Skip to content

Smoke and Release Gates

Truthound AI uses explicit gates before new functionality is treated as release ready.

Release Discipline

The core release boundary is:

  • proposal compilation stays reviewable
  • run analysis stays read-only
  • approval/apply remains explicit
  • provider integration is exercised through opt-in smoke paths
  • core hot-path contracts continue to pass without AI extras installed

Smoke Coverage

Two live OpenAI smoke paths exist as opt-in checks:

  • proposal smoke for suggest_suite(...)
  • run-analysis smoke for explain_run(...)

Both use temporary workspaces, explicit environment gates, and minimum viable contract assertions rather than brittle exact-output expectations.

Prompt hardening also has a deterministic acceptance gate that does not call a provider:

  • 200+ Korean golden prompts
  • 50+ mixed Korean/English prompts
  • 50+ ambiguous prompts
  • 90%+ ready-or-partial acceptance for golden and mixed prompts
  • 95%+ clarification behavior for ambiguous prompts
  • zero route/compiler crashes for unsupported or malformed fixtures

Live provider checks stay manual-only. Use the single-model smoke for a quick canary, or TRUTHOUND_AI_SMOKE_MODEL_MATRIX when comparing structured-output and JSON-mode behavior across models.

TRUTHOUND_AI_RUN_LIVE_SMOKE=1 \
TRUTHOUND_AI_SMOKE_MODEL=gpt-5.4-mini \
truthound ai smoke openai

TRUTHOUND_AI_RUN_LIVE_SMOKE=1 \
TRUTHOUND_AI_SMOKE_MODEL_MATRIX='[{"model":"gpt-5.4-mini","expected_format":"json_schema"}]' \
truthound ai smoke openai-prompt-canary --json

Why Manual Gates Matter

AI functionality is more temporally unstable than the kernel. Provider SDKs, models, and structured-output behavior change. The release process therefore leans on:

  • contract tests
  • doctor validation
  • deterministic prompt acceptance tests
  • live smoke runners
  • explicit dependency probes

This is how the additive AI surface stays evolvable without turning the kernel or docs portal into a moving target.