# S1: vanilla PE vs GAB (matched steps)

Date: 2026-08-03  
Worktree: `.worktrees/chessformer-lite`  
Command: `python -m experiments.chessformer_lite.run_s1_compare --train-path /mnt/c/Users/jun/chessdb/data/train/data.parquet --out runs/chessformer_lite/s1 --steps 20000 --device cuda`  
Artifact: `runs/chessformer_lite/s1/compare.json`

## Setup

| Knob | Value |
|---|---|
| steps | 20,000 |
| batch | 256 |
| warmup | 2,000 |
| seed | 0 (shared trunk init matched vanilla→GAB) |
| d_model / layers / heads / d_ff | 256 / 6 / 8 / 384 |
| probe | 4,096 held-out train rows |

## Results

| Arm | params | wall (s) | ms/step | final loss | policy CE (final) | probe top-1 | probe policy CE | train FLOPs |
|---|---:|---:|---:|---:|---:|---:|---:|---:|
| vanilla | 2,992,071 | 1074 | 53.7 | 3.134 | 2.455 | 0.464 | 1.758 | 5.97e15 |
| **GAB** | 7,566,471 | 1171 | 58.6 | **2.776** | **2.127** | **0.488** | **1.626** | 1.51e16 |

ms/step ratio GAB/vanilla = **1.09** (cap 1.30) → speed OK.

## Decision

**GO — scale GAB.**

- Better final loss and better probe policy CE.
- Quality win under matched steps/data/seed with shared trunk init.
- Step-time overhead well under 1.3× despite ~2.5× params (GAB templates dominate param count, not step cost).

## Next

S2: `run_s2_scale --use-gab true` with flop_cap 4.665e17, then `eval_play` vs v1-final.