Founder ontology – how to read E1, E2, and E3

Status: plan-of-record for intent (how the architecture was meant to be read). Authority for behaviour: claims registry and linked experiments. Where this doc and a claim conflict, the claim wins until governance updates it.

Audience: contributors, reviewers, and agents working on REE. Read this before proposing “standard” ML stacks (world model + policy + critic, monolithic transformer planner, etc.).

1. Why this document exists

REE’s three-engine split (E1 / E2 / E3) was defined from a biology-first picture of cortex and cerebellum, plus a separate commitment and comparison loop – not from default deep-RL or LLM-agent templates.

Early sessions often re-expressed that picture in familiar AI terms (extra planners, trajectory generators in the wrong module, “E2 as long-horizon world model”). Much of the claims index and repeated review cycles exist to record the founder reading and demote drift.

This page is the short, stable map between:

Founder concepts (what you meant),
REE names (modules and claim subjects),
Biology analogs (functional, not homology).

2. The three functions in founder terms

E1 – perception-association matrix (cerebral base)


Founder concept	The base perception-association matrix: slow integration of “what is going on” in the world and self. Not optimized for fast action; holds context when movement and attention pause.
REE names	`E1`, `e1.*`, `predictors/e1_deep.py`, L-space / `z_world`, ARC-001 (persistent predictive substrate).
Biology analog	Cerebrum – especially deep and associational cortex (slow predictive tissue), not “a chat model in a box.”
Not	A separate ethics module, a fast policy, or the module that commits action.

Goals, values, and “what matters” live in the cerebral world in biology: accidents of structure that specialize repeatedly. In REE they are distributed (PFC, vmPFC, residue terrain, hippocampal valence maps, ghost-goal cue system) – not a single goals/ package. That distribution is intentional.

E2 – fast forward predictor (cortical edge + cerebellum)


Founder concept	The fast forward predictor: given current state and intended action, what happens next at short horizon? Efference-copy style bridging so perception and action stay coupled.
REE names	`E2`, `e2.*`, `predictors/e2_fast.py`, ARC-002 (fast forward predictor).
Biology analog	Outer cortical fast inference plus cerebellar forward model (state + action -> next state, ~one step). Literature in the registry often cites cerebellum explicitly (MECH-057, MECH-231).
Not	The deep world model (that is E1), the multi-step rollout engine (hippocampus / E3 complex), or a long-horizon planner. Experiments such as EXQ-132 / EXQ-212 reinforce short E2 horizon vs E1.

Historical implementation drift: E2 was sometimes wired to do hippocampal trajectory-proposal work under the E2 label. Claims and refactors treat that as misplacement, not as the founder design (see MECH-057 note on E2 vs hippocampal module).

E3 – commitment and comparison (basal ganglia system)


Founder concept	Hard to name at first; settled as commitment and comparison: perception and action as a giant experiment – generate candidates, compare predicted vs actual / harm vs goal, commit or withhold.
REE names	`E3`, `e3.`, `predictors/e3_selector.py`, comparator / harm streams, `policy/`, `heartbeat/beta_gate.py`, basal_ganglia. claims, three-loop gating (MECH-057 family).
Biology analog	Basal ganglia and related gating (plus thalamic routing): which candidate gets access to execution and cognitive resources, not “another cerebellum.”
Not	A third big predictive world model parallel to E1.

E3 is the arbiter of the experiment, not the slow matrix and not the one-step forward model.

3. One diagram (founder reading)

  CEREBRAL WORLD (slow + goals + context)
  +------------------------------------------+
  |  E1  perception-association matrix       |
  |      + goal / value specializations      |
  |        (PFC, vmPFC, terrain, cues, ...)  |
  +------------------+-----------------------+
                     | priors / state
                     v
  FAST FORWARD       |         CANDIDATE EXPERIMENTS
  +----------------+ |  +---------------------------+
  | E2  next-state | |  | Hippocampus: rollouts,     |
  |     (cerebellar| |  | paths, terrain (partner)  |
  |      + fast    | |  +-------------+-------------+
  |      cortical) | |                |
  +--------+-------+ |                v
           |         |         +------+------+
           +---------+-------->| E3 / BG     |
                               | compare &   |
                               | COMMIT      |
                               +-------------+

Hippocampus is a partner for explicit multi-step candidates and path memory; it is not “E2 renamed.” Control plane modulates when updates and modes run; it is not a fourth lobe.

4. Drift catalog (what to push back on)

Drift pattern	Typical AI default	Founder / claims correction
“E2 plans the future”	Planner head, long rollout in E2	E2 = short forward; rollouts in hippocampal / E3 complex
“E3 is another predictor”	Critic or value network only	E3 = commitment + comparison + gating (BG analog)
“Goals are a module”	`GoalEnv`, reward wrapper	Goals cerebral, distributed across claims
“One world model does all”	Single transformer	E1/E2/E3 irreducible (ARC-035 family); typed error targets
Brain map shows E1-E3 as side boxes	Engineering strip only	v1.1 intent: E1 on cerebrum, E2 on cerebellum, E3/BG on gate (see brain_map.md)

When an agent proposes architecture, ask: does this preserve the experiment loop (compare then commit) and the E1/E2 split (slow matrix vs fast forward)?

5. Mapping table (quick reference)

Founder	Primary REE artifacts	Claim subjects (examples)	Biology (functional)
Perception-association matrix	`e1_deep`, L-space, `z_world`	`e1.`, `latent_stack.`, `architecture.*`	Cerebrum / deep cortex
Fast forward	`e2_fast`, action-object kernels	`e2.`, `e1_e2.`	Cerebellum + fast cortical edge
Commitment & comparison	`e3_selector`, policy, beta gate, comparator	`e3.`, `basal_ganglia.`, `commitment.`, `harm_stream.`	Basal ganglia, striatal gating
Rollouts & path memory	`hippocampal/`	`hippocampus.`, `hippocampal.`	Hippocampal formation
Goals (distributed)	PFC, residue, ghost goals	`pfc.`, `vmPFC.`, `goal.`, `hippocampus.`	Frontal + medial temporal specializations

6. For agents and reviewers

Default frame: cerebrum (E1 + goal tissue) -> fast forward (E2) -> candidate experiments (hippocampus) -> commit (E3/BG).
Before adding a module: say which founder function it serves; do not duplicate E2 as planner or E3 as world model without a claim.
Brain map / diagrams: use founder ontology for layout intent; use /api/brain-map for live claim/evidence overlays.
Promoting or demoting claims: cite experiments; founder ontology does not override evidence_direction on manifests.

E1, E2, E3 – interface contracts
Brain map – visualization and region YAML
Hippocampal systems – rollouts (not E2)
Rule apprehension / BG – commitment layer
Architecture overview – public three-engine introduction

Recorded 2026-05-19 from founder review and brain-map discussion. Update via governance when intent changes; do not let session drift silently rewrite this file.