E2 — Fast Forward Predictor (Affordance and Immediate Inference Engine)

Claim Type: architectural_commitment
Scope: Fast predictive stream; near-future sensory expectations and action affordances
Depends On: INV-013 (cognition is predictive, iterative, multi-timescale), L-space, E1, precision control
Status: stable
Claim ID: ARC-002

E2 is the fast predictive stream of the Reflective Ethical Engine (REE).
It performs rapid, low-latency forward inference from the current latent context to near-future sensory expectations and bodily/action affordances.

E2 is not the deep world-model (that is E1).
E2 is the fast forward predictor that makes the near future “reachable” as a structured set of options.

Subsystem abstract (core claims): ARC‑002 defines the fast forward predictor. Supporting context includes ARC‑001 (deep priors), ARC‑004 (latent stack), ARC‑005 (precision routing), INV‑013 (multi‑timescale prediction), and ARC‑018/MECH‑033 for the E2‑kernel → hippocampal rollout handoff.

1. Architectural role

E2 sits between incoming sensory evidence and the trajectory selection engine (E3):

Consumes current sensory/latent state (and E1 priors)
Produces near-term predicted sensory states and action-conditioned outcomes
Exposes a compact action/affordance surface for E3 to choose trajectories through

E2 is the engine of rapid perception-for-action: it keeps the system responsive while E1 maintains slower, deeper structure.

Interpretation note:

E2/E1 decomposition is a typed responsibility boundary, not an assertion of strict functional isolation.
E2 remains part of one coherent REE cognifold with bounded coupling to E1, hippocampal systems, and control-plane dynamics.

2. Inputs and outputs (interface contract)

Inputs

E2 receives:

Current sensory evidence (o_t) (or encoded features)
Current shared sensory latent (z_\gamma(t))
(see latent_stack.md)
Optional E1 priors / context
(slow predictive state and expectations)
Current precision profile components relevant to fast inference
( \alpha_\gamma, \alpha_\beta )
(see precision_control.md)
Optional motor state / proprioception (for closed-loop affordances)

Outputs

E2 emits:

Near-future predicted sensory states
( \hat o_{t:t+\epsilon} )
Action-conditioned outcome predictions
( \hat o_{t:t+\epsilon}\,|\,a )
Affordance representation / action surface
( z_\beta(t) ) (or equivalent)
Fast prediction error signals for updating (z_\gamma) and informing precision control

3. What E2 computes

E2 computes a temporally displaced bundle of predictions at short horizons:

“If I do nothing, what do I expect to sense next?”
“If I do action (a), what changes in the immediate future?”
“Which actions are locally feasible and low-cost given the current body/world state?”

This yields an affordance manifold: a structured map of immediate possible actions and their predicted consequences.

Canonical clarification (2026-02-09): E2 provides short-horizon predictions and reafference comparison. Explicit multi-step rollouts are generated by hippocampal systems, seeded by E2 predictions and E1 constraints.

Terminology: E2 performs forward prediction (local transition kernels), not explicit multi-step rollouts.

4. Relationship to E1 and the latent stack

E2 and E1 cooperate rather than compete:

E1 provides deep priors and longer-horizon structure (narrative, goals, affective regime predictors)
E2 provides short-horizon predictive closure (rapid sensorimotor anticipation)

In REE terms:

E2 primarily populates the fast horizons of the latent stack (sensory (z_\gamma) and action/affordance (z_\beta))
E1 primarily stabilises deeper horizons (workspace (z_\theta), regime/value (z_\delta), etc.)

E2 should remain perceptually corrigible: it must be able to update against sensory error rather than enforcing top-down overwrite.

SD-002: Mutual constitution (confirmed 2026-02-27)

E1 and E2 are not merely parallel systems with orthogonal outputs. They are co-constitutive:

E2 scaffolds E1. E2’s transition sequence is the temporal evidence stream from which E1 distills associations. The association “arm-up → throwing” only becomes reliably learnable because E2 provides the ordered kinematic chain that instantiates it. Without transition structure, E1 would require many more raw observations to converge.
E1 primes E2. E1’s associative prior conditions E2’s transition predictions. Knowing “this is a throwing pattern” (E1 association) lets E2 anticipate upcoming transitions more accurately. The full bidirectional loop is:
```
E2 transitions → temporal scaffolding → E1 association distillation
E1 association → predictive prior → E2 transition smoothing / context
```
V2 implementation (2026-03-06): The E1 prior is now wired into HippocampalModule.terrain_prior, which expands its input from (z_beta, residue_scalar) to (z_beta, e1_prior, residue_scalar). The E1 prior conditions the terrain search’s initial action distribution on long-horizon associative context. See ree_core/hippocampal/module.py and ree_core/agent.py generate_trajectories() for implementation. SD-002 wiring is complete.
E2→E1 autotrain pathway (V3 aspiration). E2’s forward predictions can generate synthetic temporal sequences that E1 learns from, without waiting for ground-truth observations. Not yet implemented.

