{"model_details":{"what":"a three-role LLM council (macro, technician, narrative) and a judge, over a hashed evidence pack; deterministic code sizes the practice trade (nota.risk), never the model","models":{"qwen3-235b-a22b-instruct-2507":159},"prompt_versions":{"v5":141,"v7":8,"v2":6,"v3":4},"sizing_policies":{"v1":148,"v3":8,"v2":3},"receipts":159,"first":"2026-09-07T09:29:57+00:00","last":"2026-11-10T03:16:15+00:00"},"intended_use":["a second opinion on a 7-day direction for a crypto major, with every cited number traceable to its source","grading RYO's own verdicts and trade plans against exchange candles","teaching: a decision that can be replayed, challenged and tampered with on purpose"],"out_of_scope":["placing orders: Nota never trades; every trade is practice money","advice to any person: not financial advice","tokens outside the evidence RYO returns, or horizons other than seven days"],"factors":{"symbols":{"SOL":55,"BTC":52,"ETH":52},"evidence_sources":{"price_crosscheck":158,"technicals_crosscheck":158,"market_overview":153,"narrative_convergence":153,"positioning_check":153,"liquidity_check":149,"crowd_odds":141,"analyze_token":96,"deep_analysis":96,"compare_tokens":96,"monitor_market_sentiment_shift":92,"news_verify":11,"network_check":8,"attention_check":8,"onchain_check":8}},"metrics":{"brier_judge":0.2433,"brier_base_rate":0.2387,"skill_vs_base_rate":-0.0194,"scored_calls":82,"independent_weeks":15,"brier_index":50.7,"brier_index_definition":"100 * (1 - sqrt(Brier)); 100 perfect, 50 always 0.5, 0 always certain and wrong (ForecastBench)","murphy_decomposition":{"reliability":0.067,"resolution":0.0198,"uncertainty":0.2018,"base_rate_up":0.28,"residual":-0.0057,"reading":"calibration is the larger problem"},"reliability_bins_outside_90pct_band":3,"enough_to_read":false,"reading":"the judge does not beat the base rate on these calls"},"evaluation_data":"every receipt whose seven-day horizon resolved on OKX hourly candles (nota resolve); drill and recorded-fixture receipts are never scored","training_data":"none: Nota trains no model; the council is prompted, and its calibration weights are fitted only on calls already resolved before the receipt they weight","ethical_considerations":["a confident-sounding model can mislead: the receipt shows the probability, its Brier history and the falsifier beside every call","the evidence comes from one vendor (RYO); its failures are recorded, not hidden"],"caveats":"see /limitations; scores on fewer than the meaningful number of independent weeks are provisional","see_also":{"scores":"/api/scores","benchmarks":"/api/benchmarks","status":"/api/status","limitations":"/limitations"}}