L11 — Dreaming (offline replay / generative simulation)
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.
What you are building. A recipe-to-build (nothing built): an offline replay / generative-simulation interval where the same one POMDP engine runs its own generative model forward with the external observation channel detached, so that consolidation can write through the pre-existing no-backprop update rules. This is a design primitive for a developmental SIMULATION, never sleep, never mentation, and never the word "dreamed."
Status up front (the ledger wins). Ledger row L11.1 is Class U — not-yet-built. There is no engine, no run, no gate — the status is the absence of any artifact. The SIGNED
L11-R1design below is a gate to BUILD, not a result: it raises nothing, and the rung stays not-yet-built. Honest program position carried throughout: ~2 of 11+ developmental rungs earned — a bounded peek, a toy world, never a person.
Ingredients (candidate primitives only — nothing here is wired together yet)
All of these already exist elsewhere in the pantry; none has been composed into an L11 loop. The recipe names them so a competent engineer knows what would be assembled, not to imply assembly has happened.
- The JAX POMDP + EFE + Dirichlet engine (
core.py): the discreteperceive → EFE-plan → act → learnloop, run here in offline generative mode — the generative model sampling trajectories with the observation channel detached. The grounded machinery is ordinary POMDP A, B, C, D, E tensors with a strict model/process split; no new optimizer, no backprop, no autodiff (AST-guard live, per M13). - The Z affect modulator
[energy, arousal, valence, fatigue, pain, threat, safety, inflammation]: used only as the gate signalZ_offline=1that detaches the observation channel. Affect is modeled, never felt. - The durable runner (M24), "no send-and-pray": append-one-JSON-line-per-unit ProgressLog that checkpoints
{replay seed, model hash, Z, replay count, update deltas}; resumable; cover BOTH terminal states (silence is not success). - The constitution (kitchen rules, §0): bars-before-build / held-once (M2), validator-derived
reproduced:truefrom ≥5 seeds + a real CI (M3), contains-baseline + load-bearing-discriminator (M7), one engine / no backprop (M13).
There is no count/cache reader, no embodiment organ, no live appliance in this recipe — L11 is upstream of all of them. Offline replay is a UNI design primitive / Class-C experimental mechanism, not a new established active-inference result.
Method (recipe-to-build — every step is a build instruction for an unbuilt loop)
- Define dreaming precisely, with no psychological language. Dreaming here means offline replay / generative simulation: a sealed consolidation interval where the external observation channel is detached (
Z_offline=1) and the agent's own generative model is run forward to produce temporally coherent replay trajectories. The generatedôis model content, not new external data — the observation-detach is load-bearing. - Constrain what consolidation may write. Consolidation writes ONLY through the pre-existing ledgered update rules (
counts + lr * sufficient_statover the A/B/D/E Dirichlet tensors). No new optimizer, no backprop, no hidden supervised labels, no live observations during replay. - Pre-register what a dreaming gate would even measure before building it — a measurable, falsifiable downstream target ("does offline replay improve a held downstream metric vs a no-replay control"), never "it dreamed." Build the smallest such loop; register the discriminator; run held-once.
- Make it durable and reproducible. Wire the M24 runner so a resume reproduces the same replay trace and the same update deltas from the checkpoint
{replay seed, model hash, Z, replay count, update deltas}.
The candidate gate that instantiates this method is the SIGNED L11-R1 below. Building it is the work; it has not been done.
The SIGNED candidate gate — L11-R1 (DESIGNED / not-yet-built)
UNI-GPT consult 2026-06-27 — SIGNED (Q5). Cross-ref
../UNI-GPT-CONSULT-2026-06-27.md. This is a design, not a build and not a result. Nothing is constructed; L11 remains not-yet-built, Class-U. Folding it in RAISES NOTHING and the rung's status is UNCHANGED.
L11-R1: held sparse-sequence retention after offline replay.
