XLOG has several validation layers. They answer different questions and should not be collapsed into one test count.
Do not use stale fixed pass-count snapshots as public certification statements. The authoritative gate is the command that ran, the hardware it ran on, and the evidence it produced.

​
Validation Layers

LayerWhere it runsWhat it proves
GitHub CIGitHub-hosted runnersFormatting, workflow hygiene, package metadata, no-GPU CUDA build, and non-GPU checks.
Docs site CIGitHub Actions for docs-site/**Mintlify validation, broken-link checks, and static export publication.
GPU release validationMaintainer-run CUDA hostActual CUDA behavior through scripts/validate_release_gpu.sh.
Subsystem reliability gatesSubsystem-specific suitesStatistical or staged reliability for neural-symbolic and other higher-level engines.
Green GitHub CI does not certify GPU correctness. GPU certification requires a CUDA machine.

​
GPU Release Gate

The canonical manual gate is: 
scripts/validate_release_gpu.sh --mode release
The script:
  • sets XLOG_REQUIRE_CUDA=1, so CUDA initialization failures cannot be skipped;
  • requires a visible NVIDIA GPU through nvidia-smi;
  • runs release doctor checks;
  • builds the workspace and xlog-cli release binary;
  • stages Python and CLI kernel artifacts;
  • builds the pyxlog wheel and CLI archive;
  • runs xlog-cuda-tests certification in release mode;
  • runs a basic xlog run smoke command;
  • verifies that packaged artifacts include the expected kernel files.
Use --mode smoke for a shorter CUDA smoke gate. Use --dry-run only to inspect the command sequence; it does not certify GPU behavior.

​
Docs Gate

The docs workflow triggers on edits under docs-site/** or the workflow itself. It uses Node 22, installs [email protected], runs: 
mint validate
mint broken-links
mint export
On main, the exported static bundle is pushed to the docs-dist branch. The DigitalOcean App Platform site serves that branch at xlog.md.

​
Reliability Gates

Reliability gates are not the same as CUDA kernel certification. The real staged reliability labels used in the repository are:
  • alpha: 5/5;
  • beta: 20/20, defined as 5 seeds across 4 stages;
  • GA: 50/50 with Clopper-Pearson confidence accounting.
Use those labels only for the subsystem that defines and runs that gate. Do not reuse them as global CUDA test counts.

​
Epistemic Candidate Bounds

Epistemic execution does not use a public fixed-literal or large hardcoded candidate-count bound. The source computes max_candidates = 2^(number of epistemic literals) and caps models per reduction at MAX_MODELS_PER_REDUCTION = 1024. Document that concrete bound when discussing epistemic planning limits.

​
Evidence Requirements

A certification claim should include:
  • exact command;
  • commit or release tag;
  • CUDA toolkit and GPU class;
  • whether XLOG_REQUIRE_CUDA=1 was active;
  • route counters or transfer telemetry when the claim depends on a specific optimized path;
  • artifact or log location when the evidence is durable.
If those details are missing, phrase the result as a local check, not a release certification.