A common language for AI risk assessment output
SREL is a normative context-free grammar for structured, machine-readable AI risk assessment documents. Vendor-neutral. Geometrically grounded. Built for downstream consumption by model risk management, monitoring, and remediation systems.
Structured risk output that machines can read and humans can audit
AI risk assessments today are PDFs, spreadsheets, and PowerPoint decks. They cannot be ingested by monitoring systems, compared across vendors, or automatically escalated when thresholds are crossed. SREL is the interchange format that closes this gap: a context-free grammar specifying required sections, required keys, value types, and semantic invariants for AI risk assessment documents.
Normative grammar
Defined in ABNF per RFC 5234. Any document either conforms or does not. Conformance is machine-checkable.
TOML wire format
Preferred lexical realization is TOML 1.0.0. Alternate lexical embodiments (JSON, YAML, XML) satisfying the same grammar are also conformant.
URN-identified
Documents declare their profile by URN per RFC 8141. No DNS dependency. urn:srel:profile:ai:v1 is the AI Risk Profile.
Geometric substrate
Scores are derived from n-simplex risk decomposition. Nine dyadic groups × six compositional periods = 54 named risk primitives, each at a structurally located matrix position.
Meta-condition veto
Six binary structural prerequisites (KILL, SAF, HITL, AUT, TRU, MAN) gate the composite. Breach of any one collapses the score to zero regardless of all other measurements.
Versioned profiles
The meta-language hosts multiple domain profiles. AI risk is the first. Cybersecurity, climate, model risk (SR 11-7), biological risk, and infrastructure resilience profiles are planned.
Eight required section classes in the AI Risk Profile
[capability][meta][primitive."g.p"]g.p (group I–IX × period P1–P6 = 54 positions). Each carries code token, score, shadow distance, lambda, and weight.[register_weights.*][[compound]][composite][explanation]headerOne meta-language. Multiple risk domains.
Each domain profile inherits the meta-language grammar and adds domain-specific required sections. Profile identifiers follow the pattern urn:srel:profile:<domain>:v<N>.
Open specification. Protected methodology.
Patent License Commitment
The inventor commits, conditional upon issuance of any patent on the applications disclosing the SREL specification, to grant a royalty-free, world-wide, non-exclusive, irrevocable license under any claim of the issued patent that would necessarily be infringed by implementing the SREL specification as published — solely to the extent necessary to implement, validate, transport, or consume SREL-conformant documents.
This commitment covers the format. It does not extend to claims directed to the geometric risk-scoring methodology from which the format derives. The scope is the interchange standard, not the engine.