run.

Carry the calibrated figures (ONE box, infra-only continuity, 4 distinct stacks / 5 runs), never the inflated headlines. The substrate is a body the simulation runs on — a bounded peek on real metal, a toy world, never a person. Honest program position: ~2 of 11+ developmental rungs earned. Falsify any step.

sha256 f91558f074207064 — 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 chapter is about the body a simulation runs on, not about a mind. It describes the physical substrate — real servers on real hardware — that lets a developmental simulation persist. The machine saves its state so a restart resumes it bit for bit. It reads its own telemetry as a kind of sense, learns from an operator's decisions, and acts only after a human approves. The page insists throughout that this is engineering evidence and never evidence about the science, and that it earns no developmental rung.

The one thing it says is where the ladder stops. The mind has not been shown to survive a kernel swap. That step is recorded as owed, and it is written into the middle of the page rather than buried at the end.

It is equally plain about what is missing elsewhere. The rule that the body should embody only what has been earned in the science is signed in principle but not literally true, and that gap is parked behind a design that has not yet been ported. A sensorium is never awareness.

Plain · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is f91558f074207064

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 chapter builds a body, not a mind, and it says so in its own closing words. Its subject is the substrate that lets a developmental simulation keep running on real metal. Its first line and its last are the same: this is engineering evidence, never general evidence about the science, and it earns no developmental rung.

It opens with the same constitution the rest of the book carries. Every claim gets an evidence class and a falsifier naming what would sink it. Claims live in an append-only ledger that nobody edits and that wins any disagreement. Wording is calibrated only downward, negatives are treated as content, and the four status words apply here too. A signed design folded in here enters as designed and not run, and raises nothing.

The recipe itself is a ladder of separate, falsifiable engineering claims, and the chapter is unusual in giving buildable steps. Bring up a small fleet from one image, mesh the machines privately, and give each one its own identity at first boot. Make the mind-state durable. Write it out as bytes and anchor it with a digest, then show that reading it back gives exactly the same bytes and the same next step across a real process boundary, with corruption refusing rather than proceeding. Wire the sensorium so the box reads its own telemetry and files it into a small fixed set of categories rather than into a vector of numbers. Add a mode where the machine can escalate a reasoned request and the operator approves or declines with a reason. A running count of those verdicts shifts what gets proposed next, with the set of safe actions fixed and never self-executed. Gate every change that crosses a machine behind exactly one human approval. Seal a numerical trace and check it comes out identical across different stacks. Ship the isolation pieces.

Each of those has a pass condition and, immediately after, a matching falsifier that names what would sink it. A machine that will not boot. A replay that diverges. A sense channel that stops flowing, or turns out to be the wrong kind of contract. A prior that does not update. A cross-machine change that lands without its approval. A stack that produces a different trace. An isolation piece that is absent.

The negatives are inline and first-class. The central one is that continuity of the mind across a kernel swap has not been shown; what was shown is infrastructure continuity, and a first attempt failed before recovering. A live update on a single box is a seconds-long freeze rather than no downtime, and some external legs do not survive it. An over-compressed sensory bottleneck lost on data it had not seen, so the fuller set of channels is kept. A fluency-for-calibration trade is described as honest rather than a win. The bare substrate lacks tenant isolation and any decay on learned counts.

The largest open item gets its own section. The coupling rule — that the body should embody only primitives earned in the science, frozen at the version that passed — is signed in principle but not literally true today, and the chapter lists exactly what is absent. A signed design proposes porting one learning primitive first, vendored exactly rather than reinterpreted, behind an equivalence gate. It has not been ported, it discharges nothing, and the gap stays parked.

Clear · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is f91558f074207064