Run autonomous overnight research loops on AXIOM's engine parameters. Uses Hermes Agent to iterate E1 hyperparameters, test corpus additions, and validate backtest accuracy — generating merge/reject decisions logged as research trajectories.
Autonomous overnight parameter research on AXIOM's core engines. Inspired by Karpathy's autoresearch pattern — fixed evaluation budget, single-metric verdict (adjacent accuracy on held-out backtest), keep or revert.
Each experiment loop:
backtests/ held-out periodREGIME_K factor tuning (currently hand-estimated)Adjacent regime accuracy — same metric used in Backtest_2024.py.
Counts as correct if predicted regime is the true regime OR an adjacent regime
(e.g. predicting late_cycle when true is recession_risk counts as adjacent correct).
Target: >65% adjacent accuracy on held-out 2023–2025 period.
Run an overnight research loop on AXIOM's E1 regime classifier.
Target: improve adjacent accuracy from current 68% baseline.
Budget: 10 experiments maximum.
Working directory: D:/AXIOM/Backend
Backtest script: Run_5_Year_Backtest.py
Metric: adjacent_accuracy field in backtest output JSON.
Keep changes if adjacent_accuracy improves by >1pp. Revert otherwise.
Log each experiment result to research_log.jsonl.
Experiments are logged to:
D:/AXIOM/Backend/trajectories/research_runs.jsonl
Each entry contains: hypothesis → parameter change → backtest result → verdict. Accumulates into a research history that can be used for meta-learning over AXIOM's own improvement process.