Your hallucination checker only sees the final paragraph.
That’s the bleed. A fluent wrong number often starts earlier: empty retrieval, a swallowed tool error, a stale cache after a fact update. Score the prose alone and you miss the cause — then the model fills the gap confidently.
Most “anti-hallucination” tools optimize one job: grade the answer text (encoder / HHEM-class model / LLM-as-judge). Useful. Incomplete for agents.
What they usually can’t do:
0.87
After the model speaks (and optionally before it does), run a verdict:
ShineVerdict
: decision
-
named
resolution_gate -
evidence_hash
That product is PrismShine (Apache-2.0, pip install
).
It is not a prompt-injection firewall (that’s PrismGuard). It is not an agent runtime (that’s ChorusGraph). One job: verify answers against evidence.
Public comparative vs Vectara HHEM-2.1-Open on HaluEval
(Azure ACI · ONNX Tier-3 · receipt 2026-07-20_run4_onnx
):
| System | B1 QA F1 | B2 numbers F1 | B1 p50 | LLM |
|---|---|---|---|---|
| prismshine-fast | ||||
| 0.831 | ||||
| 1.000 (0 FP) | ||||
| ~90 ms | ||||
| 0 | ||||
| hhem-2.1-open | 0.746 | 0.926 | ~216 ms | 0 |
Receipt folder on GitHub: benchmarks/progress/2026-07-20_run4_onnx
(link in first comment).
bash
pip install "prismshine==0.2.2"
prismshine verify --demo
prismshine capabilities
Core path = Tiers 0–2, CPU, 0 LLM calls by default.
from prismshine import EvidenceBundle, PreloadChunk, ShineGate
gate = ShineGate.build(profile="default")
bundle = EvidenceBundle(
run_id="demo",
question="What was revenue?",
answer="Revenue was $1000 in Q1.",
preload=[
PreloadChunk(
chunk_id="c1",
text="Revenue was $1000 in Q1.",
source="retrieval",
)
],
)
verdict = gate.verify(bundle)
print(verdict.decision, verdict.resolution_gate, verdict.evidence_hash)
0.2.2 drop-in helpers: validate_grounding · get_gate · enforce_mode_from_env
Shadow without blocking: PRISMSHINE_ENFORCE=0
Docs: INTEGRATION.md §0 (link in first comment).
There’s also a no-API-key browser demo (link in first comment) — pass → fabricated number block → empty-retrieval halt.
Honest limits (BIP)
PASS ≠ world-true — grounded in your preload only
Buffered answers (not mid-stream token verification)
Bare pip install prismshine ≠ Tier-3 span SotA; use prismshine[spans] + ONNX when you need that path
Wired runtime moat is a separate Docker receipt — don’t mix it unlabeled with the HHEM table above
Soft ask
If you try verify --demo and hit a snag, paste the traceback in the comments — I’ll help.
Where would you wire the gate first — after the LLM node, before generation (halt empty retrieval), or both?
### First comment (after publish)
Links (kept out of the lesson above on purpose):
Landing → https://www.insightits.com/products/prismshine.html Demo → https://insightitsgit.github.io/PrismShine/demo.html GitHub → https://github.com/insightitsGit/PrismShine PyPI 0.2.2 → https://pypi.org/project/prismshine/0.2.2/ Receipt → https://github.com/insightitsGit/PrismShine/tree/main/benchmarks/progress/2026-07-20_run4_onnx YouTube → https://www.youtube.com/watch?v=OZCelVhP844 Product Hunt → (paste live PH URL)
Smoke: pip install "prismshine==0.2.2" && prismshine verify --demo