A real person becomes a lens only after the tape survives a chain of increasingly strict contracts. The path is tape → mentions → actual-call lens → style model → CLV-ranked leaderboard. Each step adds structure. None is allowed to rewrite the step before it.
That separation matters most in NFL Week One, when a short clip can travel farther than its original context. A speaker may discuss a team, praise a player, reject a price, and never make a bettable call. Turning that segment into a side after the game would create a record the speaker never earned and the reader cannot audit.
Start with the tape, not the conclusion
The source object is the original segment. It needs the speaker, publication, recording time, game context, and a link that returns the reader to the tape. A transcript helps search, but the transcript does not replace the recording when tone, qualification, or a correction changes the meaning.
The first pass records mentions. A mention says the person discussed a team, game, market, player, coach, injury, matchup, or price. Mentions are useful because they preserve coverage without forcing every sentence into a forecast. They also create the retrieval set for a later reviewer who needs to decide whether an actual call exists.
A mention must not inherit a side from nearby language. “I like how this defense is built” is not automatically a spread call. “I need a better number” may be a rejection, not support for the current side. A prediction about the season is not a call on the Week One market unless the speaker ties it to that market before the grading cutoff.
Promote only explicit, gradeable calls
An actual-call lens begins when the tape identifies enough of the decision to settle it later. The record needs the event, the market, the chosen side or outcome, the time of the statement, and any stated condition. If the call depends on a price, that price belongs in the contract. If the speaker withdraws or changes the call before the cutoff, the record must preserve that change rather than choosing whichever version graded better.
Ambiguity resolves against promotion. A sentence that could mean a moneyline lean, a spread lean, or a general matchup preference remains a mention. A conditional statement remains conditional. A host repeating another person's view is not the host's call. A panel graphic is evidence only when it clearly attributes the selection and corresponds to the tape.
The actual-call lens grades at final under the rules of the market it captured. Settled calls increase n. Pushes and voids retain their own states. Missing tape, late statements, unclear markets, and retrospective summaries do not enter the scored sample. Exclusions should be visible because a clean record built through silent exclusions is not a clean record.
The actual-call record comes before the style model
A style model is a different product. It learns patterns from the person's attributed history: which matchup features draw attention, how price sensitivity appears in language, which team or market contexts produce a call, and when the person declines to choose. The output is a prediction of how that person would likely see any game, including one the person never discussed.
That output is simulated. It can cover games the person never discussed, which is exactly why it must not be blended into the actual-call ledger. The actual-call lens answers, “What did this person publicly choose?” The style model answers, “Given the learned pattern, what view would the model assign?” Both can be useful. Combining them would make neither auditable.
The label must appear at the point of use, not in a distant methodology page. “Actual call” means linked tape and a gradeable statement. “Simulated take” means a model-generated view with no claim that the person said it. The visual distinction should survive a share card, an embed, and a leaderboard row.
Quotes are evidence, not decoration
Real quotes should be no more than fifteen words, attributed to the named speaker, and linked to the source tape. The limit forces the card to use the quote as a locator rather than reproducing a segment without context. The surrounding summary must distinguish exact language from editorial paraphrase.
A short quote does not relax the grading contract. The linked tape still has to support the market and side assigned to the call. When the quote is too compressed to carry the condition, the card should state the condition separately. When the tape contains no clean call, the product should publish the mention and stop there.
No real person's record belongs in an explainer without the underlying ledger. A plausible scoreboard is still fabricated if the settled calls are not present. The live surface should carry the current record, the tape links, and the exclusions. This article describes the method instead of inventing a named result.
Rank process, then show outcomes
A CLV-ranked leaderboard asks whether the captured call beat the market's later closing number. That is a process comparison. It does not erase the final ATS grade, and it does not turn every favorable move into a winning bet. A call can beat the close and lose the game. It can also win after taking a worse number than the close.
The public row should therefore keep the measures separate. CLV can determine rank because it tests the quality of the captured price against the market's final consensus. The ATS spread record should remain visible with n because readers need to see settled outcomes. If the row uses a headline rate, it should be the ATS spread win rate with n, not a vig-laden ROI figure. A confidence interval belongs beside the rate so a narrow run is not presented with the certainty of a mature sample.
As an illustrative comparison, “100% on one call” does not outrank “62% on fifty-six”; at -110, the break-even rate is 52.4%.
The comparison is deliberately mechanical. The first record contains almost no information about repeatability. The second contains more settled decisions and can support a narrower estimate, although it still does not remove uncertainty. Ranking rules should enforce a sample-size floor before a row is eligible, then show n and the interval rather than hiding them behind the order of the table.
The sample-size floor is a publication rule
The floor answers an editorial question before it answers a statistical one: when is there enough settled work to rank a person publicly? A row below the floor can still exist. It should appear as grading, unranked, or insufficient sample rather than being allowed to lead the board after a favorable opening run.
The threshold must be fixed before reviewing the current leaders. Moving it until a preferred name qualifies converts the floor into a selection tool. The same rule should apply to tails, fades, market types, and time windows. A specialized split may be informative, but it needs its own n and interval and should not borrow the parent sample.
Confidence intervals also prevent false precision between neighboring rows. When intervals overlap substantially, the product should not imply that a small ranking difference establishes a durable gap. The correct reading may be that the available sample cannot distinguish the two records yet.
Week One is a labelling stress test
Week One produces many opinions before it produces many settled calls. That is exactly when mentions, actual calls, and simulated takes are easiest to blur. A preview segment may contain roster analysis but no market. A style model can still generate a view, but the card must say simulated. A later explicit pick can enter the actual-call lens with its own tape and timestamp.
The early board also makes sample context important. A person's historical actual-call record may span different markets, decision windows, and competition phases. The Week One card should show the context used for the displayed record rather than presenting the broadest available sample as if every call were interchangeable.
The misses should remain first-class evidence. A reviewer should be able to open a losing call, hear the original condition, inspect the captured number, and see the final grade. The same path should exist for a favorable call. Auditability cannot depend on the result.
What a trustworthy pundit lens publishes
The minimum public object is compact but strict: named speaker, linked tape, short attributed quote when useful, call or mention status, market, captured condition, decision time, final grade, n, confidence interval, and CLV treatment. A style-model output adds its model label and never claims to be tape.
The leaderboard then becomes a view over evidence rather than a personality contest. It ranks eligible rows under a declared rule. It exposes insufficient samples. It keeps actual and simulated material separate. It lets the reader move from a row back to the calls that created it.
Sources and method
This article contains no named pundit record and no live leaderboard result. Its source contract is the auditable chain described above: original tape, indexed mentions, promoted actual calls, final grades, style-model outputs, closing-number comparisons, sample size, and confidence intervals. Current figures belong on the live surface where each row can link back to its evidence.
The method is conservative by design. Preserve the tape. Record the mention. Promote only the explicit call. Grade it at final. Label every simulation. Rank only after the sample rule is met.
Model calibration: predicted vs observed
Predicted win probability bucket vs the empirical win rate inside that bucket on the test set. Points on the y=x reference line are perfectly calibrated; points below mean the model is overconfident in that bucket.
EV per $100 across win rate × odds grid
Expected value of a $100 stake at each combination of true win rate and market odds. Anywhere the cell is positive you have a long-run profitable bet; the magnitude shows how aggressive Kelly will size it.
