A working description of the system: the parts, the rules, and what we guarantee. This is the same list we hold ourselves to when a number looks off.
The design rule that makes the system checkable: language models propose, deterministic math decides. The ensemble is deliberately small, and each member has one job.
Produces the core numeric projections from the week's ingested data. The strongest model on the roster. Its output is a candidate, never a final answer.
Its only job is to find the failure modes that cost money: hidden traps, committee risk, matchup cliffs. It proposes vetoes, not picks.
Fast and cheap. It parses breaking news and market movement into bounded signals, each capped so a single story can move a projection by only a fixed amount.
Integer-linear programming over hard constraints, no language model involved. The same inputs always produce the same output. It is the only component allowed to make the final decision.
Every number we report is pulled during the session and carries a source and a timestamp. If a figure is not in the week's manifest or the results database, the answer is "no data." We do not fill the gap from memory.
Forward data is scoped to a single week and valid only for that week. Historical results live in a separate store, keyed by their own week. Nothing leaks across the boundary.
No single source is trusted alone. Rankings, projections, and market signals are checked against each other; disagreements are surfaced, not averaged away.
Every reported figure carries its source and a fetch timestamp. A projection with no source tag is treated as garbage and discarded.
If the data is older than it should be, we say so and mark the affected figures. We do not quietly proceed on stale numbers.
After actuals land, we measure each model's mean absolute error and reweight the projection blend toward what has actually been right. The system gets sharper with every week, and it can tell you which source to trust.
Model weights are a function of measured accuracy, not confidence. A model that has been wrong loses influence automatically.
Backtests verify that no future data leaked into the forward set. A result the guard cannot certify is not trusted, even if it looks good.
We establish a simple baseline and measure every model against it before scaling spend. A model cannot justify its cost until it beats the cheap alternative.
Entries, stakes, wins, and costs are kept as a real ledger. Edge is measured against money actually spent. If wins do not exceed cost, the system is told so.
Aperture Analytics is the statistics and data science division of The Camera Group.