ctde_v0.run_slam_attn

run_slam_attn

SLAM × neighbor-attention head-to-head @32²/10 — the 'ready' transformer experiment.

Two new knobs, on the shootout-winning base (role split + learned-Lagrangian connectivity): AGG (2): max (current pooling) vs multihead (softmax attention over neighbours = GAT). PERCEPTION (2): --no-sense-walls (old wall-blind belief) vs --sense-walls (SLAM occupancy).

The 2×2 isolates the attention lift, the wall-perception lift, and their interaction. max_blind ≈ reproduces the connectivity-shootout's role+lag (sanity anchor at ~42cov/85conn). FRESH (no warm-start) — the screen; the winner gets the laddered + multi-seed follow-up.

PYTHONPATH=.:../../../FiedlerValueEstimation ~/ZymeraLab/.venv/bin/python         ctde_v0/run_slam_attn.py --out runs/slamattn --seeds 3 --iters 2000 --jobs 4

Unit

Unit(uid, run_dir, seed, extra)

Single-scale 32²/10, fresh (no warm-start) — mirrors run_conn_shootout.Unit.

Source code in experiments/ctde_v0/run_slam_attn.py

def __init__(self, uid, run_dir, seed, extra):
    self.uid = uid; self.run_dir = run_dir; self.rung = (32, 10, 5)
    self.seed = seed; self.extra = extra
    self.init_from_dir = None; self.needs_dir = None; self.proc = None