ALPHA FORGE adds sandbox that skips the prediction_ledger

The root of the problem was structural: ALPHA FORGE records every prediction in a prediction ledger, and that ledger feeds both continuous learning and the displayed real-world win rate. Any casual trial of a stock through the normal path would mix subjective, biased samples into the same pool as the strict morning selection process. The developer had Claude Code audit the code first, which revealed that the inference orchestrator only returns results and that the daily batch is the sole writer to the ledger. That separation is what made a clean sandbox possible without duplicating the scoring logic.

The second guard came from a setting that had been sitting unused: a $5 daily LLM budget defined in the config file but never referenced. The sandbox run checks usage against it and raises an error once the cap is reached. The developer also kept a rule that any run incurring API charges must pause for human confirmation immediately before execution, so approving a design never amounts to a blank check on spending.

On the first real run, a large-cap stock came back rejected at 38% confidence, just under the 40% threshold. The developer noted that only the normal "adopt" path had been seen in the test environment, so the rejection flow reaching the screen intact was itself useful. The ledger's 76 entries were unchanged, with six trace rows added separately.