Implication for interpreting MECH-058 FAILs (EXQ-002, EXQ-006): The E1_FROZEN and E2_FROZEN conditions in those experiments disrupted both directions of the mutual-constitution loop, not just the targeted component. Additionally, SD-001 contamination (E2 doing hippocampal work) made E2_FROZEN equivalent to “no trajectory proposal” rather than “no transition prediction.” These results should be interpreted as system-level baselines only. Retesting on V2 substrate after SD-001 resolution (Step 2.1) may yield cleaner mechanistic results.

5. E2 as self-attribution substrate (SD-003)

Because E2 is a pure action-conditioned transition model f(z_t, a_t) → z_{t+1}, it is the natural substrate for counterfactual self-attribution:

z_actual_next = e2.predict_next_state(z_t, a_actual)
z_cf_next     = e2.forward_counterfactual(z_t, a_counterfactual)
causal_delta  = z_actual_next - z_cf_next   # agent's causal signature

Transitions that are similar under any action are environment-caused. Differences between z_actual_next and z_cf_next are attributable to the agent’s action choice.

This is only possible when E2 is a pure, independently-callable transition model. The SD-001 conflation (E2 doing hippocampal trajectory search) prevented this in V1.

V2 implementation: E2FastPredictor.forward_counterfactual(z, a_cf) added. Experimental validation (V2): EXQ-027/028 — calibration_gap = 0.027 (FAIL, threshold 0.05). Root cause: z_gamma conflates proprioceptive self-effects with world-directed contamination signal; E2’s predict_harm head received only indirect supervision (harm prediction is E3’s domain, not E2’s — MECH-069).

V3 redesign (MECH-095): SD-003’s V3 form is a joint E2+E3 pipeline:

E2 provides dynamics on z_self: z_self_{t+1} = E2(z_self_t, a_t) (efference copy)
TPJ comparator (MECH-095) checks z_self_predicted vs z_self_observed → agency_signal
E3 evaluates harm on z_world: harm_actual = E3(z_world_{t+1}(a_actual)), harm_cf = E3(z_world_{t+1}(a_cf))
World-delta causal signature = harm_actual − harm_cf (E3’s domain, clean signal)

E2 no longer carries a predict_harm head in V3. See tpj_agency_comparator.md.

6. Precision coupling and cognitive regimes

Precision control shapes E2’s behaviour:

High sensory precision ((\alpha_\gamma\uparrow))
→ rapid correction to incoming evidence, perceptual anchoring
High action-depth precision ((\alpha_\beta\uparrow))
→ sharper affordance selection, faster commitment when E3 engages
Noradrenaline-like interrupt signals
→ reset fast expectations, break maladaptive local attractors
Serotonin-like anti-collapse signals
→ keep local options open, avoid premature narrowing of affordances

E2 is therefore a key surface where global regulatory state becomes immediate cognitive style.

(see precision_control.md)

7. Interaction with E3

E2 does not decide. It proposes.

E3 consumes E2’s affordance manifold and action-conditioned predictions, then selects a coherent trajectory that can be committed and lived:

E2: generates near-future options and local predicted consequences
E3: selects and commits a trajectory across horizons, constructing the unitary present when licensed

(see E3.md)

8. Failure modes (fast-stream pathologies)

Common failure regimes for E2-like systems:

Over-narrow affordances
premature pruning → compulsive loops, reduced flexibility
Over-permissive affordances
insufficient pruning → distractibility, unstable control
Excessive sensory gain
overload → fragmentation, derealisation-like states
Weak sensory correction
top-down drift → false local coherence, hallucination-like errors

These are typically driven by precision misallocation rather than representational “content errors”.

9. Summary

E2 is the fast forward predictor that converts current context into the immediate future.

It:

predicts near-term sensory flow,
maps action-conditioned outcomes,
generates an affordance manifold,
and provides the fast substrate that E3 can collapse into lived experience.

Without E2, REE is slow and brittle. With E2, REE is responsive.

Open Questions

None noted in preserved sources.

ARC-002
ARC-001
ARC-004
ARC-005
INV-013
ARC-018
MECH-033

References / Source Fragments

docs/processed/legacy_tree/docs/architecture/e2.md
docs/processed/legacy_tree/architecture/E2.md
docs/thoughts/2026-02-09_e2_hpc_interface.md