lattrack

Working name. When a latent reasoning model's decoded leaning changes mid-computation, is it revisiting an earlier latent state, overriding it, or is the change only in the decoder? BRIEF.md has the question, WORKSTYLE.md how we work, lit-sweep.md the literature, LOG.md the running record.

The pilot runs on the two-layer COCONUT-style model from RRR/phoenix (graph reachability, trained from scratch). Infrastructure files under src/lattrack/ are copies from the phoenix repo; each carries its origin in a header line.

Setup (Mac, CPU)

git clone --depth 1 https://github.com/Ber666/reasoning-by-superposition.git vendor/reasoning-by-superposition
git -C vendor/reasoning-by-superposition fetch --depth 1 origin 72d22af7ec1533aef6bd208e3f08a819d3e1b47f
git -C vendor/reasoning-by-superposition checkout -q FETCH_HEAD
uv venv --python 3.12 .venv
uv pip install --python .venv/bin/python torch==2.5.1 numpy==2.1.3 transformers==4.46.2 datasets==3.1.0 tqdm==4.67.0 pyyaml

Checkpoints go in ckpts/<run>/best.pt (gitignored). seed0 and seed1 are the paper's accepted s1_seed0_win and s1_seed1_win; their sha256 prefixes are in LOG.md.

Tests (run before anything else, after any change to fast_coconut.py)

.venv/bin/python tests/test_fast_equivalence.py   # EQUIVALENCE: PASS
.venv/bin/python tests/test_prompts.py            # PROMPTS: PASS
.venv/bin/python tests/test_attn_hooks.py         # HOOKS: PASS

Runs

.venv/bin/python src/lattrack/lens.py --run-name seed1 --graphs pilot   # smoke, 10 graphs
.venv/bin/python src/lattrack/lens.py --run-name seed0
.venv/bin/python src/lattrack/lens.py --run-name seed1
.venv/bin/python src/lattrack/lens.py --run-name random
.venv/bin/python src/lattrack/lens.py --summarize seed0 seed1 random

Results land in results/<run>/. Every number in LOG.md points at a file there.