PAi Claims Evidence Architecture

The complete claim lives here.

Every source file stays connected to the claim. PAi Claims records what the file is, where it came from, how it was read, how confidently it was read, and whether independent readings agree.

The goal is simple: information should not enter the claim merely because an AI model said it saw something.

A claim begins with the source material.

PAi Claims does not begin by asking AI to summarize a folder of documents.

The original files become part of the claim workspace first. For a new insurance claim, the policy packet is required before the claim workspace is completed.

Documents are stored with the claim so later evidence records point back to files that actually travel with that claim.

Source file→ Stored with claim→ Content hash→ Text extraction→ Verification→ Evidence record

The original file is preserved.

When a document is attached, PAi Claims copies the file into the claim workspace before building evidence from it.

A cryptographic content hash is recorded for the stored file. That gives the system a way to recognize identical content even when the second copy has a different filename.

A renamed copy of the same document does not become a second independent source.

This matters because evidence strength should come from independent support, not from counting the same document twice.

PAi uses the most direct reading available first.

The ingestion engine does not send every document straight to a language model.

It first tries to obtain text deterministically from the file itself.

Digital documents

Plain text, structured text, HTML, spreadsheet-style delimited text, DOCX and ODT content can be read directly from their stored document data.

PDF documents

PAi reads the PDF's embedded text layer page by page and preserves page boundaries so later values can be traced back to the page where they appeared.

Scanned pages and images

Pages without enough usable embedded text can be rendered and passed through OCR. OCR confidence remains attached to the resulting reading.

One bad scanned page does not erase the rest of the packet.

A mixed policy packet may contain twenty clean digital pages and one scanned endorsement.

PAi Claims handles the document page by page. Clean text can remain readable while the scanned page is separately marked as requiring OCR or another reading method.

The ingestion record keeps track of the page count, which pages required OCR, the reading method, confidence information, and whether anything still needs visual review.

OCR is a reading, not automatic truth.

OCR reconstructs characters from pixels. That makes it useful, but fundamentally different from reading text that was embedded directly in the original document.

PAi Claims keeps that distinction.

UNVERIFIED The system does not yet have a reliable reading.
SPECULATIVE A reading or inference exists, but it does not yet have enough independent support.
EXPLAINED The value has a traceable direct basis, but has not necessarily been independently corroborated.
VALIDATED Independent readings or direct support agree strongly enough to establish the proposition at the higher state.

AI can become another reader. It does not become the authority.

When AI is available, PAi Claims can ask the model to perform a second reading.

The task is intentionally narrow: reproduce the document text as read. The model is not being asked to improve it, summarize it, fix spelling, normalize numbers, or decide what the document probably meant.

Reader 1

Embedded document text

The characters supplied directly by the digital file when available.

Reader 2

OCR

A reconstruction from the page image, carrying OCR confidence information.

Reader 3

AI transcription

An additional reading used for comparison, not permission for the AI to rewrite the source.

The readers are compared. The AI does not get to declare itself correct.

Agreement strengthens the reading. Disagreement stays visible.

PAi Claims reconciles independent readings of the same page.

When readers agree, that agreement can strengthen the resulting text. When they disagree, PAi does not resolve the problem by asking which value looks more reasonable.

Reading A Reading B Reading C → Compare word by word → Agreement or visible discrepancy

This is especially important for numbers: policy limits, deductibles, dates, form numbers, quantities, invoices and dollar amounts.

$312,000 and $412,000 are not "close enough." If the available readers cannot establish which figure is actually on the page, PAi preserves the conflict and asks the user to check the source document.

The software does not silently average or pick a winner.

An unresolved reading remains unresolved.

The user can be shown the conflicting values, the page where the conflict occurred, and the surrounding context needed to locate it.

A human confirmation resolves the immediate reading problem, but the confirmation remains identified as a direct human check rather than being magically promoted into multiple independent sources.

Model-only image transcription has its own gate.

Sometimes neither a usable text layer nor OCR can read a page.

A local model can still help transcribe or describe that image. But PAi Claims does not treat a model-only transcription the same way as a directly readable document.

The user sees the model-only transcription before claim figures are allowed to depend on it.

Unresolved document problems can stop estimation.

This is where ingestion becomes more than document storage.

If readers disagree on unresolved text, or a model-only page is still waiting for human review, the estimation-readiness check can identify those pages as blocking.

Document read→ Readers reconciled→ Conflict?→ User verification if needed→ Ready for figures

The idea is simple: do not perform arithmetic on a number nobody has actually established.

The evidence record remembers how it got there.

The ingestion result does not end as anonymous text.

PAi Claims can retain information such as:

The claim does not merely remember the answer. It remembers the path from source file to answer.

Supported document types.

The current ingestion pipeline can directly handle common claim and business-document formats including:

Source Primary reading method
PDF Embedded text first; scanned pages can fall back to OCR.
PNG, JPG, TIFF, BMP, WebP and other common images OCR where available; visual transcription can be requested when needed.
DOCX Document text extracted directly from the stored file structure.
ODT Document text extracted directly from the stored file structure.
TXT, Markdown, logs, JSON, XML, YAML Direct text reading.
CSV and TSV Direct delimited-text reading.
HTML Readable text extracted while decoding document entities.

Why this matters for AI.

Modern AI can process an enormous amount of claim material very quickly.

That power becomes much safer when the AI is reasoning over a record whose source, reading method, confidence, conflicts and provenance already exist outside the model.

PAi Claims gives AI the evidence to work with without giving AI authority to manufacture the evidence.

The AI can then compare, organize, analyze, research and draft against the claim record while the evidence architecture remains responsible for what is allowed to become established.

From file to supported claim.

Policy / photo / estimate / invoice→ Preserve original→ Hash→ Read→ Cross-check→ Resolve conflicts→ Evidence record→ Claim analysis

The complete claim lives here because the evidence never loses its connection to where it came from.