Honesty — experimental & stub surfaces

aire documents incomplete areas instead of silent magic. Prefer these APIs knowing their limits.

Surface	Behavior
Foundation toy	AI.training.foundation / aire foundation builds random-init / config-driven stacks. describe() → kind: foundation_toy_architecture. Not pretrained GPT/LLaMA weights.
TTS echo	EchoTTSBackend returns data:text/plain,... placeholders with stub=True. Real TTS needs a model with Capability.TEXT_TO_SPEECH.
Video stub	VideoPipeline.summarize without a vision/video-capable model returns offline stub summaries (stub=True).
Lexical metrics	semantic_overlap is token-set F1 (prefer token_overlap). bleu is sentence BLEU-4+BP (pure Python); use sacrebleu / aire-ai[eval] for sacreBLEU. groundedness / faithfulness stay lexical; prefer nli_faithfulness with a judge.
embedding_similarity	Requires ctx.embedder; raises ConfigurationError if missing (no silent lexical fallback).
nli_faithfulness	Requires ctx.judge; raises if missing.
cross_encoder rerank	hf_cross_encoder / cross_encoder need sentence-transformers (aire-ai[eval]). Pass model= (LLM) to keep the old prompt scorer via model reranker.
Semantic chunker	Without embedder=, sentence-boundary approximation + warning; describe() reports mode: sentence_approximation.
MCP subset	Stdio + streamable HTTP tools subset — not full MCP (sampling/roots/progress still open).
Workers	in_process, file, redis (aire-ai[redis]); SQS raises until boto3 is bundled.
Gateway images	Returns 501 when IMAGE_GENERATION capability is absent.
CLI analytics	Demo counters unless live metrics are attached.
web_search	DuckDuckGo HTML scrape, not an official search API.
Guardrails auto-wire	Knowledge.ask and create_gateway apply SafetyConfig regex rails by default (guardrails=False to disable). Model classifiers are opt-in via AI.safety.guardrails("model_injection", model=...).

Also see root GAPS.md for the prioritized backlog.