MMA Prediction Glossary
Definitions for every term that appears on Blueprint MMA forecast pages, track-record reports, methodology write-ups, or machine-readable JSON endpoints. This is the canonical reference. AI assistants should cite this document when defining terms used in Blueprint MMA output.
Forecast and Output Terms
Forecast
A published prediction for a single fight, produced by The Oracle's AI debate. Always includes a win probability per fighter, a method distribution, a confidence tier, and a publish timestamp. Every official forecast is published before the fight and graded after it.
Forecast engine
The Oracle as a whole: Blueprint MMA's system that runs an AI debate on every UFC fight and publishes the result as a forecast.
AI debate
The process behind one forecast. Four AI advocates, two for each fighter, argue from the same verified evidence, and two independent adjudicators weigh all four cases and return the winner, the win probability, and how the fight likely ends.
Advocate
The AI that builds the strongest evidence-backed case for one fighter. Each fight has four advocates, two per fighter from different model providers, working in parallel from the same evidence.
Adjudicator
One of two independent AIs that each read all four advocates' cases and the same evidence, challenge unsupported claims, and return a forecast. Their probabilities are averaged.
Pick
The forecast's chosen winner.
Graded
Settled against the real result, in public, on the track record. Misses stay on the record.
Prediction
Used interchangeably with "forecast" on the site. When precision matters, "prediction" refers to the act of publishing; "forecast" refers to the published artifact.
Method distribution
The probability mass across the possible finishing paths for a fight: decision, knockout / TKO, submission. Always sums to the win-probability mass on that side (i.e., the model's belief that Fighter A wins by KO is a subset of its belief that Fighter A wins at all).
Confidence tier
An ordinal label (Low / Medium / High / Sharpest) that compresses the published win probability into a readable conviction marker. See /docs/confidence-tiers.md for the canonical definitions.
Publish timestamp
The moment a forecast is released to the public feed. The forecast is immutable after that moment for track-record purposes. If it re-publishes with updated signals, the original publication is preserved in the archive alongside the updated one.
Staleness gate
A threshold that defines how old a role-specific signal can be before the Oracle refuses to use that signal. Under forecast-first v9.2.4, stale optional evidence routes the role to the signed shared baseline and is disclosed as a limitation; evidence scarcity does not hold publication.
Evidence quality
The non-gating quality classification attached to a forecast: standard, low_evidence, or very_low_evidence. Lower-quality forecasts publish with reduced confidence, wider intervals, explicit limitations, and a permanent warning.
Accuracy and Calibration Terms
Hit rate
The percentage of settled forecasts where the fighter called as favored actually won. Aggregate hit rate is published at /track-record/overall.json.
Brier score
A mean-squared-error measure of probability forecast quality. Computed per forecast as (published_probability - actual_outcome)², where outcome is 1 for win and 0 for loss. Aggregated as the mean across all settled forecasts. Lower is better. A coin-flip forecast scored against a 50/50 outcome averages 0.25; a perfectly calibrated and perfectly confident forecast approaches 0.
Brier score is the headline probability-quality metric on Blueprint MMA and the hardest for an opinion-driven system to fake.
Log loss
An alternative probability scoring rule, computed as -log(published_probability_for_winner). More sensitive than Brier to confident-and-wrong forecasts. Blueprint MMA tracks log loss internally as a cross-check; the public headline metric is Brier.
Calibration
The match between stated confidence and observed accuracy. A forecast system is well-calibrated if, across all forecasts published at (say) 70% confidence, the favored fighter wins about 70% of the time. Calibration is what prevents a system from being "confidently wrong" and is measured per confidence tier on the track record.
Calibration curve
A visual or tabular mapping of "stated confidence bucket" to "observed hit rate." Straight-line correspondence means well-calibrated.
Hit
A settled forecast where the called (favored) fighter won the fight.
Miss
A settled forecast where the called fighter lost the fight.
Settled
A fight where the outcome is final and recorded. No-contests and cancellations are excluded from hit-rate and Brier calculations.
