PAi Claims Evidence Architecture

The claim is built from traceable evidence — not by AI.

PAi Claims is designed to let artificial intelligence do what artificial intelligence does best: read, compare, recognize patterns, reason across large records, find gaps, estimate, research, explain and draft.

Then software outside the AI keeps the factual record grounded in real sources, real provenance and real evidence.

Do not weaken the AI. Control its authority.

One approach to AI reliability is to constantly tell the model to be cautious, avoid assumptions and refuse to move until it is certain.

That can also remove much of what makes AI useful.

PAi Claims takes a different approach.

Let the AI think broadly. Let the evidence system decide what is allowed to become fact.

The AI can investigate possibilities, notice relationships, suggest repair methods, compare estimates, find inconsistencies, identify missing information and help build arguments.

But an AI conclusion does not become stronger merely because the model sounds confident.

Two different jobs.

AI supplies intelligence.

  • read large amounts of material
  • inspect photographs
  • identify possible damage
  • compare documents
  • look for inconsistencies
  • recognize patterns
  • suggest missing evidence
  • help determine possible repair scope
  • generate preliminary estimates
  • research complicated questions
  • reason across the entire claim
  • draft explanations and correspondence

The software controls authority.

  • where each proposition came from
  • which source supports it
  • whether the source is independent
  • whether the proposition is provisional
  • whether another source corroborates it
  • whether evidence contradicts it
  • whether stronger evidence replaced it
  • which facts remain unresolved
  • whether the workflow can proceed
  • the history of how the proposition changed

An AI observation is allowed to be useful before it is proven.

Suppose the AI examines a photograph and sees what appears to be damaged shingles.

That observation is valuable.

PAi Claims does not need to throw it away just because the AI was the first thing to notice it.

AI sees possible roof damage → Provisional proposition → Look for independent support

The important distinction is that the AI observation begins as an observation with a known source.

It is not silently converted into: "the roof is damaged and therefore this repair is owed."

One observation does not prove everything attached to it.

If AI identifies possible roof damage, PAi Claims does not automatically promote all the other statements that might follow from it.

Damage exists? → Cause? → Quantity? → Repair method? → Material? → Price?

Those are separate factual propositions.

They can have different sources, different support and different levels of certainty.

The software does not let one true statement drag six unproven statements across the line with it.

Evidence promotes the proposition — not the AI.

A proposition can become stronger when independent evidence arrives.

AI observation
The photograph appears to show wind-damaged shingles.
Inspection
A physical inspection independently documents lifted and creased shingles.
Measurement
Roof measurements independently establish the affected quantity.
Contractor
A contractor provides a repair scope and pricing tied to the observed condition.
Invoice
Actual completed work establishes a real transaction and actual cost.

The AI may have discovered the issue first. The later evidence is what strengthens the factual record.

The AI cannot promote itself.

Asking the model the same question repeatedly does not create independent evidence.

Having two AI prompts repeat the same conclusion does not turn the conclusion into two sources.

A summary generated from a photograph does not become independent evidence from that photograph.

More AI output is not the same thing as more evidence.

PAi Claims tracks provenance so the system can distinguish genuinely independent support from information derived from the same underlying source.

Contradictions are information too.

Real claims contain disagreements.

A photograph may suggest one condition. A contractor may report another. A carrier estimate may omit the item entirely. An invoice may later reveal something different again.

PAi Claims does not need the AI to hide those conflicts in order to produce a clean answer.

A contradiction remains a contradiction until something actually resolves it.

The AI can analyze the disagreement and explain what additional evidence might resolve it.

What it cannot do is quietly choose the version it likes best and rewrite the claim history.

Confidence is not evidence.

Language models are designed to produce coherent answers.

A coherent answer can still begin from an incorrect observation, incomplete source material or an unsupported assumption.

PAi Claims therefore does not treat the tone or confidence of the model as a measure of factual strength.

The model can be extremely confident. The evidence still has to be there.

Arithmetic does not strengthen the source either.

A calculation can be mathematically perfect while one of its inputs is still uncertain.

For example, multiplying a speculative quantity by an accurate material price does not magically make the quantity proven.

Uncertain quantity × Reliable price = Useful provisional estimate

The result may be useful for working the claim. Its evidentiary strength still depends on the strength of the underlying propositions.

This lets PAi use AI aggressively.

Once factual authority is controlled outside the model, AI can be used for far more than simple document summaries.

Investigate

Look across photographs, documents, estimates, correspondence and pricing for connections a person may otherwise miss.

Challenge

Compare competing scopes and values, identify omissions and find places where the evidence does not support a position.

Explain

Translate a complicated evidence graph into a readable explanation without replacing the evidence underneath it.

Research

Look for relevant technical or legal information while keeping external authority distinct from facts about the physical loss.

Estimate

Develop a useful preliminary estimate even when some values remain provisional and clearly marked as such.

Draft

Prepare supplements, comparisons, explanations and dispute documents from the structured claim record.

The software can stop the workflow when reality is still unresolved.

Some uncertainty is acceptable during investigation.

Other uncertainty becomes important enough that the next step should not proceed.

If a critical policy value has conflicting readings, a document still requires human verification, or necessary evidence remains unresolved, application logic can prevent later claim work from treating that information as settled.

The AI can keep thinking. The workflow does not have to pretend the missing fact is known.

The history stays attached.

Facts in a real claim change as better evidence arrives.

PAi Claims can preserve the difference between:

what AI first suggested → what was observed → what was measured → what was invoiced → what was paid

Better evidence can supersede an earlier proposition without destroying the record of how the claim developed.

That is the separation.

The AI can be intelligent.

Let it notice things. Let it ask questions. Let it compare. Let it reason. Let it search. Let it estimate. Let it explain. Let it draft.

The software keeps it real.

Preserve the source. Track provenance. Keep propositions separate. Require independent support. Preserve contradictions. Block unresolved critical facts. Maintain the history.

All the intelligence of AI. Factual authority stays with traceable evidence.

The AI does the heavy analysis. The evidence decides what survives.

That is the architecture behind PAi Claims.

It is not about making AI less capable.

It is about separating the thing that is extraordinarily good at finding answers from the thing responsible for deciding whether those answers are actually supported.