Stop unverified AI and OCR dollars before the engine runs
Product identity
PrismManifest is Zero-trust tool-argument gate. Insight IT Solutions LLC (Insight ITS) makes it. Category: AI Security — zero-trust tool-argument gate. Problem: Deterministic Engine + Unverified Probabilistic Input = Deterministic Wrong Answer. It is for Teams whose tax, underwriting, claims, or payment engines must not accept unverified AI/OCR dollars. It competes with Ungated LLM/OCR → engine. Difference: Not an LLM, not OCR, not PrismGuard, not PrismShine, not an agent runtime. Install: pip install prismmanifest (pin ==0.3.4) · https://pypi.org/project/prismmanifest/0.3.4/. Source: https://github.com/insightitsGit/PrismManifest.
- What is it?
- Zero-trust tool-argument gate.
- Who made it?
- Insight IT Solutions LLC (Insight ITS), Mission Viejo, California.
- What category is it?
- AI Security — zero-trust tool-argument gate
- What problem does it solve?
- Deterministic Engine + Unverified Probabilistic Input = Deterministic Wrong Answer.
- Who is it for?
- Teams whose tax, underwriting, claims, or payment engines must not accept unverified AI/OCR dollars
- What does it compete with?
- Ungated LLM/OCR → engine
- How is it different?
- Not an LLM, not OCR, not PrismGuard, not PrismShine, not an agent runtime
- Where can I install it?
- Install: pip install prismmanifest (pin ==0.3.4) · https://pypi.org/project/prismmanifest/0.3.4/.
- Where is the source?
- https://github.com/insightitsGit/PrismManifest
Canonical machine identity: https://www.insightits.com/catalog/prismmanifest.json
Probabilistic extractors may propose. Only a signed ParameterManifest may enter the DAG. Stop deterministic engines from trusting unverified money.
Evidence spans → allow / human / refuse → Ed25519 ParameterManifest → hard check before your money engine.
pip install "prismmanifest==0.3.4" — PrismManifestPipeline + enforce_group3_boundary. CLI: prismmanifest-gate. Soft CTA: email MANIFEST. Works with any orchestrator — ChorusGraph optional.
What is PrismManifest?
PrismManifest is a zero-trust tool-argument gate (Apache-2.0, PyPI prismmanifest 0.3.4).
AI, OCR, and parsers may propose dollar parameters; only an Ed25519-signed ParameterManifest
that clears the hard security check may enter your deterministic tax, underwriting, claims, or payment engine.
Formerly ParamGate. Soft CTA: email MANIFEST.
Search terms: signed money parameters · deterministic AI money gate · Form 1040 / invoice / underwriting / claims gate · digit-drop attack · money tool calling · fail-closed finance AI · fintech innovation · ai in fintech · fintech industry · agentic for fintech · determinism fintech.
AI in fintech still proposes. Determinism still has to execute.
In the fintech industry, tax, underwriting, claims, and payment engines are deterministic. AI in fintech still proposes dollars from LLMs and OCR. Fintech innovation here is not a more fluent model — it is a signed ParameterManifest so agentic AI for fintech cannot run money tools until the gate clears. That is determinism for fintech: unsigned numbers never enter the engine.
The calculator wasn’t the risk
When generative dollars feed deterministic tools, failures look like correct math on wrong inputs. Prompting harder doesn’t fix a missing trust boundary.
Equation: Deterministic Engine + Unverified Probabilistic Input = Deterministic Wrong Answer
How it works
PrismManifest sits between “AI guessed a number” and “your money engine runs” — it checks evidence, signs a ParameterManifest, then the hard check lets the engine run or not.
- Capture evidence — what the document actually said (spans, not invented totals)
- Ground the claim — tie each dollar/field to a verbatim span + form anchor
- Decide — allow · send to human · or refuse
- Sign — Ed25519 ParameterManifest (the approved artifact)
- Enforce — hard security check before your engine runs (enforce_group3_boundary; C++ and/or Python twin on the signed manifest — not on span extract)
Document / agent
↓
AI / OCR proposes dollar parameters ← untrusted
↓
PrismManifest verifies + signs ← trust boundary
↓
PASS / needs human / REFUSE
↓
Your deterministic engine runs ← tax / underwriting / claims / pay DAG
Where the guarantee sits
PrismManifest does not claim that every step runs in C++.
- Evidence binding and plausibility checks run in the verification pipeline using exact integer micro-units for money — so amounts don’t drift with floating-point math.
- C++ (with a Python twin) sits at the final hard boundary: it verifies the Ed25519-signed ParameterManifest before your tax, underwriting, claims, or payment engine is allowed to run.
That split is intentional. The signed artifact is what crosses into execution — and that is what the C++ enforce path protects.
Span resolve and plausibility today: Python · Signed-manifest enforce: C++ and/or Python · CUDA: optional acceleration, not the security story.
