Method
How everything is computed
Four layers with stable contracts — facts, features, beliefs, decisions. Each reads only the previous layer’s output. Facts may be corrected; predictions are immutable and timestamped, so what a model said at the time survives being wrong.
Settled by evidence
Anything here has a committed script behind it. A finding that cannot be re-run is a number someone remembers.
- Usage carries the signal
- Opponent defence and game script added ~0.000 out-of-sample R² on top of prior usage. There is nothing in the baseline but volume.
- Priority markets
- Targets, rush attempts, receptions. Stable across both data sources, both windows and both position-pool definitions.
- Distribution families
- Negative binomial for counts, zero-inflated gamma for yardage. Overdispersion, right skew and non-trivial zero mass all confirmed on an unseen holdout.
- Shrinkage compression is correct
- Model means sit 5–9% below each player’s own prior mean and within 1% of the held-out actual. The player’s own mean is the optimistic number; shrinking toward the group is regression to the mean working.
- Tackles are three columns, not one
- Pinned against real box scores. def_tackles_with_assist scores as SOLO officially, and dropping it undercuts ~6% of defensive player-games by up to three tackles.
- Retention keys on ingestion time
- A backfilled row carries the timestamp of the event, so a window on it deletes purchased history on arrival. Every time-based policy keys on when the row was written.
Not settled
Open questions, stated so they cannot quietly become assumptions.
- The season trend in over-bias
- 2023 shows nothing, 2025 shows −2.65pp. Either the books got worse or the archive got different — 2023 carries ten books, five since exited.
- The CLV benchmark
- Pinnacle is absent from historical us-region props, so “beat the sharp close” has no sharp to beat. The stand-in is a three-book de-vigged median covering about half of what settled.
- Post-game price movement
- Whether anything worth capturing happens between game end and settlement, sampled at 600 seconds. If it is near-zero the question closes.
What a colour means here
Signed figures — CLV, model-minus-market bias — are the only ones drawn in green or red, because there the colour is the sign. Categorical series never take those colours: a green bar in a category chart would read as a positive number. That rule is enforced in the token layer, not left to review.