# vyre-conform — Internal Spec
> This file is gitignored. It exists for agents and internal development. Never committed to public repos.
## Identity
Conformance suite and specification definition for the vyre GPU compute bytecode standard.
## Purpose
Without vyre-conform, there is no objective measure of vyre correctness. A runtime refactor or alternative backend could silently break bytecode semantics. This crate defines the spec independently so that vyre (the implementation) is tested *against* it, not the other way around.
## North Star
Become the CTS (Conformance Test Suite) that any third-party vyre implementation must pass. Legendary looks like Khronos-level conformance: exhaustive opcode coverage, property-based parity tests, canary regressions for every past bug, and a published backend trait that lets anyone plug in their own GPU runtime.
## Role in Ecosystem
- **Depends on:** nothing (zero dependencies by design)
- **Depended on by:** vyre (dev-dependency only)
- **Relationship to warpscan:** none direct; indirectly ensures the GPU condition layer that warpscan uses is correct.
- **Standalone value:** YES — anyone implementing a vyre-compatible runtime can pull this crate in as their test harness.
## Invariants
- Must never depend on vyre (the implementation). Any spec leakage into the implementation crate is a bug.
- All golden outputs are deterministic across platforms.
- A backend passes conformance only if it matches the reference CPU semantics on every test vector.
## Boundaries
- Does not implement GPU dispatch — it only defines the `VyreBackend` trait and test vectors.
- Does not generate WGSL shaders — that's vyre's job.
- Does not handle file I/O or scanning orchestration — that's warpscan/yaragpu.
## Quality State
- Tests: 0 explicit targets, 3 inline tests, 0 test files (the crate itself IS the test framework; test counts are measured by consumers)
- Lint preamble: yes (`#![warn(clippy::pedantic)]`, `#![forbid(unsafe_code)]`, `#![warn(missing_docs)]`, unwrap deny block)
- `#![forbid(unsafe_code)]`: yes
- Doc coverage: ~80% (backend trait and pipeline modules are well documented)
- Known issues: still expanding opcode coverage; some advanced quantifier cases need more adversarial vectors.