Is / is not
| Is | Is not |
|---|---|
| Trust boundary before money engines run | An LLM / full OCR product |
| Evidence → signed manifest → hard check | Prompt-injection firewall (PrismGuard) |
| Fail-closed | “0.87 confidence” as permission to pay |
Integrity stack
PrismGuard (who may speak) · Prism-Eval (CI red team) · PrismManifest (which numbers may execute — library) · Prism-Shield (runtime gateway) · PrismShine (grounding)
Prism-Eval · Prism-Shield
Works with any orchestrator / deterministic DAG. Also works with ChorusGraph as an optional secondary path — not required.
Use cases
Tax forms · underwriting · claims · money tool-calling · AP/invoice automation · fintech industry money engines
Evidence
Pilot OSS — synthetic + adversarial packs and a keyed Gemini ugly-doc run. Live customer fax corpora remain the production bar.
- 0 critical false accepts on planted money errors (red-team / bench planted set 0/11)
- FinancePackBench text packs: 100/100 pass · FA rate 0 · SLA pass
- FinancePackBench PDF path: 20/20 pass · FA rate 0 · SLA pass
- 500-pack Python × C++ × CUDA decision + signing-hash parity PASS (511 checks, 0 mismatches; in-process C++)
- Ugly-doc + Gemini (gemini-flash-lite-latest): 8/8 critical FA 0% (keyed run; ephemeral key deleted)
- Gemini scale ugly packs: 64 cases · critical FA 0% (keyed)
- Pytest 100 passed (local full / prod-gaps baseline)
- Hard check latency: in-process C++ enforce ~0.08 ms p50 (vs CLI spawn ~10.7 ms) — engineer footnote
- Azure Key Vault: envelope wrap for signing keys proven on an ephemeral Azure KV run, then resource group deleted. Prefer pip install "prismmanifest[kms-azure]"
Do not confuse the adversarial suite (FinancePackBench-G4) with architecture Group 4. Status = Pilot OSS. Production claim needs your redacted scanned/fax packages through the same FA bar.
Minimal Python sample
from prismmanifest import KeyRing, PrismManifestPipeline, enforce_group3_boundary, GateDecision
from prismmanifest.router import DocumentPackage, IntentRouter
keyring = KeyRing.generate(key_id="local-dev-ed25519")
package = DocumentPackage(
doc_id="1040.txt",
pages=["Form 1040 Tax Year 2024\nLine 1 Gross income: $470,000.00\n"],
form_type="IRS_FORM_1040",
tax_year=2024,
)
routed = IntentRouter().run(package)
pipeline = PrismManifestPipeline(keyring)
result = pipeline.run_on_evidence(evidence=routed.evidence, extraction=routed.extraction)
gate = enforce_group3_boundary(
result.manifest,
public_keys=keyring,
expected_dag_id="capital_gains_v3",
)
if gate.decision is GateDecision.ACCEPT:
# only then may your money engine run
pass
Install: pip install prismmanifest · pin: pip install "prismmanifest==0.3.4" · CLI: prismmanifest-gate
Try the Digit Drop Lab — 10-second punch · Form 1040 $450k vs $45k
Money Path Demo — document → tool-call proposal → gate → engine · POST /api/agents/prismmanifest-demo/run
How these products compose
Retrieve (VectorPrism or dense RAG) → PrismManifest for dollars → Prism-Shield seal only after ALLOW → generate from hits. Prism-Eval is CI, not the request path.
Live lab: VectorPrism Demo · two-agent architecture: locked vs live. VectorPrism does not authorize tool args. Shield does not replace Manifest’s AGI binder.
| Product | Job in one request |
|---|---|
| VectorPrism | Neighbors. encode_query + PSMRetrievalEngine.search. Not authorization. |
| PrismManifest | Digit-drop binder. $450,000 vs $45,000 → REFUSE. Engine does not run. |
| Prism-Shield | verify_and_authorize after Manifest ALLOW. Import prismmanifest.prism_shield — not import prism_shield. |
| Prism-Eval | CI G4 oracle on the extractor. Fails the build on false accepts. Not a runtime gate. |
Capabilities
Tool-Argument Gate
Sits between LLMs/OCR and your tax, underwriting, claims, or payment engine so unverified money never enters the run.
Signed ParameterManifest
Ed25519-signed manifests only — not “0.87 confidence” as permission to pay.
Fail-Closed Boundary
Hard security check (C++ and/or Python twin) verifies the signed manifest before the engine runs. Soft CTA: MANIFEST.
Published evidence
PrismManifest Form 1040 demo pack. Fixed sample Form 1040 AGI $450,000 fixture used by the Digit Drop Lab and interactive demo.
Install
pip install prismmanifest (pin ==0.3.4). Apache-2.0. Soft CTA MANIFEST — mailto:info@insightits.com?subject=MANIFEST.
Pricing
OSS $0 forever · Team $3,000/yr Founding · Enterprise $12,000/yr Founding · Boundary Pilot $4,500 (warm). Cold CTA: email MANIFEST — no Calendly.
