LinePointer / Methodology
The current app shows experimental NFL forecasts and their results. A useful method explains what was known, when the prediction was saved, and how the outcome is counted.
Open predictions takes you to LinePointer’s public Picks page on linepointer.fulcrumtechnologies.io, leaving these company pages. Browse saved forecasts and game explanations without signing in or generating a new prediction. Results & audit opens the separate final-pick evidence record.
The research workflow
Follow the inputs, the saved times, and the result. Each stage adds context to the number.
A model's selection is the team it expects to win. Its win chance describes estimated uncertainty. A 60% estimate is not a measured win rate or proof of calibration. A projected score is an estimate, not a promised final score.
Different model families and forecast editions keep separate records. Market consensus is a comparison baseline, not an independent trained-model vote.
Saved context can include team performance, matchup information, public player-availability reports, and applicable weather. The app records the source vintage and shows missing or unverified inputs. An unpublished injury report does not mean a team is healthy.
The October 2 input correction added separately verified current-season passing, rushing, defensive matchup measures, and recent results to future AI score forecasts. Earlier forecasts were preserved. Improved accuracy has not been established by that repair.
The original saved forecast, the selected final pick, and its actual publication receipt have distinct times. Early and game-day editions are not interchangeable. A late public posting is not represented as a public pre-kickoff commitment.
In the current experimental record, eight exact forecasts and owner selections saved before kickoff were finalized and published late because of technical delivery issues. They count on the disclosed saved-forecast basis; no pre-kickoff publication receipt is claimed for them.
Accuracy is wins divided by wins plus losses for the record's eligible final selections, including the eight explicitly disclosed late-finalized selections. Ties and voids are excluded. Missing and unpublished picks remain visible; games without a qualifying counted pick stay outside the accuracy denominator.
This is an experimental final-pick record. It is separate from the future governed Tracked Calls and Official Picks program; entries are not retrospectively promoted into that program.
The public record links each ledger entry to the previous entry using SHA-256. The commitment identifies the saved forecast, model request, prompt, and input packet by their hashes. Result corrections append a link to an earlier grade.
A tamper-evident chain can reveal inconsistent entries. It is not independent timestamp proof and does not prove the operator could never rewrite the whole store. The requested AI model identity is not independently provider-attested.
The forecasts remain unpromoted. An early season sample does not establish reliable probability calibration, superiority to the market, or profit. Software checks prove software behavior; they do not establish predictive quality.
Stronger claims require registered out-of-time evaluation, prospective evidence, appropriate comparisons, and uncertainty. No forecast guarantees an outcome.
Reading the app
The selection, estimated score, important data warnings, and results. Start here.
The same saved forecasts, with win chances, model comparisons, and more game context.
The protected research workspace, source evidence, model tests, and system diagnostics. A view preference does not grant access.
Probability quality
Evaluate on later, unseen periods so future information cannot leak into the past.
A scoring rule that penalizes confident errors. Lower is better.
The mean squared difference between a probability and its binary outcome. Lower is better.
Whether predictions near a given probability occur at roughly that rate across enough observations.
Public browsing uses saved output. Visitors can browse saved predictions and their explanations in the public Picks view, and inspect published results in the separate audit record. Research, new forecasts, and publication remain controlled by the owner. Private source text, prompts, service credentials, and raw provider feeds are not part of this website's results view.