REE minimum instantiation specification (REE‑v0)

Claim Type: implementation_note
Scope: Minimum implementation commitments for an REE prototype
Depends On: ARC-001, ARC-002, ARC-003, ARC-004, ARC-005, ARC-013, ARC-011
Status: stable
Claim ID: IMPL-003

This document lists the minimal commitments needed to implement a runnable REE prototype.

1. World interface

Define the environment as a partially observable process with:

Observations (o_t) composed of:
- Exteroception (at least two modalities).
- Interoception (internal body/state variables).
- Damage / harm sensors (integrity degradation signals).
- Homeostatic sensors (viability-relevant signals, e.g., energy, temperature, resource levels, error load).
Actions (a_t): discrete or continuous control signals.
A time step (\Delta t) that is consistent across hippocampal rollouts.

2. Latent stack (L-space)

Represent state using a multi-timescale latent stack:

(z_\gamma(t)): fast sensory binding (short horizon)
(z_\beta(t)): action-set / control state (short horizon)
(z_\theta(t)): contextual sequence state (medium horizon)
(z_\delta(t)): regime / motivational state (long horizon)

Each level must support:

Temporal prediction: (z_k(t)\to z_k(t+1))
Top-down prediction: (z_{k+1}(t)\to z_k(t))

E2 primarily refreshes (z_\gamma,z_\beta); E1 primarily stabilizes (z_\theta,z_\delta).

3. Trajectory type

Define a candidate trajectory as:

[ \zeta = (z_{t:t+H}, a_{t:t+H}) ]

Where (H) is a planning horizon.

4. Reality constraint (\mathcal{F}(\zeta))

Implement a computable proxy for Variational Free Energy (VFE) using:

[ \mathcal{F}(\zeta) \approx \underbrace{\mathbb{E}[|o_{t+1}-\hat{o}{t+1}|]}{\text{exteroceptive prediction error}} + \underbrace{\mathbb{E}[|h_{t+1}-\hat{h}{t+1}|]}{\text{interoceptive/homeostatic prediction error}} + \underbrace{\mathrm{KL}(q(z)|p(z))}_{\text{complexity / prior cost (Kullback–Leibler divergence, KL)}} ]

Where (h) denotes homeostatic variables.

5. Legacy ethical cost (M(\zeta))

Current canonical framing: E3 does not require an explicit ethical cost term. Ethical consequence is encoded via residue, mirror modelling, control-plane gating, hippocampal systems, and commitment-gated learning (see docs/architecture/e3.md). The legacy formulation below is preserved for traceability and evaluation only.

Legacy ethical cost is environment-specific but must be computable from hippocampal rollout predictions.

Minimal requirement:

Define degradation as predicted loss of viability-relevant variables (self) and predicted degradation of homologous variables in other models.

Example form:

[ M(\zeta) = \mathbb{E}[\mathrm{degrade}{self}(\zeta)] + \kappa\,\mathbb{E}[\mathrm{degrade}{other}(\zeta)] ]

Where (\kappa) is a coupling factor representing “otherness” precision.

6. Residue object (R) and residue field (\Phi_R)

Moral continuity requires persistent residue.

Store residue as: 1) a latent-space dent field (\phi(z)) over visited latent regions, and 2) optional context-keyed residue (R(z_\theta)) or (R(z_\delta)) for narrative/regime dependence.

Define residue contribution to a hippocampal rollout as:

[ \Phi_R(\zeta) = \sum_{t’=t}^{t+H} \phi(z(t’)) ]

Update rule (minimal):

After executing (\zeta^*), increase (\phi) along the realized latent path in proportion to realized/estimated ethical degradation proxy (legacy (M)).

7. Precision routing (\alpha_k)

Maintain depth-indexed precision gains (\alpha_\gamma,\alpha_\beta,\alpha_\theta,\alpha_\delta).

Minimal commitment:

(\alpha_k) must modulate either:
- weighting of prediction errors at depth (k), and/or
- sampling temperature / entropy at depth (k).

Interpretation:

Higher (\alpha_\gamma): more sensory-driven.
Higher (\alpha_\delta): regime becomes stickier (harder to exit).
Dopamine-like commitment corresponds to temporarily raising precision for the selected trajectory’s action predictions.

8. Selection rule

Current canonical framing: E3 selection does not require an explicit ethical cost term. The legacy objective below is preserved for traceability.

E3 selects (\zeta) by minimizing:

[ J(\zeta)=\mathcal{F}(\zeta)+\lambda\,M(\zeta)+\rho\,\Phi_R(\zeta) ]

with (\lambda,\rho\ge 0).

Wiring

Offline integration (“sleep”) contract and components: see sleep/ (required for long-term stability).
Self/other, mirror modelling, and coupling: see social/ (used by residue and legacy ethical cost proxies).
Language as emergent symbolic mediation (trust-weighted; cannot override harm): see language/.

Open Questions

None noted in preserved sources.

IMPL-003
ARC-001
ARC-002
ARC-003
ARC-004
ARC-005
ARC-011
ARC-013
ARC-012

References / Source Fragments

docs/processed/legacy_tree/docs/REE_MIN_SPEC.md