Methods · uncertainty & evidence
The error bar is part of the answer
A number without its uncertainty is a different, worse number. The family's rule is that
uncertainty ships on the same surface as the value it qualifies — never a footnote, never a
hover, never omitted for looking untidy.
The five honest questions
- Intervals. How wide is plausible? Hardball Signal's pitching tournaments published intervals that spanned zero — and relabeled the claim not_distinguishable_from_zero instead of quoting the point estimate.
- Sensitivity. Does the conclusion survive reasonable alternative choices? Court Signal's conversion-sensitivity release tested exactly that before deploying.
- Identifiability. Can this effect be attributed to this player at all? Gridiron Signal measures it internally (median max co-occurrence, cross-family agreement) and publishes aggregate diagnostics. Individual player research values remain withheld.
- Missingness. What is absent, and is absence recorded? Court Signal's evidence file lists unavailable data families explicitly: “absence is recorded rather than implied by an empty table.”
- Evidence tier. How reliable is the whole estimate? The SIGNAL metric (football and baseball) profiles six axes — sample, identifiability, granularity, noise, availability, lineage — and its overall band is the minimum across axes, never an average and never an accuracy percentage.
Status labels are load-bearing
The words concept, scaffold, research, provisional, and validated are a claim ladder, not decoration. SIGNAL never softens a label it authors: failed never becomes beta, blocked never becomes coming soon, and research never becomes current. Where a sibling's own enum appears — a raw coming_soon, a snake_case state — it is printed in quotes as the sibling's word, not rewritten. The test suites of the producing repositories — and of this site — enforce it.