Comparing the two studies against results…
Head-to-head overall
Winner-pick accuracy across both studies, each graded against the same ESPN results on its own eligible subset — the denominators differ (the automated pipeline predicts per model/prompt/timing on live matches; the manual study grades 5 frontier models × 3 prompt conditions per fixture), so treat the delta as directional, not a controlled experiment.
Overall accuracy detail
Full counter breakdown for each study side by side.
Per timing
Accuracy as the match approaches (24h before → at kickoff → halftime) — both studies use the same three timings.
Group vs knockout
Winner-pick accuracy split by stage. Knockout "winner" is the team that advanced (incl. extra time + penalties) in both studies.
Confidence calibration
Are high-confidence picks actually more accurate? Bands over each study's reported confidence (%).
Per model
Union of both studies' models — the populations differ (pipeline cloud models vs the study's Gemini/GPT/Claude/Kimi/Llama), so a model present in only one study shows an em-dash on the other side.
3-way Brier score
Multiclass Brier score of each study's 3-way outcome distributions — lower is better (random = 0.667). Delta = manual − automated; negative means the manual study scored better on its own (different) fixtures.
Market baseline (manual study only)
Only the manual study injected raw Kalshi odds into its prompts, so "did the model beat the market?" is only measurable there — on every paired fixture the market favorite is graded the same way. Full per-model breakdown on the Manual page.
Bet performance (market odds, $1,000 bankroll)
Each study's bankroll allocations graded at the market price captured for the bet's timing: a bet of stake on an outcome priced at implied probability p wins stake × (1/p − 1) if the outcome occurred, loses −stake otherwise (bets without a usable price fall back to even-money). ROI is identical across bankroll sizes — absolute dollars differ. Per-model/per-prompt bet tables live on the detail pages.
Drill-down
Per-match/per-fixture breakdowns live on the detail pages — the two studies key matches differently (internal match IDs vs study Excel IDs), so there is no joinable per-match table here.