prism-core (pip install prism-honesty)

A fail-closed honesty gate for LLM answers grounded in user-provided sources.

The audit checks a draft against the source material (plus the user's own statements) and flags only unsupported facts — invented credentials, metrics, dates, scope — while leaving motivation, tone, framing, and hedged language alone. Deterministic backstops keep it honest in both directions:

• fail-closed — if the auditor can't produce a usable verdict (truncated, malformed, transport error), the draft is not certified. No silent pass.
• verbatim backstop — a flagged claim whose exact phrase appears contiguously in the source is rescued (auditor false positive); a same-meaning inflation is never rescued.
• no truncation games — an over-long source is fitted by relevance to the draft, not head-cut (head-cutting was the bug where critical content slid past the cut).

Standard library only. No provider, framework, or chat-UI dependency: the model call is injected by you.

Usage

import asyncio
from prism_core.audit import fact_audit

async def my_complete(messages, model, *, max_tokens, temperature,
                      reasoning_effort, session, label, return_finish):
    # call whatever provider you like; return (raw_text, finish_reason)
    ...

result = asyncio.run(fact_audit(
    full_request="Write my resume summary. I mentored two interns.",
    source="Mentored two interns on an internal tool.",
    candidate="Senior engineer who led a team of 50 engineers.",
    complete_fn=my_complete,
    model="your-model-id",
    max_tokens=2000,
    reasoning_effort="max",
))
# -> {"unsupported": ["led a team of 50 engineers"], "verdict": "FABRICATION"}
# or {"verdict": "ERROR", ...} when no usable verdict arrives (fail closed)

Contents

License

MIT (this directory). The parent project, PrismAI, is AGPL-3.0 — only prism_core/ is MIT.

Status

This is what I use in production, extracted so my own code has a real boundary. Published without promises: no roadmap, PRs may not get merged.