Previously _encode_global() always computed:
total_power = creatures * 3 (heuristic)
Now reads p.get('total_power', 0) first, falling back to the heuristic
only when the key is absent (legacy JSONL snapshots without Step 4 data).
Also wires total_toughness, artifacts, and enchantments into _empty_player()
so padding players have consistent keys.
- ml/eval/forge_evaluator.py: reusable eval harness wrapping
MultiThreadBatchRunner + PolicyInferenceService; collects win rate,
avg turns, life differential, commander damage, illegal action rate,
policy entropy per game; writes results/eval-{run_id}.json
- scripts/eval_policy.py: CLI to run N Forge eval games with any
checkpoint vs forge-builtin or policy-server opponent; supports
--threads, --clock, --seed, --greedy, --temperature
- scripts/compare_eval.py: loads two result JSONs, computes absolute
win-rate delta + 95% CI, runs Fisher's exact test for significance,
prints formatted comparison table and pass/fail verdict
Wires forge_episode_generator into the PPO training loop:
- New --data-source flag: synthetic (default) | forge | mixed
- New --forge-results-dir, --forge-batch-id, --forge-weight flags
- ForgeDataSource class: runs run_forge_pipeline() in background thread,
drains episodes into RolloutBuffer via episodes_to_buffer()
- Mixed mode: blends Forge episodes + synthetic self-play per iteration
using --forge-weight ratio (0.0–1.0)
- Synthetic path unchanged — no regression for existing training runs
Implements the three approved Phase 5 architecture decisions:
1. Zone embedding boundary: add explicit documentation in StateDimensions
and StateEncoder._encode_zones making the 6144-dim zero-padded zone
contract visible to future maintainers.
2. SQLite WAL online learning store: new OnlineLearningStore class backed
by WAL-mode SQLite for concurrent read/write on single-machine baseline.
New /api/policy/decide endpoint collects decisions during inference.
Stats endpoint at /api/policy/decide/stats.
3. Forge-only default weights: change ppo_weight default from 0.5 to 0.0
across scope.py, DistillationConfig, DatasetConfig, and CLI defaults.
Add MIXED_MODE_PRESETS dict with forge_only, forge_90_10, forge_80_20
presets. Distill/start endpoint now accepts a "preset" parameter.
27 new tests covering all three decisions (all passing).
Co-authored-by: Claude <claude@anthropic.com>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
The distillation pipeline was looking for JSONL decision files in
data/results/ but they are stored in results/ (matching CFG.results_dir).
This caused all generations to fail at Step 1 with:
'No ml-decisions-*.jsonl files found in data/results'
Fixed:
- routes/ml.py: results_dir and history_dir now use 'results/' not 'data/results/'
- ml/training/distillation_loop.py: updated DistillationConfig defaults to match
- History endpoint paths updated to match
- Create ml/eval/elo_tracker.py:
- EloTracker class with standard ELO formula (K=32, start=1200)
- EloResult dataclass for tournament output
- EloHistory class for JSON persistence (data/elo_history.json)
- run_elo_tournament(): round-robin ELO-rated matches using
existing run_match() from tournament.py
- run_generation_tournament(): discovers gen-*/best_policy.pt
checkpoints and runs cross-generation ELO tournament
- CLI entry point: python -m ml.eval.elo_tracker
- Add ELO API endpoints to routes/ml.py:
- POST /api/ml/elo/tournament — start background ELO tournament
- GET /api/ml/elo/tournament/status — poll progress
- GET /api/ml/elo/history — load ELO history from disk
- Integrate ELO convergence in distillation_loop.py:
- New config fields: elo_convergence_enabled (default False),
elo_convergence_threshold (20.0), elo_tournament_episodes (30)
- New field: GenerationRecord.elo_rating
- _run_elo_mini_tournament() runs after quality gate pass
- _check_convergence() checks ELO delta when enabled
- Add ELO UI to distillation page:
- ELO Ratings panel with canvas multi-line chart showing per-model
ELO progression across generations
- Run Tournament button with progress indicator
- ELO leaderboard table with rank, rating, and visual bars
- Responsive chart resize handling
Closes#70
Implements the closed-loop distillation pipeline that automates:
1. Supervised train on current dataset (Forge + PPO data)
2. PPO self-play using supervised model as starting policy
3. Export winning PPO decisions to JSONL
4. Quality gate check (win rate threshold)
5. If passed, merge into dataset and repeat
6. If failed, discard batch, adjust hyperparameters, retry
New files:
- ml/training/distillation_loop.py — DistillationLoop orchestrator,
DistillationConfig, GenerationRecord, CLI entry point
Updated files:
- ml/config/scope.py — DistillationDefaults dataclass with all
default hyperparameters for the pipeline
Features:
- Generation tagging: data and models tagged with gen-NNN
- Convergence detection: stops when win rate plateaus over N gens
- Automatic retry with LR reduction + entropy boost on gate failure
- Per-generation history saved to JSON for observability
- Full CLI with all tunable params
- progress_callback for UI integration (Phase 5 ready)
Closes#68
Phase 2 of closed-loop distillation pipeline.
- Scan both ml-decisions-sim-*.jsonl (Forge) and ml-decisions-ppo-*.jsonl (PPO)
- Add source_weights config for configurable sampling ratios
- Add min_reward_threshold filter for PPO data quality
- Tag each sample with source for downstream analysis
- New CLI args: --forge-weight, --ppo-weight, --min-reward
- NPZ output now includes 'sources' array
Phase 1 of closed-loop distillation pipeline.
Adds DecisionExporter class that captures PPO self-play decisions
as JSONL files compatible with the Forge DecisionSnapshot schema.
Includes PPO-specific metadata fields (source, model_version,
reward, episode_return) and configurable win-only filtering.
Replace all print() calls with structured logging using the Python
logging module. This provides:
- Named loggers per subsystem: commander_ai_lab.api, .batch, .sim,
.coach, .deckgen, .collection, .scanner, .ml, .cache, .pplx
- Dual output: console (StreamHandler) + rotating file (RotatingFileHandler)
at logs/commander-ai-lab.log (25 MB x 5 backups = 125 MB max)
- Configurable via env vars: CAL_LOG_DIR, CAL_LOG_MAX_MB, CAL_LOG_BACKUP_COUNT
- --verbose / -v flag for DEBUG-level output
- overnight-run.py: separate rotating log at logs/overnight-run.log (10 MB x 5)
- Sub-module loggers in ml/, scanner/, deck_sources/, src/ all propagate
to the root commander_ai_lab logger
- Stripped redundant [TAG] prefixes from messages (logger name carries context)
- Format: '2026-03-13 14:00:00 INFO [commander_ai_lab.batch] message'
Files changed: lab_api.py, overnight-run.py, and 11 sub-module .py files
176 total print() calls converted (101 in lab_api.py, 76 in sub-modules)
KoalaTrapLord