Director-Lite — streaming halt in 3 lines¶

director-ai-lite is a standalone, dependency-free package: it stops an LLM token stream before a hallucination finishes generating, with zero heavy dependencies and no director-ai requirement.

pip install director-ai-lite

from director_ai_lite import guard

result = guard(
    token_stream,
    facts={"capital": "Paris is the capital of France."},
    prompt="What is the capital of France?",
)
print(result.output)       # surviving text (halted tokens removed)
print(result.halted)       # True if the stream was stopped
print(result.halt_reason)  # why it was stopped

token_stream is any iterable of string tokens — wire it straight to your LLM's streaming response.

Priority

Director-AI publicly shipped streaming contradiction-halt surfaces in early 2026 and deposited the related artefact on Zenodo (March 2026). Treat that as provenance for the mechanism, not as a standalone production accuracy claim.

How it works¶

The default path is model-free. Each accumulated prefix is scored by a grounding heuristic (content-word overlap against the supplied facts) and the same calibrated coherence combination the full package uses in its no-model path. The stream hard-halts on the first token whose coherence drops below threshold (default 0.5). With no facts, scoring stays neutral and nothing is halted.

This is great for a first look and runs anywhere, but it is approximate — the heuristic has no model behind it.

One call, or a reusable guard¶

from director_ai_lite import StreamGuard, streaming_guard

# one-shot
result = streaming_guard(token_stream, facts={...}, prompt="...")

# reusable
g = StreamGuard(facts={...}, threshold=0.6)
result = g.guard(token_stream, prompt="...")
text = g.safe_text(token_stream, prompt="...")  # surviving text only

Parameters¶


`facts`	mapping of key → grounded statement for factual scoring
`threshold`	coherence floor in `[0, 1]`; the stream halts below it (default `0.5`)
`scorer`	optional `review(prompt, text)` scorer (e.g. model-backed NLI) that overrides the heuristic

Upgrade to model-backed scoring¶

The grounding heuristic and coherence calibration match the full package's no-model path, so upgrading does not change the call site — install the full package and pass its scorer:

pip install "director-ai-lite[full]"

from director_ai_lite import StreamGuard

g = StreamGuard(facts={...}, scorer=my_nli_scorer)  # any review(prompt, text) scorer

Tiers¶

Director-Lite is the free, standalone entry point. The wider product ladder:

Tier	What it is
Director-Lite	Free. This package — standalone, model-free streaming halt, zero dependencies.
Director-AI	The full runtime — model-backed NLI/RAG scoring, REST/gRPC server, framework integrations, sealed evidence packets, tamper-evident audit.
Director-AI Pro	Production-tier licence and support on top of the full runtime.
Director-AI Full	The complete advanced + labs capability set.
Director-Class AI	Separate action-control and evidence product for high-impact agent operations; scoped commercial order.

License: Apache-2.0.