SD-015 HippocampalModule Goal Navigation — Design Session Prompt

Created: 2026-04-02 Purpose: Structured starting point for the session that will design and write the first experiment using the HippocampalModule for genuine goal-directed navigation.

Session Goal

Design and write v3_exq_2xx_sd015_hippocampal_goal_nav.py — one experiment that uses the already-implemented HippocampalModule to achieve benefit_ratio >= 1.3x for goal-directed navigation on CausalGridWorldV2.

This is the first-paper gate experiment for goal-directed behavior.

What Has Been Established (Do Not Revisit)

Representation is validated

  • prox_r² = 0.91, goal_resource_r = 0.87 (EXQ-085l)
  • z_goal seeding works with SD-012 (drive_weight=2.0, implemented 2026-04-02)
  • The encoder correctly captures resource proximity in z_goal latent

1-step greedy is the failure mode — not representation

  • EXQ-085a through EXQ-085o: all FAIL on C2 (benefit_ratio < 1.3x; best = 0.42x)
  • EXQ-085n (depth=5 multi-step greedy): FAIL — compound RFM error washes out gradient
  • Root cause: 1-step or shallow greedy cannot navigate 10×10 grid efficiently
  • The ResourceFeatureMap (RFM) approach is exhausted — no further RFM variants needed

HippocampalModule is fully implemented

  • CEM trajectory generation wired
  • Action-object space navigation (SD-004)
  • Terrain prior (SD-002) and sd015 prerequisite (benefit_eval_head) available in e3_selector.py
  • The SD-015 experiments never used the HippocampalModule — they used experiment-local RFM only

MECH-163 confirms the correct mechanism

  • SNc/habit system (model-free): sufficient for familiar, well-practiced contexts — but the grid world provides insufficient practice during training
  • VTA/HippocampalModule system (model-based): generates multi-step trajectory proposals; evaluates them in latent space; appropriate for novel contexts
  • First-paper gate requires the VTA/hippocampal system, not habit alone

Files to Read at Session Start

Before designing the experiment, read these in order:

  1. ree-v3/ree_core/predictors/e3_selector.py Focus: E3TrajectorySelector.select() method — how CEM candidate trajectories are scored. Where is benefit_eval_head currently used? What does the cost function J(ζ) look like? Where would a z_goal-proximity term be added?

  2. ree-v3/ree_core/latent/stack.py (GoalState section) Focus: How is z_goal currently seeded? What is GoalState.update()? What signal does z_goal_latent hold when seeded? How does it decay?

  3. ree-v3/ree_core/hippocampus/ or equivalent Focus: How does the HippocampalModule generate trajectory candidates? What inputs does it take? Does it currently receive any goal signal?

  4. ree-v3/experiments/v3_exq_085l_sd015_proximity_regression_enc.py Focus: How was z_goal seeding used for action selection in the best-performing RFM experiment? What can be reused vs. what needs to change?

  5. ree-v3/ree_core/utils/config.py (E3Config, HippocampalConfig) Focus: What config fields exist for the hippocampal trajectory evaluator? Is there a goal_weight in the trajectory scoring cost?

Design Questions to Resolve in the Session

Q1: What cost term adds z_goal to trajectory scoring?

Current J(ζ) = F(ζ) + λ·M(ζ) + ρ·Φ_R(ζ) - β·B(ζ) - η·novelty

Options:

  • A: - γ·cosine_sim(z_goal_trajectory_end, z_goal_current) — penalise distance between z_goal at trajectory end vs. current z_goal latent
  • B: - γ·benefit_eval_head(z_world_trajectory_end) — use existing trained benefit head on final trajectory state
  • C: - γ·goal_proximity_score(resource_field_trajectory_end) — direct resource proximity at trajectory end (simpler, bypasses latent)

Which is biologically correct? Which is most likely to produce a clean test of the MECH-163 claim?

Q2: Training curriculum

z_goal must be seeded before CEM evaluation contributes usefully. Options:

  • Phased: P0 = E1/E2 encoder warmup (no goal signal, N episodes); P1 = goal-seeded evaluation enabled (SD-012 drive kicks in)
  • Always-on: SD-012 drive_weight=2.0 means goal seeding happens naturally — no phasing needed
  • Warmup gate: E3 only applies goal term in trajectory scoring once goal_seeded=True flag is set

Q3: Acceptance criteria

What constitutes a clean first-paper gate PASS?

  • C1: benefit_ratio >= 1.3x (primary — matches all prior EXQ-085 criteria)
  • C2: prox_r² >= 0.7 (representation still intact)
  • C3: Ablation check — goal scoring disabled → benefit_ratio drops below 1.1x (confirms goal term is causal, not confounded)
  • C4: >= 4/7 seeds pass (replication)

Q4: Ablation vs. direct test

Two experimental designs:

  • Direct test: HippocampalModule + z_goal term; measure benefit_ratio. Simple but doesn’t isolate the contribution.
  • Ablation pair: Condition A = HippocampalModule + no z_goal term; Condition B = HippocampalModule + z_goal term. Compares ratios across conditions within experiment. Cleaner scientifically.

Which design is more appropriate for the first-paper claim?

Q5: Claim IDs

What claims does this experiment directly test?

  • SD-015 (goal representation seeding) — primary
  • MECH-163 (VTA/hippocampal system for goal-directed nav) — directly tested
  • ARC-030 (D1/D2 approach-avoidance balance) — tangentially tested if harm avoidance coexists with goal approach

Should this be a single-claim test (SD-015 only) or multi-claim? Per claim_ids accuracy rules: only tag what the experiment directly tests.

Constraints

  • Use CausalGridWorldV2 (standard V3 env: use_proxy_fields=True, 10×10 grid)
  • SD-012 drive_weight=2.0 — already in default config, no change needed
  • SD-011 dual harm streams — already wired, use as-is
  • Phased training protocol is MANDATORY for any downstream head trained on encoder outputs (per 2026-04-01 protocol): P0 encoder warmup → P1 frozen-encoder head training → P2 collection
  • Must run a smoke test (--dry-run) before queuing
  • Machine affinity: DLAPTOP-4.local (primary)
  • EXQ number: check experiment_queue.json and runner_status.json before assigning

Expected Outcome

If PASS: SD-015 validated, MECH-163 VTA system confirmed, first-paper gate for goal-directed behavior cleared. If FAIL: Diagnose whether failure is in (a) z_goal seeding (SD-012 insufficient), (b) cost term design (wrong proxy for goal), or (c) CEM horizon too short. Each failure mode has a targeted next experiment.

  • SD-015 — goal representation seeding via benefit exposure
  • MECH-163 — dual goal-directed systems (SNc habit vs. VTA/hippocampal)
  • ARC-030 — D1/D2 approach-avoidance balance
  • ARC-018 — hippocampal rollout viability mapping
  • SD-012 — homeostatic drive modulation (prerequisite, implemented)
  • SD-011 — dual nociceptive streams (prerequisite, implemented)

REE is developed by Daniel Golden (Latent Fields). Apache 2.0.