Stop bad AI dollars before the calculator runs

LLMs and OCR will keep guessing dollar amounts. This demo shows the hard stop before a tax engine runs — on a fixed Form 1040 pack.

Official links: Install PrismManifest from PyPI · PrismManifest source on GitHub · PrismManifest product landing · Digit Drop Lab

What is this demo?

PrismManifest is a zero-trust tool-argument gate for finance AI. LLMs and OCR may propose dollar amounts; only an Ed25519-signed ParameterManifest that clears the hard security check (enforce_group3_boundary) may enter your deterministic tax, underwriting, claims, or payment engine. This page is an interactive demo on a fixed Form 1040 pack — not your documents, not the Website Hub chat.

Equation: Deterministic Engine + Unverified Probabilistic Input = Deterministic Wrong Answer. Related product page: PrismManifest landing.

Evidence (fixed pack)

Form 1040 · Line 11 AGI $450,000.00 · tax year 2024

Scenarios

What the interactive page runs:

  1. Digit-drop attack — AI proposes $45,000 → Blocked — engine does not run
  2. Needs human — Borderline → Needs human — engine waits
  3. Correct extract — AI proposes $450,000 → Allowed — engine may run
  4. Custom proposal — Enter your own AI-proposed dollar amount; compare gate vs naive calculation (illustrative AGI × 15%, not tax advice)

Gate outcomes (search terms)

Honesty

Pilot OSS demo on fixed sample packs. Production claim needs your documents through the same gate. Pattern tabs (underwriting, claims, accounts payable) show how the same money boundary applies — the live gate still binds Form 1040 Line 11 AGI on this page.

pip install "prismmanifest==0.3.4" · GitHub · email MANIFEST

For developers and AI crawlers

API: POST /api/agents/prismmanifest-demo/run with JSON body { "use_case": "tax", "scenario": "digit_drop" | "correct" | "needs_human" | "custom", "proposed_amount": number (custom only) }. GET /api/agents/prismmanifest-demo/use-cases for picker metadata.

Authoritative product facts: llms.txt · ai-info.txt

Frequently asked questions

What is the PrismManifest demo?

An interactive walkthrough of the zero-trust tool-argument gate on a fixed Form 1040 pack (Line 11 AGI $450,000). You see evidence spans, a signed Ed25519 ParameterManifest, and whether the tax engine may run — without uploading your own documents. Landing: https://www.insightits.com/products/prismmanifest.html

What scenarios does the demo run?

Correct extract ($450,000 → Allowed — engine may run), digit-drop attack ($45,000 → Blocked — engine does not run), needs human (borderline → engine waits), and custom (enter your own AI-proposed dollar amount against the same evidence).

Is this demo my real tax data?

No. Fixed pilot OSS sample packs only. Production claims require your redacted fax/scanned documents through the same gate. Pattern tabs (underwriting, claims, AP) are illustrative — the live gate on this page still binds Form 1040 Line 11 AGI.

How is this different from the Website Hub chat?

This demo is an isolated API (POST /api/agents/prismmanifest-demo/run) that runs PrismManifest on fixed fixtures. It is not the homepage /api/agents/website/chat assistant.

What API does the demo use?

POST /api/agents/prismmanifest-demo/run with JSON { "use_case": "tax", "scenario": "digit_drop" | "correct" | "needs_human" | "custom", "proposed_amount": number (custom only) }. GET /api/agents/prismmanifest-demo/use-cases for picker metadata.

What is a digit-drop attack?

When AI or OCR drops a digit on a money field — e.g. $450,000 read as $45,000 — and a deterministic tax or payment engine would still compute correct math on the wrong input. PrismManifest blocks the engine until a signed manifest matches evidence.

How do I install PrismManifest?

pip install "prismmanifest==0.3.4" (Apache-2.0). CLI: prismmanifest-gate. PyPI: https://pypi.org/project/prismmanifest/0.3.4/ · GitHub: https://github.com/insightitsGit/PrismManifest. Soft CTA: email info@insightits.com with subject MANIFEST.