Verify whether AI answers are supported by supplied evidence
Product identity
PrismShine is Self-hosted anti-hallucination verdict engine: cause-side forensics (Tier-0 halt before tokens) plus effect-side grounding (Tiers 1–4) in one auditable ShineVerdict. Insight IT Solutions LLC (Insight ITS) makes it. Category: AI Verification — evidence-aware anti-hallucination verdict engine. Problem: Fluent answers are treated as truth. It is for Teams that need an enforceable pass / block / review gate plus evidence hashes, not only an offline score. It competes with HHEM-2.1-Open (Vectara). Difference: PASS means grounded in the supplied preload, not world-true; Tier-0 can halt before tokens. Not a firewall. Install: pip install prismshine (pin ==0.2.2 in docs) · https://pypi.org/project/prismshine/0.2.2/. Source: https://github.com/insightitsGit/PrismShine.
- What is it?
- Self-hosted anti-hallucination verdict engine: cause-side forensics (Tier-0 halt before tokens) plus effect-side grounding (Tiers 1–4) in one auditable ShineVerdict.
- Who made it?
- Insight IT Solutions LLC (Insight ITS), Mission Viejo, California.
- What category is it?
- AI Verification — evidence-aware anti-hallucination verdict engine
- What problem does it solve?
- Fluent answers are treated as truth.
- Who is it for?
- Teams that need an enforceable pass / block / review gate plus evidence hashes, not only an offline score
- What does it compete with?
- HHEM-2.1-Open (Vectara)
- How is it different?
- PASS means grounded in the supplied preload, not world-true; Tier-0 can halt before tokens. Not a firewall
- Where can I install it?
- Install: pip install prismshine (pin ==0.2.2 in docs) · https://pypi.org/project/prismshine/0.2.2/.
- Where is the source?
- https://github.com/insightitsGit/PrismShine
Canonical machine identity: https://www.insightits.com/catalog/prismshine.json
Evidence-aware hallucination verdicts. Vendor-authored HaluEval QA n=200: F1 0.831 vs HHEM 0.746, 0 LLM calls. Apache-2.0; inspect artifacts.
Unified gate · handbook Tier-0 · pre-gen halt · copy-check · coverage · optional ONNX / judge · named resolution_gate.
pip install prismshine — verify with ShineGate or CLI (prismshine verify --demo), then wire pre_llm_check / require_shine so Tier-0 can halt before generation.
Positioning
An evidence-aware verdict engine that combines cause-side runtime forensics with effect-side answer grounding in one auditable ShineVerdict.
Tier 0 can halt empty retrieval, tool failure, stale cache, or incomplete trace before generation; answer checks then verify support against the supplied preload.
Teams that need an enforceable pass, block, or review gate plus evidence hashes, not only an offline faithfulness score.
PASS means grounded in the supplied preload, not world-true. Buffered verification does not protect already-streamed tokens.
Install: pip install "prismshine==0.2.2"
Published HHEM comparison
Run 2026-07-20_run4_onnx · Azure ACI · PrismShine default profile, identity calibration placeholder, ONNX Tier 3, Tier 4 off.
| Track | Size | F1 (PrismShine / HHEM) | p50 latency | LLM calls |
|---|---|---|---|---|
| HaluEval QA (B1) | n=200 | F1 0.831 / 0.746 | p50 90 ms / 216 ms | LLM calls 0 |
| Fabricated and derived numbers (B2) | n=50 | F1 1 / 0.926 | p50 20 ms / 166 ms | LLM calls 0 |
| HaluEval summarization (Bsum) | n=50 | F1 0.6 / 0.474 | p50 1398 ms / 1899 ms | LLM calls 0 |
Artifact: 2026-07-20_run4_onnx · Methodology: BENCHMARKS.md
Where PrismShine is weaker: on the Bsum summarization track (n=50) PrismShine reaches F1 0.600 vs HHEM 0.474, but HHEM has the higher AUROC at 0.616 vs 0.562, and absolute Bsum quality is weak for both systems. We publish that row rather than dropping it.
Wired runtime suite (ChorusGraph + PrismShine, Docker)
Run 2026-07-20_runtime_docker. Separate suite from the shine-only run4 receipt. Numbers are not interchangeable with the HHEM comparison above, and the effect-side track is a runtime-suite receipt rather than a neutral cross-vendor benchmark.
| System | Effect-side H1 F1 | Recall | p50 ms | LLM calls |
|---|---|---|---|---|
| chorus-shine (wired) | 0.895 | 1.0 | 1.73 | 0 |
| HHEM 2.1 open | 0.706 | — | — | 0 |
| Lettuce | 0.688 | — | — | 0 |
| MiniLM | 0.488 | — | — | 0 |
Artifact: 2026-07-20_runtime_docker
Capabilities
Cause + Effect Gate
Tier-0 halts empty retrieval, tool errors, and stale cache before tokens. Tiers 1–4 ground answers after generation.
