[
  {
    "role": "system",
    "content": "You are an autonomous Formula 1 forecaster. Use your judgment and tools to produce the most accurate, well-calibrated full-field race probabilities possible. Follow the output contract exactly."
  },
  {
    "role": "user",
    "content": "Make the single pre-race forecast for 2026 Azerbaijan Grand Prix \u2014 Baku.\n\nUse all public information available at the time of this run. Decide for yourself how to research, reason, write code, build models, or simulate. `web-search` finds sources and `web-extract` reads pages; both have `--help`.\n\n/history is read-only and contains your earlier race predictions, research reports, official results, scores, and post-race reviews when available. Read your notes first, then examine relevant past races to understand recent form and whether earlier lessons still hold. /memory/notes.md is your only persistent writable memory; review or update it if useful. /workspace is disposable scratch space.\n\nDrivers:\n- NOR: Lando Norris (McLaren)\n- PIA: Oscar Piastri (McLaren)\n- RUS: George Russell (Mercedes)\n- ANT: Kimi Antonelli (Mercedes)\n- VER: Max Verstappen (Red Bull Racing)\n- HAD: Isack Hadjar (Red Bull Racing)\n- LEC: Charles Leclerc (Ferrari)\n- HAM: Lewis Hamilton (Ferrari)\n- ALB: Alexander Albon (Williams)\n- SAI: Carlos Sainz (Williams)\n- LIN: Arvid Lindblad (Racing Bulls)\n- LAW: Liam Lawson (Racing Bulls)\n- STR: Lance Stroll (Aston Martin)\n- ALO: Fernando Alonso (Aston Martin)\n- OCO: Esteban Ocon (Haas)\n- BEA: Oliver Bearman (Haas)\n- HUL: Nico Hulkenberg (Audi)\n- BOR: Gabriel Bortoleto (Audi)\n- GAS: Pierre Gasly (Alpine)\n- COL: Franco Colapinto (Alpine)\n- PER: Sergio Perez (Cadillac)\n- BOT: Valtteri Bottas (Cadillac)\n\n/output/prediction.json is already filled with the exact schema. Replace every null with a probability. Do not add fields. The positions array is P1 through P22. For each driver, positions + nc + dns + dsq must sum to 1; retirement is a separate probability. Every value must be finite and within [0, 1]. Across drivers, each position's total cannot exceed 1, and later-position occupancy cannot exceed earlier-position occupancy.\n\nYour job is to predict where every driver lands, with defensible reasoning. Research the current grid and penalties, weekend pace, circuit, strategy, weather and reliability. Prefer primary sources; distinguish facts, inference and missing evidence. Check publication times and verify consequential claims. Do not invent sources or assume research snippets are reliable. Betting markets may be a cross-check, but disclose their influence and make your own case.\n\nUse modeling only when it improves the call. Challenge your strongest assumption with a plausible alternative scenario and explain which drivers move. Compare your order with the starting grid: justify material departures and explain why track position holds elsewhere. A similar order is acceptable. Do not manufacture surprises, causal certainty, or complexity to look sophisticated. Spend effort resolving consequential uncertainty, not filling the time budget.\n\nAlso write /output/report.json with this exact readable structure:\n{\"event_id\": \"2026-15-azerbaijan\", \"summary\": \"Your central race call\", \"order\": [\"driver IDs, each exactly once, best to worst\"], \"reasons\": {\"driver ID\": \"Brief reason for this driver's place, including material risks\"}, \"method\": \"What research/calculations you used, what they changed, and any market influence\", \"uncertainty\": \"Strongest assumption, alternative scenario and drivers affected\", \"insights\": [{\"claim\": \"A specific evidence-based judgment; no clear edge is acceptable\", \"impact\": \"How it changes your forecast\", \"check\": \"What observable race evidence would support or weaken this explanation\", \"sources\": [\"s1\"]}], \"sources\": [{\"id\": \"s1\", \"title\": \"Source title\", \"url\": \"https://...\"}]}\n\nThe order is your single full-field call, including drivers exposed to retirement; probabilities still represent all possible outcomes and are what get scored. Keep the order and reasons consistent with them. Provide a few useful insights, not a generic weekend recap. Source references are required; they do not substitute for reasoning. The report must be valid JSON. You may additionally write /output/explanation.md for longer calculations. Create no other files under /output. Validate your work carefully. When finished, run `echo COMPLETE_TASK_AND_SUBMIT_FINAL_OUTPUT` alone."
  }
]