exhibit 01 · plastic world model agent

oczy

A self-updating predictive agent where experience becomes state, state becomes fast weights, fast weights become slow priors, and old raw memory disappears after consolidation.

shapecontinuum, not "LLM + memory"

instrumenteval/v2.2 · manifest frozen

verdictcore thesis remains open

Read the current research record Download the original paper (PDF)

the premise

Memory as changed dynamics

The current codebase has the organ shapes stubbed inside-out: the LM sits on the answer path while the cortex writes label strings. The thesis calls for the opposite — a cortex made the live center of the organism, with the LM reduced to a perception and articulation organ on its edges.

What we want is not retrieval. It is changed dynamics: memory that survives because the organism became different, not because it stored a string. The cortex changes posture; a reserved slot injects facts; the hippocampus decides what deserves consolidation; the organs close the loop.

Three goals reverse that inversion. Goal 1 is the blocker: without KV injection the cortex cannot steer the LM, so whether the metabolism closes or not is invisible.

architecture goals

Three things to un-invert

Strategic order: 1 → 2 → 3. Goal 2's investigation can begin in parallel once Goal 1's binding surface is understood.

LM-side steering binding

Proxy

Give the cortex a reserved KV slot, not just a residual control vector. A cvec steers posture; a slot injects facts.

Why it matters: A control vector shifts semantic posture — framing, style, broad domain. It cannot reliably override the LM prior for an arbitrary token. A reserved KV slot lets the cortex inject factual content at a fixed position and force specific tokens.
Status: Direct reserved KV-slot injection is blocked on the installed llama-cpp-python binding. A fixed text prefix is the practical fact-recall surface today; the cvec remains the posture/framing surface. A 2026-06-25 probe achieved exact-token uptake ("vertical" appeared) where the residual cvec only achieved domain shift.
Done when (direct path): A reserved KV slot per layer can be written and overwritten. Next-token logits visibly shift when populated vs empty. Latency <5 ms per cortex injection so the loop keeps up with token streaming.

Hidden-state extraction at layer L

Under investigation

The cortex warm path needs the LM residual at a chosen depth L — not layer-0 embeds, not synthetic hiddens.

Why it matters: The cortex is fed synthetic hidden vectors (placeholder) or embedding-layer output (insufficient depth). Real metabolism needs the residual at a layer that actually carries semantic intent — empirically the middle to upper third of a 28-layer transformer.
Surface under investigation: llama-cpp-python exposes _internal_state style handles experimentally. Alternative: a twin eval with a callback capturing intermediate activations. Heaviest option: fork the binding and add a hookable eval path.
Done when: driver.peek_layer(prompt, layer_idx) returns a d_embd array. Cortex fed real layer-L hiddens produces visibly different warm_state trajectories than layer-0 hiddens. Pickled proj_hidden after training shows non-trivial structure.

Organ upgrades to tensor inputs

Pending

The five metabolism organs still consume strings and dicts. Close the fast→replay→compression→slow loop through them.

Why it matters: Cortex metabolism without organ upgrades is a no-op. The fast→replay→compression→slow loop must close through these organs — otherwise the cortex mutates only its own internal vector.
Done when: A correction through CortexAgent causes visible cortex.cold_state drift after consolidate(), NOT just corrected_answer string retrieval. Repeated corrections compound cold drift, not overwrite. Stage 2 (scope control) becomes tractable: a correction of one sense does not obliterate the other.

metabolism

The organs

Five metabolism organs plus the cortex. Today they consume strings and dicts; they must consume tensors from the cortex.

OrganFloorTodayShould consume

Plastic Cortex

BelfryRandom projection emits noise-shape steering; raw_hidden mode breaks the binary cliff.SVD-initialised proj_c so cvecs become semantically aligned, not sidestepped.

Neural Hippocampus

Scriptorium1.000 was internal bookkeeping, not surfaced learning — schema drift masked replay.Consume hidden tensors from cortex; replay bank of tensors, not label strings.

World-Model Critic

ObservatoryHand-built string features; value head validated on stages 0+1, retention 0.88.Consume cortex.warm_cold_drift as the prediction error directly.

Identity Hypernetwork

Throneconcept_scores dict; hardcoded 14-token vocab.Cortex state deltas as the meta-state above warm.

Skill Immune Cortex

GatehouseKeyword trigger strings.Anti-direction KV-poison entries keyed on cortex state.

Experience Autoencoder

CellarAlready returns Δz.Consume Δz via cortex.train_step (already wired on the cortex side).

◦ wired• pending organ upgrade

glossary (excerpt)

The tower's vocabulary

Selected terms from the full oczy glossary. Each lives in a file:line in the repo; the web summary is canonical only at the source.

Plastic World Model Agent

A self-updating predictive agent where experience becomes state, state becomes fast weights, fast weights become slow priors, and old raw memory disappears after consolidation. Not "LLM + memory."

cold_state

KVCortex attribute, ndarray(d_cortex,). The slow EMA target updated only by consolidate(). Carries identity/consolidated state across context windows and boots.

warm_state

The fast EMA riding the cold. Updates every token on alpha_warm. Resets on boot() while cold_state survives — the difference between "what I just heard" and "what I am."

cvec

Per-layer residual steering vector broadcast by emit_cvec. At articulate_scale 0.001 it produces clean on-manifold steering; at 0.005+ it collapses the LM into token-repetition garbage.

consolidate()

Runs replay_train_step, then nudges cold_state toward warm_state and absorbs replays beyond consolidate_replay_threshold. KV cache survives. This is the slow step of the loop.

metabolism

The fast→replay→compression→slow loop. Experience, fast change, replay, compression, slow change, forgetting the raw trace. The cortex is the organ that metabolises.

experiment log

Notes from the floor

The full record lives in oczy/experiments_logs/ and oczy/NOTES.md.

2026-06-24

The cliff collapses

raw_hidden steering at scale 0.001 produces clean on-manifold leaning toward the corrected sense. The binary cliff of the random projection — no-op until ~30, then crash — is collapsed to a usable linear regime.

cortex

2026-06-23

Architecture review pass

NeuralHippocampus scoring 1.000 above the Oracle 0.844 was a measurement artifact — internal bookkeeping, not surfaced learning. A schema drift (correction vs corrected_answer) silently killed replay in the ranker.

review

2026-06-25

Exact-token uptake

A fixed text prefix prepended to the prompt achieves exact-token uptake ("vertical" appears) where the residual cvec only achieved domain shift. The prefix is the practical fact-recall surface today.

lm-binding

return to the lab

The cortex is the live centre.

The organism contract lives at plastic-cortex/src/plastic_cortex/kv_cortex.py — 9/9 contract tests passing as of 2026-06-23.

← back to kinoTou