How to read Pick6Labs

1. Identify the kind of information

A player’s recorded passing yards are a source statistic. A projected future passing total is a model estimate. A bookmaker’s offered line and price form a market quote. An optional Groq explanation is an interpretation of supplied inputs. One does not prove the others: a detailed AI narrative cannot establish a missing injury update, route count or actual payout.

Basic player profiles do not require an AI request. A profile can load with verified identity even if multi-season statistics or scouting evidence are unavailable. Missing data should remain visibly missing.

2. Player coverage comes before a grade

The ranking population is built from current rosters plus completed regular-season game participation and box scores. Valid, explicitly non-DNP game records establish appearances; roster membership alone does not. Numeric source identities deduplicate players across weeks and transfers. Historical participants remain included after leaving a current roster. Preseason, future and unfinished games do not contribute.

The Players Hub reports its checked teams, completed games and verified participant count. If a source refresh fails, the previous complete snapshot remains available with stale status rather than being silently replaced by a partial population. These checks cannot guarantee that an upstream provider never makes an error; corrections still matter.

3. The ranking rule is not a universal talent grade

The box-score index compares available production and efficiency within position/role peer groups. It combines weighted feature percentiles and shrinks small samples toward neutral before mapping them to a 70–99 OVR scale. At least three comparable peers are needed. The scale is a presentation convention, not a probability of success.

Established samples require two completed games and position-specific exposure: 50 quarterback pass attempts, 30 running-back touches, 15 receiver/tight-end targets, five kicker field-goal attempts or five punts. Defensive ratings need recorded defensive evidence. Smaller positive samples can receive provisional ratings. Participants without defensible grade inputs remain ranked by verified games played, with ties; linemen and long snappers are not assigned fictional blocking or snapping scores.

Established ratings sort first, provisional ratings next, then participation-only ranks. Rank is assigned within the selected category before searching, stat sorting or pagination. A category rank and an all-players rank have different scopes. Weekly movement compares the same rule at adjacent completed-week cutoffs, not a change in independently measured talent.

See the worked small-sample guide and expand a player’s rating evidence in the Players Hub.

4. Game forecasts and market comparisons

The game engine combines team-strength information, matchup/context adjustments, market-aware components and simulation/distribution calculations. Different components use different sampling budgets. It is not accurate to describe every response as 10,000 complete play-by-play simulations, or every output as independent of market information.

Winning the game, covering a handicap and going over a total are different events. Integer lines can push. Evaluating a quote requires its exact event, market, side, bookmaker, line, price and observation time. A default price is not a substitute for a missing actual quote. Expected value depends on both the forecast and the payout, and apparent positive EV does not by itself establish that the forecast is calibrated.

Readiness labels are part of the result. A WATCH or PASS result should not be promoted into a confident pick merely because a model has a numerical lean. Read the spread and push example.

5. Freshness is separate from page refresh time

Prepared boards serve compatible, already-evaluated results while background computation handles updates. They preserve underlying source ages and invalidate when relevant inputs change or their validity window ends. A temporary preparing response must not display expired selections as actionable. Reopening or refreshing a page does not make an old quote new.

Ranking snapshots refresh in the background after their two-hour lifetime. Their source caches have separate lifetimes: schedule checks are shorter-lived, current rosters are cached for two hours, and older completed-game records can be cached for a day to limit repeated provider requests. Those choices trade correction latency against load; they are not a claim of instantaneous upstream data.

6. Historical records and honest evaluation

Recorded pregame snapshots preserve the information saved at their cutoff. Later live forecasts answer a different question and should not overwrite the earlier grading pick. A season record needs its population, number of graded events, cutoff and treatment of pushes or ties. It is not proof of performance on markets outside that record.

Reliable predictive claims need a defined test population, information available before each event, a held-out evaluation and uncertainty. We make no blanket guarantee of calibration, profitability or coverage of every historical market. Experimental ideas and qualitative explanations do not become validated production improvements just because they are detailed.

Sources, corrections and further reading

• ESPN public athlete, roster, schedule and completed-game data: ESPN NFL.
• Open football data and documented metrics: nflverse and the nflfastR beginner’s guide.
• Provider reporting, our derived rules and illustrative examples are distinct; a citation does not imply endorsement.

Report an issue with the game/player ID, page, relevant timestamp and a source link through Contact & corrections. Do not post credentials or sensitive account information in public discussion.