AI ASSISTED BASEBALL RESEARCH WITHOUT THE AI SLOP
Baseball research · AI workflow · Evidence boundaries
A public control system for AI-assisted baseball research and analytical products, built to keep questions, evidence, verification, and human judgment connected.
A capable model can draft code or analysis quickly, but fluency does not prove that the question is defined, the evidence is complete, or the result deserves to influence a decision. Agent+ puts a visible workflow around the model so each task has a contract, bounded inputs, deterministic checks, independent review, and a clear stop condition.
Define the baseball question first.
Does prior pitch-family exposure improve a defined chronological prediction target versus a count-and-location baseline?
Decide whether the feature may advance to a pre-specified out-of-sample evaluation.
Null if there is no improvement. BLOCK if date coverage or temporal order cannot be verified.
Question → contract → evidence → review.
Question
Define the decision before the work starts.
Name the decision, population, date boundary, target, and null path.
- Authority
- Human owner
- Artifact
- Question brief
- Stop when
- No decision, population, or null path.
Release files.
Synthetic baseball question
A historical prediction question with a decision, comparator, time rule, null path, and BLOCK path.
Research contract
A reusable contract template that freezes population, target, evidence, stop conditions, and permitted claims.
Authority-aware work
Deterministic tools, bounded assistance, implementation work, independent review, and human ownership have different jobs.
Evidence record + BLOCK
A missing partition or invalid receipt stays visible. It is not rewritten into a positive baseball finding.
What this package does not prove.
Agent+ shows how to structure, check, and stop a research task. The package is synthetic. It is a reference implementation, not evidence that a baseball feature works, improves a forecast, or authorizes a decision.