[ Public algorithm ]

Every Score can be checked.

Score v2 uses only observed signals. Fabricated seed and intake metrics were purged, so Scores are lower and flatter than v1. A factor without evidence reads “insufficient evidence”, not zero.

Weights

FactorWeight

Score factor

github health

18% of Score v2 when the factor is observed. Unobserved factors are omitted.

18%

Score factor

agent use

16% of Score v2 when the factor is observed. Unobserved factors are omitted.

16%

Score factor

human notes

14% of Score v2 when the factor is observed. Unobserved factors are omitted.

14%

Score factor

composability

16% of Score v2 when the factor is observed. Unobserved factors are omitted.

16%

Score factor

checklist

10% of Score v2 when the factor is observed. Unobserved factors are omitted.

10%

Score factor

evidence quality

14% of Score v2 when the factor is observed. Unobserved factors are omitted.

14%

Score factor

benchmark

12% of Score v2 when the factor is observed. Unobserved factors are omitted.

12%

100 × sum(weight × value) over observed factors only. Unobserved factors are omitted, not zeroed.

Worked example: Model Context Protocol

FactorCalculationPoints
GitHub health18% × 0.305.4
Agent useinsufficient evidenceinsufficient evidence
Human notes14% × 0.091.2
Composabilityinsufficient evidenceinsufficient evidence
Checklist10% × 1.0010.0
Evidence quality14% × 0.466.5
Benchmarksinsufficient evidenceinsufficient evidence
Rounded Score23
Open full explanation

[ Rules ]

  • Score v2 is evidence-only. A factor without an observed signal is insufficient evidence, not a zero.
  • Purging fabricated seed and intake metrics lowers and flattens Scores. That is intended.
  • Humans and agents are first-class. Neither base can invent the other.
  • Composability enters Score only when an observed signal exists. A rationale sentence is not a number.
  • Checklist starters are MCP, API, and cost range. Those are declared claims on the record, not usage metrics.
  • Benches enter only when Leyline or an external bench is observed with source, method, and observedAt.
  • Intake must not write stars, installs, success, abandons, resolve wins, trust, or benches unless they were observed.