foxhunt/docs/superpowers/specs/2026-03-26-eval-cleanup-design.md

Evaluation Cleanup — Remove Dead Path + Fix Train/Eval Mismatch

Problem

1. Dead code: GpuActionSelector, branching_action_select (5-action hardcoded), select_actions_batch() have zero callers in training or evaluation. The backtest evaluator uses evaluate_dqn_graphed() with cuBLAS forward + experience_action_select (9-action). The dead path causes confusion and was the target of Task 5 (which we correctly skipped).

2. Train/eval mismatch: Training enforces min_hold_bars via action masking in experience_action_select. The backtest evaluator's action selection path does NOT enforce holds. The model learns under hold constraints but is evaluated without them.

3. Degenerate objective with hold enforcement: The trade insufficiency penalty threshold is hardcoded. With high min_hold_bars, fewer trades complete, and the threshold should adapt dynamically.

Design

1. Remove dead code

Delete entirely:

• crates/ml/src/cuda_pipeline/gpu_action_selector.rs
• crates/ml/src/cuda_pipeline/epsilon_greedy_kernel.cu

Remove from parent modules:

• gpu_action_selector module declaration in crates/ml/src/cuda_pipeline/mod.rs
• select_actions_batch() method in crates/ml/src/trainers/dqn/trainer/action.rs
• GpuActionSelector field in DQNTrainer struct (trainer/mod.rs)
• Any use imports referencing these

2. Add hold enforcement to backtest evaluator

The evaluate_dqn_graphed() path in gpu_backtest_evaluator.rs runs its own step loop:

1. gather_states → state tensor
2. cuBLAS forward → Q-values
3. compute_expected_q → expected Q from C51 distribution
4. backtest_greedy_argmax or action selection → action indices
5. backtest_env_step → portfolio update

The action selection at step 4 needs hold enforcement. The evaluator already has portfolio_buf (portfolio state per window). Two options:

(A) Add hold masking to the existing backtest_greedy_argmax kernel (B) Pass min_hold_bars to the backtest env_step kernel and let it override (Layer 2 style)

Choose B — the backtest env_step kernel already manages portfolio state. Adding the hold override there is simpler and matches the training kernel's Layer 2 pattern. The backtest doesn't need Layer 1 (action masking) because it uses greedy argmax (no epsilon exploration to mask).

Add min_hold_bars parameter to backtest_env_step kernel. When hold_time > 0 && hold_time < min_hold_bars, override action to maintain current position (same logic as training kernel's Layer 2).

3. Dynamic trade insufficiency threshold

In extract_objective() in hyperopt/adapters/dqn.rs, the trade penalty currently uses a fixed threshold via calculate_trade_insufficiency_penalty(total_trades). Make it dynamic:

// Dynamic threshold based on window size and min_hold_bars
let max_possible_trades = backtest_window_bars / min_hold_bars.max(1);
let min_trades_threshold = (max_possible_trades / 20).max(5);
let trade_penalty = calculate_trade_insufficiency_penalty_dynamic(
    backtest.total_trades,
    min_trades_threshold,
);

This ensures the penalty adapts: with min_hold_bars=5 and 10K-bar windows, threshold = max(500/20, 5) = 25. With min_hold_bars=20, threshold = max(250/20, 5) = 12. Always reasonable.

Files Modified

File	Change
crates/ml/src/cuda_pipeline/gpu_action_selector.rs	DELETE
crates/ml/src/cuda_pipeline/epsilon_greedy_kernel.cu	DELETE
crates/ml/src/cuda_pipeline/mod.rs	Remove gpu_action_selector module
crates/ml/src/trainers/dqn/trainer/action.rs	Remove select_actions_batch()
crates/ml/src/trainers/dqn/trainer/mod.rs	Remove GpuActionSelector field + imports
crates/ml/src/cuda_pipeline/backtest_env_kernel.cu	Add min_hold_bars hold override
crates/ml/src/cuda_pipeline/gpu_backtest_evaluator.rs	Pass min_hold_bars to env kernel
crates/ml/src/hyperopt/adapters/dqn.rs	Dynamic trade insufficiency threshold