{ "target_metric": 2990, "metric_mode": "lower_is_better", "noise_floor": 10, "min_seeds": 3, "source_refresh_hours": 12, "warden_threshold": 0.45, "harness": { "agents": { "rules_file": "AGENTS.md", "autonomy": "keep_loop_running_unless_warden_blocks" }, "goal": { "file": "goal.md", "target": "reach benchmark target in fewer steps with valid statistical receipt" }, "plan": { "file": "plan.md", "state": "mutable frontier" }, "thread": { "file": "scratchpad/THREAD.md", "state": "durable mission log" }, "source_refresh": { "cadence_hours": 12, "sources": ["upstream PRs", "records", "papers", "local best"] } }, "experiments": [ { "id": "contra_muon_stack_v3", "idea": {"name": "Contra-Muon stack", "lineage": ["upstream PR", "previous local frontier"]}, "patch": {"files": ["train_gpt.py"], "summary": "optimizer/schedule/init mutation"}, "config": {"schedule": 3050, "lr_power": 1.2}, "run": {"log": "runs/contra_muon_stack_v3.log"}, "metric": {"value": 2930, "target": 2990}, "judge": {"seeds": 31, "noise_floor": 10, "rule_pass": true, "reproducible": true}, "warden": {"source_stale_hours": 2, "local_loop_score": 0.1, "overfit_risk": 0.15}, "compute": {"gpu_hours": 42, "tokens": 120000000} }, { "id": "novelty_only_deadend", "idea": {"name": "unvalidated novelty optimizer"}, "patch": {"files": ["optimizer.py"], "summary": "novel but weak idea"}, "config": {}, "run": {"log": "runs/novelty_deadend.log"}, "metric": {"value": 3105, "target": 2990}, "judge": {"seeds": 1, "noise_floor": 10, "rule_pass": true, "reproducible": false}, "warden": {"source_stale_hours": 20, "local_loop_score": 0.7, "overfit_risk": 0.4}, "compute": {"gpu_hours": 12, "tokens": 50000000} } ] }