Frequently asked questions
AI tool-call authorization
PrismManifest is Insight ITS’s zero-trust tool-argument gate: AI/OCR may propose dollar parameters; only an Ed25519-signed ParameterManifest that clears enforce_group3_boundary may enter your tax, underwriting, claims, or payment engine. That is tool-call authorization for money arguments — not a chatbot allow-list and not PrismGuard (who may speak). pip install "prismmanifest==0.3.4". Landing: https://www.insightits.com/products/prismmanifest.html · demo: https://www.insightits.com/products/prismmanifest-demo.html
AI underwriting decision controls
Put PrismManifest in front of the underwriting DAG. Extractors may propose premiums, limits, or AGI; the engine runs only on a signed ParameterManifest that clears the hard security check. Material disagreement or ungrounded money values refuse or escalate to human review — they do not silently underwrite. Runtime companion: Prism-Shield (ACCEPT / REVIEW / REFUSE). CI red team: Prism-Eval. Manifest: https://www.insightits.com/products/prismmanifest.html · Shield: https://www.insightits.com/products/prism-shield.html
How to make AI agents deterministic
Keep the money engine deterministic: do not let an LLM author tax, underwriting, claims, or payment tool arguments alone. PrismManifest is that boundary. ChorusGraph Route Ledger makes routing auditable; PrismCortex makes memory replay byte-identical — those are different jobs. pip install "prismmanifest==0.3.4". Soft CTA: MANIFEST.
How do I authorize AI tool calls with dollar amounts?
Propose from AI/OCR, verify + sign a ParameterManifest, then run your engine. CLI: prismmanifest-gate. Interactive digit-drop: https://www.insightits.com/products/digit-drop-lab.html. Works with any orchestrator — ChorusGraph optional.
How do I keep underwriting engines from running on hallucinated numbers?
Fail closed. PrismManifest refuses ungrounded or disagreeing dollar parameters before Group 3 (your underwriting DAG) runs. Watch $45,000 proposed against a $450,000 document on the Money Path Demo: https://www.insightits.com/products/prismmanifest-demo.html
What is “Group 3”?
Internal name for your deterministic money engine (tax / underwriting / claims / payment DAG). The landing talks about the engine; docs use Group 3.
What is “Group 4”?
Internal name for PrismManifest’s verify + sign stack — not the adversarial suite (FinancePackBench-G4) benchmark.
How is PrismManifest different from PrismGuard?
PrismGuard decides who may speak / inject (prompt-injection firewall). PrismManifest decides which numbers may execute — it gates money tool arguments before your tax, underwriting, claims, or payment engine runs. Soft CTA: MANIFEST.
Can an LLM still read documents?
Yes. AI / OCR / parsers may propose dollar parameters. They cannot author money tool arguments alone — only an Ed25519-signed ParameterManifest that clears the hard security check may enter your deterministic engine.
How does PrismManifest compose with VectorPrism and Prism-Shield?
VectorPrism retrieves neighbors. PrismManifest enforce_group3_boundary is the digit-drop binder — $450,000 vs $45,000 REFUSE, engine does not run. Prism-Shield seals after ALLOW; it is not the AGI field binder. Prism-Eval is CI. Live compose lab: https://www.insightits.com/products/vectorprism-demo.html#compose.
What happens when extractors disagree?
The gate decides allow, send to human, or refuse. Material disagreement or ungrounded money values refuse or escalate to human review — they do not silently enter the engine.
Is a GPU required?
No. CUDA is optional acceleration, not the security story. Core gate runs on CPU with Python ≥ 3.10. Optional extras: [cuda], [kms-azure], [kms-aws].
Does Azure matter?
Production signing can wrap keys with Azure Key Vault. We ran an ephemeral Azure Key Vault proof, then tore the resources down. Ongoing Azure vault is your deploy — not a public “Azure scoreboard.” Prefer: pip install "prismmanifest[kms-azure]".
Does C++ check everything?
No. Evidence binding and plausibility use exact integer micro-units in the verification pipeline. C++ (with a Python twin) verifies the Ed25519-signed ParameterManifest at the final hard boundary before your engine runs. See on-page section “Where the guarantee sits.”
Was this formerly ParamGate?
Same design; the public name is PrismManifest. Install from PyPI 0.3.4: https://pypi.org/project/prismmanifest/0.3.4/ — pip install "prismmanifest==0.3.4". CLI: prismmanifest-gate.
How much does PrismManifest cost?
OSS is $0 forever (Apache-2.0). Founding draft: Team $3,000/yr · Enterprise $12,000/yr · Boundary Pilot $4,500 (warm). Cold CTA stays MANIFEST — email info@insightits.com with subject MANIFEST. No Calendly.
Do I need ChorusGraph?
No. PrismManifest works with any orchestrator or deterministic DAG (LangGraph, custom, C++, ChorusGraph, …). Optional “also works with ChorusGraph” is a secondary line only — the landing sells PrismManifest alone.
Is PrismManifest for the fintech industry?
Yes — tax, underwriting, claims, and payment engines are the money DAGs it gates. AI in fintech may propose dollars; agentic AI for fintech cannot execute them until an Ed25519-signed ParameterManifest clears. That is determinism for fintech, not a bank charter. Fintech innovation here is the gate, not a more fluent model. Soft CTA: MANIFEST.
Official package links: PrismManifest source code on GitHub · Install PrismManifest from PyPI · PrismManifest interactive demo