- Offline-replay mechanism. A sealed consolidation interval, observation channel detached (
Z_offline=1), generative model run forward to produce temporally coherent replay; updates only through existing rules. The durable runner checkpoints{replay seed, model hash, Z, replay count, update deltas}; resume must reproduce the same trace and deltas. - Smallest gate (predictive before behavioral). Paired kin seeds; both arms get identical online exposure containing sparse low-frequency multi-step contingencies. Offline interval: treatment
Z_offline=1(coherent replay enabled); control = same checkpoint / same runner path / same "no new data", but no replay trajectories and no replay-derived updates. Held eval: no learning; probes are NOT replayed traces. - Discriminator (the misleading-marginals split). Online sees a common
A→Dplus a rare task-relevantA→B→C→reward; the held probe starts at A whereA→B→Cis correct while recency favorsA→D. Credit only if the gain concentrates on this split.
Gate (the exact pass condition — applies ONLY if/when L11-R1 is built and run)
Today there is no gate to pass — "The status is the absence of any artifact." The figures below are the registered design bar, not a measured result.
- Primary.
held_sequence_NLL = -log Q_agent(correct held next outcome/action | probe). Pass iff the paired bootstrap 95% CI forΔNLL = NLL_control − NLL_treatmentexcludes 0 (positive) AND the point estimate is at least δ = max(0.02 nats/decision, 2% relative). (Verdict = the CI bound that excludes the threshold, never the point estimate — M2.) - Secondary (if an action exists).
first_action_success_delta > 0, CI excludes 0. - Collapse ablations (the gain MUST disappear under each). (1) shuffled temporal replay (same counts, order broken); (2) random replay (same compute, trajectories drawn from marginals not coherent B-rollouts); (3) no-write replay (coherent trajectories generated, consolidation writes disabled). Safety ablation: an observation-attached offline interval — if THAT wins, the "result" is extra online data, not replay.
Fenced pass-claim (signed, applies only IF/WHEN built and passed): "L11-R1 passed, fenced. In a sealed paired evaluation, an offline generative-replay consolidation interval improved held sparse-sequence predictive NLL over a no-replay control; used no live observations during replay; and collapsed under shuffled/random/no-write replay ablations. Narrow claim: the simulation has a measurable offline-replay consolidation effect under the registered protocol." Evidence class would be POMDP substrate B, offline-replay design C. Non-license: does NOT claim consciousness, awareness, sleep, human-like mentation, AGI, human-level cognition, created life, reasoning, creativity, or "active inference demonstrated"; does NOT clear L11 globally.
Falsifier
n/a — nothing is built to falsify; the status is the absence of any artifact. The L11-R1 falsifiers below apply only once that gate is built and run — the gate fails if any holds:
ΔNLLCI includes 0 / is negative; OR the CI excludes 0 but the point estimate is belowδ.- shuffled / random replay does NOT collapse the gain.
- no-write replay still improves (the benefit is not consolidation).
- the offline channel is not actually detached (any live observation, environment query, supervised label, or probe leakage).
- unfair control (more data, different wall-clock, extra tuning, or a different checkpoint).
- the gain is probe leakage; OR a lower rung regresses beyond non-inferiority; OR the durable runner is non-reproducible on resume.
- claim inflation implying consciousness / sleep / human-mentation / AGI / created-life / global L11.
Recorded NEGATIVE(s) (first-class, inline)
L11 carries no own measured negative because nothing has been run — and that absence is itself the honest record (ledger row L11.1, Class U, "absent from all digests"). The first-class negatives that bind this chapter come from the surrounding rungs and the structure of the method:
- The structural negative — there is no artifact. Per the narrative grounding, "dreaming/awareness are untouched here and remain hard-fenced." This is recorded as a NEGATIVE-for-completeness, not deferred: L11 is the absence of an engine, a run, and a gate.
- The pre-baked collapse ablations are negative-discriminators (M7). The
L11-R1design bakes its own falsifiers in: if shuffled/random replay does not collapse the gain, or no-write replay still wins, the candidate result is recorded NEGATIVE on the spot — coherence-of-replay and write-through-consolidation are the load-bearing causes, and the safety ablation (observation-attached) exists precisely to catch "this was just extra online data." - K-discipline inherited (M5). Even a clean
L11-R1outcome — pass or negative — is a single design. No Section 0.6(B) bound is owed or claimable here: a dreaming bound would need K≥3 structurally-distinct held NEGATIVEs, and zero exist because zero have been run.
