Skip to content

Conformance

Vexor is regularly measured against the ecosystem's own conformance tooling — not just its own test suite. This page mirrors the "Conformance" section of the main repository README, which is the source of truth; it exists so the current numbers are visible from the docs site itself.

Instruction-execution corpus (official harness)

Vexor is measured against firedancer-io/solana-conformance using the firedancer-io/test-vectors instruction-fixture corpus, executed live against a version-matched Agave reference — not against baked-in expected outputs. Latest full-corpus run (47,240 fixtures):

  • 85.82% raw (40,542 / 47,240) and 92.74% under the harness's consensus-compatibility mode (-c, which normalizes error-code encoding).
  • Several families pass 100% raw: system (7,400/7,400), compute-budget (2,627/2,627), precompile (19,292/19,292), zk_sdk (3,124/3,124).
  • The raw gap is dominated by disclosed known gaps: the BPF-loader-owned ELF-loading fixture families (~2,900 fixtures) are not yet implemented to byte-match, and most of the remaining vote/vm-programs raw failures are error-code/CU-encoding differences that the harness's consensus mode recovers.

Syscall / crypto corpus (self-hosted harness)

A second, separate fixture corpus (7,571 fixtures total) exercises Vexor's syscall layer directly — including the crypto syscalls (BN254/alt_bn128, BLS12-381, secp256k1_recover, Poseidon, curve25519) — through a self-hosted runner built against the same execution seam. Latest run: 99.62% non-known-gap (6,476/6,501 runnable fixtures pass; 7,571 total including ~1,070 disclosed known-gap fixtures excluded from the denominator). bls12_381 and secp256k1 families are both fully passing (664/664, 190/190).

Reading these numbers honestly

Both corpora are self-measured — Vexor runs its own conformance harness and reports its own results; no external audit has reviewed them yet. "Known gap" fixtures (functionality not yet implemented, disclosed by name in each corpus's own results log) are excluded from the non-known-gap denominators above; they are never counted as passing. Where a family or corpus reads 100%, that means every fixture currently in scope for that family passes — not that the family has zero unimplemented functionality. The instruction-corpus numbers above (85.82% raw / 92.74% consensus) are the more conservative reading, since the official harness scores every fixture in the full 47,240-fixture corpus, known gaps included.

Independent of both conformance corpora, every deploy to the live validator replays 1,992 canonical slots spanning an epoch boundary offline and requires byte-identical bank hashes against the live cluster's before the binary ships. This is a correctness gate on the exact binary being deployed, not a fixture-corpus measurement — see Why Byte-Faithful.