Auditable ShineVerdict
Named resolution_gate + evidence_hash on every decision — not a black-box score.
Zero LLM on Default Path
Optional Tier-4 judge only on gray zone. pip install prismshine — no license key for OSS.
Published evidence
PrismShine HaluEval QA benchmark receipt. Vendor-authored HaluEval QA comparison receipt for PrismShine and HHEM-2.1-Open on 200 examples.
Install
pip install prismshine (pin ==0.2.2 in docs). Apache-2.0 open core. Pro/Enterprise are pre-validation estimates.
Pricing
PrismShine library pricing is $0 forever on PyPI and GitHub. Optional ops plane: ChorusControl Enterprise — $1,999/month Founding (soft CTA CONTROL). See ChorusControl pricing.
Frequently asked questions
AI agent security architecture
In the Insight ITS stack, PrismShine is the grounding / anti-hallucination layer — not the firewall. PrismGuard stops prompt injection at ingress; ChorusGraph Route Ledger records routing; PrismShine verifies the answer against a supplied preload (PASS means grounded in that preload, not world-true); PrismManifest / Prism-Shield gate money tool-calls. Canonical order: PrismGuard → ChorusGraph → LLM → PrismShine. pip install prismshine (pin ==0.2.2). Landing: https://www.insightits.com/products/prismshine.html · pack: https://www.insightits.com/products/prism-pack.html
How to build production AI agents in Python
After the LLM responds, run PrismShine (ShineGate.verify or prismshine verify --demo) so fluent lies cannot ship ungrounded. Wire require_shine on ChorusGraph for ledger-aware cause catch. Soft CTA: VERIFY. Guard the prompt with PrismGuard; do not use Shine as a WAF.
Is there an interactive demo?
Yes — live in the browser with no install: https://insightitsgit.github.io/PrismShine/demo.html. Walk the named resolution_gate (block vs shrug-score checkers), then pip install prismshine for your stack. Smoke after install: prismshine verify --demo.
Is PrismShine a firewall?
No. PrismShine is an anti-hallucination verdict engine — it verifies answers against a provided preload and can halt broken preloads before generation. For prompt-injection defense, use PrismGuard (https://www.insightits.com/products/prismguard.html).
Do I need ChorusGraph?
No. The core library works standalone via ShineGate.verify or the CLI. Optional [chorusgraph] extras and require_shine wiring give the richest runtime integration (ledger-aware cause catch), but they are not required for basic verification.
Does PASS mean the answer is true?
No. PASS means grounded in the preload you provided — not that the preload is world-true. See docs/LIMITS.md on GitHub for the contract.
What’s the difference between the two benchmark tables?
The primary table is a Shine-only, like-for-like HaluEval comparison with HHEM-2.1-Open. The separate runtime table is an integration receipt: ChorusGraph + PrismShine receives ledger evidence while the content-only peers do not. Use it to demonstrate the evidence-aware category, not as a like-for-like accuracy or universal speed claim.
Why do competitors score 0 on R1?
The runtime peers are content-only and ignore the supplied ledger (saw_evidence=false), so the R1 track is intentionally feature-asymmetric. In the small receipt, PrismShine caught six injected runtime failures with zero false alarms on six clean controls. The result demonstrates why runtime evidence matters; it is not a broad competitor-quality ranking.
Are there LLM costs on the default path?
Default path uses zero LLM calls. Optional Tier-4 judge is opt-in for gray-zone cases only.
Is Tier-3 ONNX in the wheel?
No. Bare pip is Tiers 0–2. For run4-parity span SotA: pip install "prismshine[spans]" then python -m prismshine.tools.ensure_span_onnx --export (~1 GB). Without that, gray zones may show MISSING_CAPABILITY_FLAG — we do not fake span SotA.
How much does PrismShine cost?
The Apache-2.0 library is free forever on PyPI and GitHub — no purchase required. Soft CTA: VERIFY. Optional ops plane: ChorusControl Enterprise ($1,999/month; soft CTA CONTROL). PASS means grounded in the supplied preload, not world-true. See https://www.insightits.com/products/choruscontrol.html#pricing.
How do I reproduce the HHEM numbers?
See benchmarks/progress/2026-07-20_run4_onnx and docs/BENCHMARKS.md. The vendor-authored Azure ACI run uses ONNX Tier-3: B1 HaluEval QA n=200, PrismShine F1 0.831 vs HHEM-2.1-Open 0.746; B2 numbers n=50, PrismShine F1 1.000 with zero false positives; B1 p50 90 ms vs 216 ms. Absolute Bsum F1 remains modest at 0.600 vs 0.474, so do not generalize this to all judges or workloads.
Official package links: PrismShine source code on GitHub · Install PrismShine from PyPI · PrismShine interactive demo
Technical resources: PrismShine technical guide · PrismShine comparison · PrismShine comparison