Skip to content

Launch Gates

Release gate checklist for launch readiness, and how to certify a BYOK model.

Use these gates before promoting release channels.

Validate The Harness For Free

Before spending any model budget, confirm the gate pipeline itself is sound:

Terminal window
bun run test:agentic:dryrun

This runs the full agentic task matrix against a canned no-model stub instead of a real provider — it validates case iteration, output parsing, file/regex checks, and pass-rate computation end to end with zero API calls or spend. A clean bun run test:agentic:dryrun means any failure in a real run below is genuine model/runtime signal, not a harness problem.

Certifying A BYOK Model

AlphaBase is BYOK — different users will connect different models, from the free testing tier up to frontier paid models. Use this to measure whether a specific model is reliable enough to be your default before committing to it:

Terminal window
MODEL=<provider/model> \
MODEL_LABEL=<short-name> \
REQUIRED_PASS_RATE=85 \
bun run certify:model
  • MODEL — the provider/model under test (defaults to the free testing model, ollama/minimax-m2.5:cloud, if unset).
  • REQUIRED_PASS_RATE — percent of graded cases the model must pass to certify (default 100, i.e. strict). Set a lower threshold, e.g. 85, to certify a capable model that may reasonably differ on one edge case rather than requiring perfection.
  • Provider-quota/billing blocks are excluded from the pass rate — they’re an environment condition, not a model capability signal.

A passing certification at your chosen threshold is the evidence that a given model is ready to be recommended as a default, rather than assuming it from the model’s reputation alone.

Required Commands

Terminal window
npm run typecheck
npm run test:smoke:launch
npm run test:agentic:gate
npm run test:tenant:isolation
npm run test:launch:gate

P0 Runtime Gates

  • no silent task drop
  • no post-success continuation loop
  • no internal lifecycle artifacts in user-visible response
  • explicit terminal completion/failure handover

These are enforced by the task-recovery and loop-limit mechanisms described in Task Execution — a task that cannot complete must stop with a clear message, not loop or fail silently.

Clean Install Gate (VS Code)

  • install latest VSIX in clean profile
  • verify diagnostics/sophia/chat runtime from clean state
  • archive prompt + screenshot evidence

Rollback Gate

  • previous known-good version remains installable
  • rollback path is tested at least once per channel promotion

Multi-Tenant Safety Gate

  • tenant isolation checks pass
  • access controls and audit logs are active