Stability matrix

Use this page to decide which capabilities are safe for production defaults and which should stay behind tighter rollout controls.

Feature maturity

Area	Scope	Status	Notes
LLM provider layer	OpenAI, Azure OpenAI, Gemini, Anthropic, AWS Bedrock, LM Studio, Groq, Ollama	Stable	Primary generation path for `run`.
Plan-first + lockfile	`run --dry-run`, deterministic planning, `.hyperlocalise.lock.json` checkpoints	Stable	Recommended default for CI and PR review flows.
TMS adapters	Crowdin, LILT AI, Lokalise, Phrase, POEditor, Smartling	Experimental	`sync push` / `sync pull` contracts may evolve. Push now uses fresh remote baselines and entry-level mutation where supported, but adapter concurrency guarantees still vary. Pin CLI versions in CI.
Eval framework	`eval run`, `eval compare`, quality/regression scoring	Experimental	Report schema and scoring can change between releases.
CI-ready status labels	`ready`, `needs review`, `missing` derived from `status` CSV	Experimental	Label mapping is stable for current docs, but classification semantics may evolve.

Treat run + lockfile as baseline in default CI checks.
Use sync and eval in controlled pipelines with explicit version pinning.
Gate releases with status-based checks before enabling write-capable sync jobs.