A remix is a fork with a burden of proof. It begins with an existing lens, model, or decision rule. It changes a declared part of that parent. It preserves the lineage needed to reproduce both versions. Then it asks a narrower question than “did this beat the market?” The question is “did this change improve on the parent?”
That distinction matters in NFL Week One, when a new filter can look persuasive after a small set of early decisions. The child may show a cleaner record because it removed difficult cases. It may also show a cleaner record because the remaining sample happened to break its way. Lineage makes those explanations testable.
A remix starts with a parent
The parent is the exact version that existed before the fork. A useful lineage record identifies the parent, the child, the author of each, the change time, the evaluation window, and the rule that changed. It also preserves the unchanged parts. Without that record, a later reviewer cannot tell whether the child added one filter or quietly replaced the entire decision process.
A remix can change a feature, a threshold, a market, a timing rule, an eligibility filter, a weighting scheme, or the way an output becomes a call. The change may be small in code and large in effect. A filter that removes games after a late price move can alter the sample, the average price, the grading context, and the kinds of misses that remain.
The parent must stay inspectable after the fork. A child should not overwrite the object it claims to improve. Preserving both versions lets the product show the branch point, the changed rule, and the evidence accumulated after the change. It also stops a strong child from erasing the misses that motivated the fork.
Lineage is part of the result
Lineage answers who changed what, from which version, and why. Attribution is not a courtesy layer added after performance is known. It is the chain that lets a reader assign credit correctly. The parent author owns the original claim. The remixer owns the declared change. The child record belongs to the child version.
A fork should carry a short change note that can be tested. “Added weather” is not enough if the implementation actually changes the market window and removes road games. “Require the parent signal plus the declared weather condition” is closer because it identifies the added gate and leaves the parent signal intact.
The lineage card should also state whether the remix uses the parent's historical data, a fresh holdout, or both. Reusing historical rows can be appropriate for exploration, but it creates a different evidentiary status from a child evaluated after the rule was frozen. A reader needs that distinction before comparing records.
The parent test is not the market test
The market test asks whether a strategy's calls performed better than a relevant market baseline. The parent test asks whether the child's change improved on the strategy that already existed. Those are separate questions. A parent and child can both compare favorably with the market while the child adds no measurable value. A child can also improve on a weak parent and still remain unfit for public promotion.
A fair parent test uses matched conditions. Parent and child should be evaluated on the same eligible decision times, the same market snapshots, the same grading rules, and the same outcome window. When the child declines a call the parent made, that abstention is part of the comparison. It cannot be discarded merely because there is no child grade.
The cleanest comparison is paired. For each parent-eligible opportunity, record the parent decision, the child decision, the price each could have taken under its contract, and the final grade. The analysis can then distinguish changed calls, retained calls, and filtered calls. That is more informative than placing two aggregate records beside each other and assuming the difference came from the remix.
Attribution follows the changed rule
A remix earns credit for the marginal contribution of its change, not for the entire parent edge. If the child keeps nearly every parent call and modifies only a narrow subset, the evaluation should focus on that subset as well as the full child record. Otherwise the parent's established performance can dominate the aggregate and make the new rule look stronger than the evidence supports.
The same principle applies to losses. When the child introduces a new miss, the record should show whether the parent would have avoided it, made the same call, or made the opposite call. A remix is not audited by collecting only the cases where parent and child disagree in the child's favor.
Public attribution should remain visible through later forks. A remix of a remix carries the full path, not just the immediate parent. That path lets the reader trace an idea back to its original claim and see which branch introduced each condition. It also prevents a popular descendant from presenting inherited work as if it began at the latest fork.
Filters spend sample size
A remix that adds a filter loses n before it has any evidence of gaining hit rate. Most remixes add a condition. The condition may be sensible: require a stronger signal, exclude unstable prices, restrict the decision window, or remove a context where the parent has struggled. But every added filter spends n. Some parent-eligible opportunities no longer reach the child. The child has fewer settled decisions before any improvement can be trusted.