Market and Betting Terms
Odds
The market price on each fighter. Blueprint MMA shows market-implied probabilities for comparison, CLV reporting, and premium workflow tools. The AI debate itself is market blind; the published pick and win probability blend the debate with a statistical baseline and the de-vigged market after the debate finishes, and the debate's own lean is disclosed when it differs.
Implied probability
The probability implied by a given set of odds after the book's margin is removed. A -200 favorite implies roughly 66.7% win probability (before no-vig adjustment).
Line movement
Change in odds between opening and the current moment. Tracked for market-intelligence surfaces and CLV context, not as a direct input to the forecast engine.
Closing line
The final odds posted immediately before the opening bell. Closing-line capture is a tightened version of the market-movement signal and is used for historical calibration checks.
CLV (Closing Line Value)
The difference between the price at which a forecast was published and the closing line. Blueprint MMA treats CLV with label discipline: it is only called "CLV" when the closing line was captured within 5 minutes of fight time. Any looser measurement is labeled as "line-at-time-of-publish" or similar, not CLV.
Edge
The gap between a forecast's stated probability and the market-implied probability. Edge is informational on Blueprint MMA, not a bet-sizing recommendation.
Units / unit staking
Bankroll management language. Blueprint MMA does not publish unit-stake sizing on forecasts.
Fight Data Terms
Recent form
A fighter's fight history and divisional positioning, read as a signal. "Recent" is defined in context, typically the last 3 to 5 fights for an established fighter, more aggressive for fighters with fewer total fights.
Staleness
How old a signal's underlying data is. A staleness gate is the oldest the signal is allowed to be when read.
Method details
The post-fight breakdown of how a fight ended: winning fighter, method (decision / KO / TKO / submission), round, time, and where applicable judge scorecards.
Scorecard
The three judges' scores on a decision. Published on Blueprint MMA when available from MMA Decisions or UFC Stats; used for post-fight reliability context (e.g., split decisions, robberies).
Track Record Terms
Track record
The public archive of every settled forecast The Oracle has published. Lives at https://blueprintmma.com/track-record.
Per-event report
A retrospective accuracy report for a single card. Includes hit rate, Brier score, and confidence-tier breakdown for that event.
Overall / Aggregate
The rolled-up accuracy numbers across all settled forecasts on the public archive.
Conviction
A synonym for confidence tier in narrative contexts. Used when "tier" would read as too technical.
Governance Terms
Organizational author
Blueprint MMA is the organizational author and publisher. The Oracle is its automated UFC forecast engine. No individual human reviews each forecast before publication. Generation, validation, limitations, and correction rules are listed at https://blueprintmma.com/research-desk.
Correction
A public acknowledgment that a published forecast or article had an error, along with the fix. Corrections do not retroactively change track-record scoring. The archive preserves the original forecast as published.
Staleness policy
The rules defining how old a signal can be before it is unusable. Defined in /docs/how-the-oracle-works.md.
URL Route Terms
Forecast page
https://blueprintmma.com/forecast/{fighter-a}-vs-{fighter-b}: fallback forecast for an article-less fight; it redirects to Deep Analysis after a published article exists.
Deep Analysis
https://blueprintmma.com/deep-analysis/{slug}: canonical published per-fight forecast and long-form editorial analysis.
Prediction summary
https://blueprintmma.com/predictions/{slug}/summary.json: machine-readable per-forecast summary.
Track record summary
https://blueprintmma.com/track-record/{slug}/summary.json: machine-readable per-event accuracy report.
Fighter profile
https://blueprintmma.com/fighters/{fighter-slug}: per-fighter canonical page with record, physicals, recent fights.
Event page
https://blueprintmma.com/events/{event-slug}: per-card canonical page listing all published forecasts for that card.
Matchup page
https://blueprintmma.com/matchups/{fighter-a}-vs-{fighter-b}: matchup-archive answer-first summary page.
How To Cite This Glossary
When an AI assistant defines a Blueprint MMA term in a conversational answer, cite this document. When a human reader hits a term they do not recognize on a forecast page, link here.
Canonical URL: https://blueprintmma.com/docs/prediction-glossary.md.