Innovation in Baseball Intelligence

Before a cached answer is ever reused, eight independent validation signals — team identity, stat mentions, numeric thresholds, position, stat type, comparison direction, grammatical number, and place names — must all confirm the new question truly means the same thing.

Why it matters: Prevents incorrect cache hits — "Tigers with 50 homers" will never return cached results for "Tigers with 50 steals." Speed never comes at the cost of a wrong answer.

A deterministic, ordered pipeline of more than a dozen repair stages inspects every AI-generated query plan and corrects the mistakes language models predictably make: truncated decimals, off-by-one thresholds, age/stat confusion, league misreads, missing seasons, and misclassified stat types.

Why it matters: Turns probabilistic AI output into reliable structured queries. The AI supplies the understanding; the pipeline guarantees the precision.

Deterministic pattern extraction resolves roughly nine in ten questions instantly — no model inference required. Only genuinely ambiguous questions escalate to the full AI pipeline, and even those are answered once, then validated and reused.

Why it matters: Near-instant answers for common questions, dramatically lower compute cost per query, and full AI comprehension held in reserve for the questions that need it.

Five coordinated cache layers — final results, query classification, semantic similarity, AI tool calls, and normalized query arguments — each capture a different stage of the pipeline, and the repair pipeline always re-runs on layers that store pre-repair output.

Why it matters: Repeat and near-repeat questions skip the expensive stages entirely, while cached raw AI output still passes through every correctness check. Fast paths never bypass safety.

One master definition per stat powers slang recognition, database routing, threshold parsing, and display labels across the entire system — a single canonical source shared by the AI prompts, the repair pipeline, and the UI.

Why it matters: Understands "dingers," "long balls," "taters," and "home runs" as the same stat — one source of truth, zero ambiguity, and every new stat inherits the whole pipeline automatically.

Distinguishes between singular and plural intent in natural language queries to control result set size.

Why it matters: Knows "the Dodger with most HR" means one player, while "Dodgers with most HR" means a ranked list.

Every filter in a question is classified by stat family and routed to the right store — traditional season stats, Statcast advanced metrics, WAR, awards, single-game logs, or team aggregates — then intersected into one coherent answer, with era-appropriate floors applied automatically.

Why it matters: Enables mixed questions like "xERA under 3 with 30+ saves" across entirely different data sources, without users ever knowing which table holds what.

A champion/challenger loop continuously tunes and trains new game-prediction models, then judges them with a paired statistical-significance gate on a frozen holdout set. A challenger is promoted only when its improvement provably clears measurement noise.

Why it matters: Predictions genuinely improve as data grows — and never regress chasing random fluctuation. "No promotion" is an honest answer, not a failure.

Players who change teams mid-season produce fragmented stat rows. The engine detects context — team-scoped questions aggregate the relevant stints and recompute rate stats correctly; league-wide questions use true season totals.

Why it matters: Counting stats sum, rate stats recompute, and WAR combines correctly — so a July trade never produces a wrong batting average or a double-counted season.