This trade is often obscured by a higher hit-rate point estimate. Removing hard cases can improve the estimate, but a smaller sample also moves more easily. The uncertainty can widen even as the headline looks better. A responsible remix card shows parent n, child n, the retained share of opportunities, and the uncertainty around the difference.
Filters can also change price quality. A rule that waits for more confirmation may enter later, after the market has moved. The child may win more often while taking worse numbers. A rule that enters earlier may improve closing-line value while accepting more final-result variance. The remix has to state which objective it is trying to improve.
How smaller samples create false improvements
The simplest illusion comes from selecting after the outcome. A reviewer notices that the parent lost in one visible context, adds a filter excluding that context, and reports the recomputed record on the same history. The child is now optimized to a known miss. The backtest can describe the rule, but it cannot establish that the rule will repeat.
A subtler illusion comes from trying many candidate filters and publishing only the best one. Each filter may look reasonable. The selection process still searched the same noise. Lineage should record the candidates or at least the fact that a search occurred, and the published child should face a holdout it did not help select.
Context leakage creates another false gain. If the child is evaluated only in the regime that inspired it while the parent is summarized across a broader regime, the comparison is not matched. The correct parent baseline is the parent's performance on the same opportunities the child could see, followed by a separate accounting of the opportunities the child removed.
What counts as a real improvement
A credible child improves a declared objective under matched evaluation and keeps that improvement outside the data used to design the change. The evidence should show the parent and child on shared opportunities, the effect of abstentions, the change in n, the uncertainty around the difference, and the behavior of the change across relevant contexts.
For a betting lens, final ATS grades and closing-line value answer different parts of the question. ATS shows settled outcomes under the spread contract. Closing-line value shows whether the captured decision tended to beat the later market number. A remix that improves one while degrading the other has made a trade, not an unqualified improvement.
Stability matters as much as the aggregate. A child that gains entirely in one narrow slice may be useful as a specialized branch, but it should not replace the parent everywhere. A child whose advantage disappears after the branch point may remain exploratory. A child whose evidence weakens over time may need a decaying gate even if the full-history record still looks favorable.
Week One rewards restraint
Week One offers many plausible reasons to fork: changed personnel, new coordinators, altered roles, different market timing, and stale assumptions carried from the prior season. Those reasons can justify a hypothesis. They do not supply the grade. The child still needs a frozen rule and a record that starts before outcomes are known.
An early remix can therefore be useful while remaining in grading. The public card can show the parent, the exact change, the calls the child would remove or alter, and the evidence still missing. That is more informative than presenting a short favorable run as proof that the fork solved the parent's weakness.
No-call behavior deserves special attention. A filter may improve the child's apparent record by declining most opportunities. That can be a valid design if selectivity is the objective, but the card should show coverage beside performance. A strategy that rarely acts answers a different operational question from a parent built to cover the full board.
What a publishable remix card shows
The minimum card includes the parent identity, child identity, authorship, branch point, declared change, evaluation status, parent n, child n, matched comparison, uncertainty, and the objective being tested. It should state whether the result comes from exploratory history, a holdout, or forward grading.
The card should lead with the trade. If hit rate rose while n fell sharply, say both. If closing-line value improved while final ATS results remained uncertain, keep those measures separate. If the child only works in a narrow context, name the context. If the parent still performs better on the full board, do not hide it.
The reader should be able to move backward through lineage and forward into later forks. That turns remixing from a gallery of disconnected ideas into an auditable research tree. It also makes failure productive: a child that does not improve the parent still records which change was tried and under what conditions.
Sources and method
This article describes the remix contract without presenting a live parent, child, or Week One record. The required evidence belongs on the live remix surface: parent version, child version, attribution, changed rule, matched opportunities, sample sizes, final grades, closing-number comparisons, uncertainty, and evaluation status.
The core test remains narrow. Preserve the parent. Name the change. Compare on matched opportunities. Charge every filter for the n it removes. Promote the child only when the evidence supports improvement beyond a smaller sample.
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.
Breakeven win % at common American odds
The win rate you need to break even at each price. Pick odds shorter than -150 and you must win >60% just to stay flat — a hurdle most casual handicappers never sustain.
