Solutions by boundary
MITHRIL · SOLUTIONS
The same one-language boundary, re-stated for four kinds of organisation. These are target patterns for the language, not customer stories: Mithril is new, and no deployed case is claimed.
Government
Sovereign and public-sector compute
- Problem
- AI-assisted software must pass through procurement and audit chains that assume a human wrote every line, while the systems themselves hold data a state cannot afford to expose.
- Approach
- The host grants only capabilities admitted as inspectable data before execution, and every artifact resolves to a canonical graph digest. The toolchain runs without a JVM and asks for no ambient network: one pinned, signed release installs on an air-gapped site, and the conformance suite - not a vendor claim - says what the compiler admits.
- Proof
-
- NIST SSDF (SP 800-218) asks for provenance and release integrity: content-addressed artifacts and admission receipts are that evidence in machine-readable form
- Air-gapped operation stated and tested by the language itself: the toolchain is JVM-free and artifacts are content-addressed
- Executable proof on this site: a .mith application and its web library, compiled and served from the public compiler
Finance
Fintech, banking and markets
- Problem
- Model output reaches customers through release cycles that answer to supervisors, and model risk management (SR 11-7 and its successors) expects documentation and effective challenge - not a vendor's reassurance.
- Approach
- Every artifact carries its admission receipt: what was compiled, under which policy, with which capabilities. Compliance evidence becomes a query over data the build already produced - model inventory and challenge material generated with the artifact, not reassembled after an incident.
- Proof
-
- Kotobase: durable, content-addressed storage for artifacts and receipts - the evidence chain outlives the release
- Per-artifact compliance exports (SBOM, admission receipts) are the Mithril enterprise distribution's stated scope
- The receipts this site's own graph agent stores by CID are the same mechanism, public
Defense
Defense and dual-use programs
- Problem
- Edge and classified programs run software that must not phone home, under supply-chain rules that quietly assume connectivity - and generated code inherits whatever the host can reach.
- Approach
- The execution boundary is fixed at compile time: a program that cannot hold ambient authority needs no perimeter to keep it honest. Offline operation is a property of the toolchain (JVM-free, content-addressed), and provenance is cryptographic - each binary resolves to a reproducible source and toolchain identity.
- Proof
-
- Deny-by-default admission: OaK revalidates ontology, transaction, Execution IR, capability and implementation identity before any effect
- The air-gap FAQ on this site: nothing phones home; a pinned toolchain and a private registry are all an offline site needs
- Amu: the JVM-free compiler/runtime beneath the language, built for verified compilation
Cybersecurity
Cybersecurity and critical infrastructure
- Problem
- Generated code inherits the network, filesystem and secrets of whatever host runs it - the assumption every breach replays - while defenders must reason over vulnerability data that arrives faster than audits.
- Approach
- Deny by default is the admission rule, not a hardening add-on: effects are typed and checked before they run. The post-quantum floor (hybrid X25519 + ML-KEM-768, ML-DSA-65) ships as machine-readable coverage, so KEV and CVE feeds become the start of an evidence query, not the boundary itself.
- Proof
-
- Published post-quantum coverage data at kotoba-lang.org (machine-readable, as deployed)
- OaK kernel: fail-closed admission before any effect - public source and conformance suites
- KEV and CVE datasets published at knowledge.kotoba.cloud - the same evidence plane the language's own admission consumes