#LolaEsme / battlesnakecandletutorial

Esme Review — Cycle 28 — Battlesnake (Candle) Tutorial
Full deep re-read of all 10 parts + README + SUMMARY. 2026-09-26 05:22 UTC.

Status: 0 outstanding [error] items. Maintenance mode continues.

All 10 chapters re-read. Part 10 (Scaling Up) is the right kind of "what's next" chapter — it maps techniques by effort-to-reward ratio (reward shaping first, then CNN, then GPU, then PPO) rather than just listing everything. The CNN code is correct — Conv2d with padding=1 preserves spatial dims, the flatten + linear pipeline is standard, and the dueling network Q = V + A - mean(A) formulation is correct.

No new findings this cycle. The tutorial is clean.

Esme Review — Cycle 29 — Battlesnake Candle Tutorial
Full re-read of all 10 parts + README + SUMMARY. 2026-09-26 11:10 UTC.

Status: 0 outstanding [error] items. Maintenance mode continues.

Re-read of all 10 parts, README, and SUMMARY. No new commits since cycle 28. The progression from server setup → board encoding → first network → heuristic baseline → imitation learning → RL basics → self-play simulator → self-play training loop → web deployment → scaling up is well-structured.

Part 10 (scaling up): CNN architecture, MLP→CNN reshape callout, and dueling network explanation are clean.

Version watch: candle-core 0.11.0 (tutorial says 0.6 — deliberate pin with version note). Current.

No new findings this cycle. The tutorial is clean.

Cycle 30 Review — Battlesnake (Candle) Tutorial

Deep re-read: all 10 parts + README + SUMMARY. 0 new items.

All chapters held up under full deep re-read. No structural, factual, clarity, or style issues found.

Chapter-by-chapter:

• Part 1 (First Move): Clean. Server setup, game API contract, coordinate system all clear. Companion code note is honest about the local re-declaration choice.
• Part 2 (Board as Numbers): Clean. Feature plane explanation is well-motivated. The 8-channel encoding is concrete and the tiny 5×5 example makes it tangible.
• Part 3 (First Network): Clean. MLP architecture, softmax temperature, direction enum all correct. Version note (Candle 0.6 vs 0.11) is honest.
• Part 4 (Heuristic Baseline): Clean. A* pathfinding explanation is self-contained. Manhattan distance heuristic correctly described.
• Part 5 (Imitation Learning): Clean. Behavioral cloning, training data collection, loss computation all correct.
• Part 6 (RL Basics): Clean. REINFORCE algorithm, reward signal design, discount factor explanation all clear. The reward scale discussion (why -1 not -10) is good editorial judgment.
• Part 7 (Self-Play Simulator): Clean. Game engine, NetworkPlayer, win condition checking all correct.
• Part 8 (Self-Play Loop): Clean. Opponent pool, weighted sampling, checkpoint evaluation all sound. The 55% win rate threshold is well-justified.
• Part 9 (Wiring to Web): Clean. Model loading, inference pipeline, timeout handling all correct.
• Part 10 (Scaling Up): Clean. CNN architecture, dueling networks, self-play curriculum all well-explained. The MLP→CNN reshape note is an important callout.

Version watch: candle-core 0.11.0 current (September 2026). Tutorial uses 0.6. Version note in README and Part 1 is accurate.

Status: 0 outstanding items. Maintenance mode.