Scoring methodology, short form
Analytical aid. Not a rating or rating recommendation. Full rules: docs/scoring/scoring_rules.md in the
repository (not rendered here; it contains case-reference wording the output vocabulary guard blocks).
Dimensions and tiers
Four dimensions per issuer: Business Risk, Financial Risk, the Liquidity gate, Governance evidence. Each scored
dimension gets an engine materiality tier: low, moderate, elevated or high (internal calibration). The overall
tier is the worst scored dimension; governance evidence can only raise it. A dimension without enough data is
not scored: it is reported as such, never treated as low or as a pass. Cyber incidents are scored per
scoring_rules.md section 4; AI infrastructure issuers per section 5 (internal calibration v0, not validated).
Ratios are as-reported XBRL proxies; S&P adjustments are not applied, which understates leverage for CoreWeave
and Applied Digital.
Numeric scale (scoring_rules.md 8.1)
Scores run 1 to 5: 1 negligible, 2 low, 3 moderate, 4 elevated, 5 severe. Engine tiers map low=2, moderate=3,
elevated=4, high=5. A dimension scores 1 only where it finds no exposure: governance with no evidence flags, or
financial risk with every computed ratio in the strongest band, none borderline, and no event shock. Governance
with one moderate flag scores 3; any flag from the top class, or two or more moderate flags, scores 4. An
unscored dimension is n/a, never 1. The scale is a presentation of the engine tiers and is never mapped to any
rating scale.
Composite (8.2)
Composite = arithmetic mean of the scored dimensions, one decimal, shown as "x.x / 5, n of 4 scored", labeled
internal calibration. Cyber issuers use the base scenario. The worst-of overall tier is unchanged and shown in
the breakdown.
Status (8.3)
Review needed: composite at or above 3.5, or at or past an S&P stated unfavorable condition on the
governing basis (D-052). Watch: any dimension score at or above 4, or distance to an S&P stated unfavorable
condition below 10%. Clean: otherwise. Distances compare the engine's as-reported proxy with S&P's stated level; where S&P's own
stated adjusted value is available it governs, and a breached proxy that S&P's stated figure contradicts is
reported as not comparable, never as a breach.