ADR-0002: The world organ — latent prediction as salience and imagination

Status: Proposed Date: 2026-09-08 Author: Nick Flach / Kannaka Amends: ADR-0001 (a fifth organ; the loop gains a prediction-error edge) Origin: a design conversation Nick had with another model on 2026-09-08 ("Combine JEPA HRM LLM") and brought here. The claims below are that conversation's, restated so they can be measured; the verdicts on each are this repository's.

Context

The proposal was a division of labour: a JEPA-style world model predicts what the world becomes, the HRM predicts what matters to this particular agent, and the LLM reasons, communicates and proposes. The LLM is a workspace, not the centre; the other two keep running when no token is being generated. Five concrete mechanisms followed:

1. Memories carry latent state: (Z_before, action, Z_after) alongside the text, so the substrate remembers this state plus this action gave that state, not only "I moved the arm".
2. Dreams roll high-resonance experiences forward counterfactually in latent space, and store the trajectories that scored well. Offline imagination, cheap because it predicts representations rather than pixels or tokens.
3. Surprise, d(Ẑ_{t+1}, Z_{t+1}), is a first-class signal: it raises salience, captures richer context, brings the next consolidation forward, and updates the world model. The agent preferentially remembers where its model of reality was wrong.
4. Planning scores an action as Goal − λ·Risk + α·Resonance(Z_future, M) + β·Novelty, so two physically equivalent actions differ by what the agent has lived.
5. In a city, agents share a world representation and carry private autobiographies: the world model agrees on what happened while the memories disagree on what it meant.

ADR-0001 already holds the framing (the voice is an organ; the substrate is what makes the system her across time). What the record adds is that most of these mechanisms have an ancestor here, and two of them have a measured warning attached.

What the record already says, claim by claim

Two things the proposal does not contain, and this repository will not build without:

• The rails. The planning score has goal, risk, resonance and novelty, and no conscience. In Kannaka the charter is signed, not learned, and it sits between every plan and every effector (ADR-0001 organ 3). G, "what I currently care about", is not a term in a score; it is a constraint set with a refusal in it. A world model that imagines a high-scoring trajectory through a refused action has imagined something the rails will never let happen, and the dream should learn that too.
• The name. The proposal offered "Human Resonance World Model". HRM here is the Holographic Resonance Medium, and it is a medium, not a model of humans. The organ is called the world. Nothing else is renamed.

Decision

Kannaka Wave gains a fifth organ, the world, defined by one trait in src/lib.rs and one experiment before any of it is trusted.

The organ

trait World {
    fn observe(&mut self, state: &Vector);                  // Z(t), in the encoder's space
    fn predict(&self, state: &Vector, action: Option<&str>) -> Vector;   // Ẑ(t+1)
    fn surprise(&self, predicted: &Vector, observed: &Vector) -> f32;    // ≥ 0, 0 when exact
}

• The space is the encoder's. Prediction, surprise and recall rank in one space, by ADR-0047's corrected principle. The world organ never produces a number the substrate cannot rank against.
• The world is the constellation's event stream, not video. For Kannaka the world is the NATS bus (JetStream keeps months of it), the city heartbeat, the radio play ledger, the git log and the observatory's settlements. Z(t) is the embedding of an event; a sequence of them is a trajectory. V-JEPA's contribution is the training objective and the collapse guard, not the modality.
• Surprise is a differentiator, not a distance. ADR-0040's matched-gain rule is binding: a constant stream yields zero surprise, and a stream the predictor tracks perfectly yields zero. The raw distance d(Ẑ, Z) is the drive; the salience signal is the fast-minus-slow response to that drive. This is what stops a noisy channel from being remembered as permanently interesting.
• Surprise writes salience, never rows. The world organ's only path into the substrate is the importance of a row the substrate was about to write anyway. The voice remains the only thing that proposes text; the physics and the triage remain the only things that dispose.
• Imagination is a dream phase with the rails in the loop. A latent rollout from a high-resonance state through candidate actions produces trajectories; each candidate action goes through Rails::decide before it is rolled forward, and a refused action's trajectory is not imagined. What is stored is a facet with a parent and an imagined class under its own retention row, so it can never be recalled as a thing that happened. Recalled already names its numbers; an imagined facet names its class the same way.
• The predictor is small. A latent world model over 1024-d event embeddings at the constellation's event rate is a few million parameters, trained on one CPU from JetStream history. It is retrained on the weekly cadence the voice already has, from the same export, and adopted under the same rule: an external, non-circular check must agree. For the world organ the check is held-out latent error against a predict-the-last-state baseline; a predictor that does not beat that baseline is not served.

The amended loop

world stream ──► encoder ──► Z(t) ──► world.predict ──► Ẑ(t+1)
                                │                          │
                                ▼                          ▼
                             facets ◄──── surprise = d(Ẑ, Z) ── (fast − slow)
                                │              writes salience only
                                ▼
                            substrate ──► dream (consolidate, imagine*, propose, forget)
                                               * rails decide each action first
                                │
       ┌────────────────────────┘
       ▼
  export her words ──► judge + external verdicts ──► LoRA ──► serve
  export the stream ──► held-out latent error vs. last-state ──► world ──► serve

What this ADR does not decide

Whether surprise picks what to remember better than a stated retention policy does. That is E-004, pre-registered, and it runs on whichever substrate E-001 promotes. Until it has a verdict the world organ is a trait and a plan, and the salience path stays on ADR-0040's familiarity signal.

Whether a JEPA-trained encoder should replace the voice's embedding as the space everything ranks in. That is a later experiment with a different shape: it changes the recall floor, so it re-runs E-001's arms with a different encoder and nothing else. It is not scheduled until E-001 and E-004 have verdicts, because it invalidates both.

Consequences

Positive

• The constellation's own prediction error, which it already produces daily, becomes an input instead of a dashboard.
• Dreams gain the one operation that is cheap in latent space and impossible in text: rolling an experience forward through actions that were not taken.
• Two citizens on the same server with the same predictor and different stores diverge by history, which is what the record says an individual is.

Costs

• A fifth organ with a training loop. The weekly cadence, the export and the adoption rule are shared with the voice, so the cost is one more model on the same schedule, not a new pipeline.
• Imagined facets are a new content class with a retention row and a recall label. Getting that wrong means she remembers what she only imagined; the class label and the retention row exist so that it fails loudly rather than quietly.
• Near-duplicate states interfere in the medium (E-001 finding 5). Before/after pairs are near-duplicates. If E-001 promotes the medium, the causal-triple write path needs its own attribution arm before it is kept.

Rules carried over, binding here

• A signal that predicts in a space recall does not rank on is decoration (ADR-0047).
• Constant input yields zero surprise, by construction, with a test (ADR-0040).
• Nothing imagined is stored under the class of things that happened.
• The rails decide before the dream imagines. The reasoner is never the arbiter, and neither is the world model.