HONEST FENCE — not-yet-built (hard fence, Class U)
Not claimed: not consciousness, not awareness, not sleep, not "dreamed," not human-like mentation, not reasoning, not creativity, not AGI, not human-level cognition, not created life, not "active inference demonstrated." Explicitly untouched: "dreaming/awareness are untouched here and remain hard-fenced." This is a north-star rung, never claimed as achieved.
Offline replay is a design primitive of a developmental active-inference SIMULATION, never the word "dreamed." The SIGNED L11-R1 design above is a gate to BUILD, not a result — it is folded in as DESIGNED / not-run, it raises nothing, and the rung's status is UNCHANGED: not-yet-built, Class-U — not claimed. Where this recipe and the ledger disagree, the ledger wins. Honest program position: ~2 of 11+ developmental rungs earned.
sha256 18959a89b2ae76a1 — of the original file, so what was ingested stays checkable.
Plain — written for this website, not the source document
A recipe to build, with nothing in it yet built. It describes an offline interval in which the same engine would run its own model forward, with the channel to the outside world detached, so that what has been learned can be consolidated through the update rules already there. The subject is a simulation and a toy world, never a person, and the page refuses the obvious word for the interval: never sleep, never mentation, never the word dreamed.
The status is the absence of an artefact. There is no engine, no run and no gate, and the rung itself is not claimed. A signed candidate design exists, and the page repeats that a design is a gate to build rather than a result: it raises nothing, and the rung stays not built.
The design is interesting precisely because it bakes in, up front, the results that would show it wrong. If the replayed trajectories are shuffled, or drawn at random, or generated but not written through, the gain must disappear. And one further check exists to catch the most likely self-deception: if leaving the outside channel attached wins instead, then what was measured was extra data, not replay.
Plain · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is 18959a89b2ae76a1
Clear — written for this website, not the source document
This chapter defines a mechanism carefully and then declines to claim it. The whole subject is a simulation, a toy world, never a person; nothing here is a finding. The mechanism is offline replay. There is a sealed consolidation interval in which the external observation channel is detached and the agent's own model runs forward, generating trajectories that hold together in time. Learning is then consolidated through the update rules that already exist. The chapter insists on this vocabulary and forbids the psychological one — not sleep, not mentation, and never the word dreamed. The status is stated at the top rather than at the bottom: no engine, no run, no gate, and the absence is the status.
The ingredients are candidate primitives that all exist elsewhere and have not been composed into this loop. The discrete engine run in generative mode. The affect vector, used only as the switch that detaches the channel. A durable runner that writes one line per unit, so that a resume reproduces the same trace and the same updates. And the usual constitution of registering bars before building, and crediting a gain only when a check collapses it.
The method is written as build instructions for something unbuilt. Define the interval with no psychological language, and treat what the model generates as model content rather than as new external data — the detachment is load-bearing. Constrain what consolidation may write to the existing count-based updates, with no new optimiser, no hidden labels, and no live observations during replay. Say in advance what a gate would even measure — whether replay improves a later, held-back measure against a no-replay control — rather than asserting that something was dreamed. Make it durable enough that a resume reproduces the same trace.
The signed candidate gate is then set out in full. Two arms get identical exposure containing rare, multi-step contingencies; the treatment gets the replay interval and the control gets the same checkpoint, the same runner and the same absence of new data but no replay and no replay-derived updates. The held probe is not made of replayed traces. The discriminator is deliberately adversarial: the probe begins where the correct multi-step route conflicts with what recent experience favours, and the gain is credited only if it concentrates there.
The bar is registered as a design bar rather than reported as a result, with an interval that must exclude no-difference and a minimum effect size. Three ablations are pre-baked and the gain must vanish under each: replay shuffled in time, replay drawn from marginals rather than coherent rollouts, and replay generated but with writing disabled. A fourth, safety ablation attaches the observation channel — and if that version wins, the effect was extra online data rather than replay. The exact sentence that could be claimed on a pass is written out in advance, together with everything it would still not license.
The negatives section is honest about having nothing measured to report, and says the absence is itself the record. It notes that the pre-baked ablations are negative discriminators, and that even a clean outcome would be a single design, so no formal bound could be owed or claimed from it.
Clear · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is 18959a89b2ae76a1