IQAI Risk

AI-assisted documents need proof trails before reliance.

IQAI Risk reviews AI-assisted documents before they become reports, filings, client deliverables, summaries, public statements, or decisions.

Memos in. Review record out. The system structures claims, evidence boundaries, review labels, verification queues, human checkpoints, and receipts.
00 · Risk at a glance

Claim-level support review before documents become liability.

A concise view of what Risk inspects, how it works, and what evidence it gives the reviewer.

What is it?

A rule-governed review system for AI-assisted documents.

It inspects claims, evidence links, unsupported exposure, external verification needs, and review state.

Who uses it?

Legal, compliance, audit, consulting, risk, and procurement teams.

Built for high-stakes files where polished language can outrun evidence.

What does it reveal?

Weak support, unsupported statements, and unresolved review risk.

It separates memo evidence fit from facts that require external lookup.

What does it produce?

Claim registers, evidence maps, queues, labels, sign-off state, and receipts.

The output is designed for human review, not automatic legal truth.

01 · The problem

AI lowers the cost of drafting. It does not lower the cost of being wrong.

AI can make a document look finished before its proof trail is complete. The risk is reliance on claims that are weak, overstated, unverifiable, or unsupported by the file.

Polished output

The document reads as complete.

Fluent language does not establish provenance, evidence, or defensible support.

Mixed claims

Facts, assumptions, forecasts, and judgments blend together.

Reviewers need to know which statements are supported, inferred, or externally checkable.

Reliance event

The exposure begins when someone trusts the file.

A memo, report, filing, or generated answer becomes risk when it becomes relied-upon work.

02 · Product architecture

Claims, gaps, checks, receipts.

Risk turns a document into a structured review object: what was claimed, what supported it, what still needs checking, and what a human decided.

1
Document in

Memo, report, filing, summary, deck, response, or AI-generated output.

2
Claims

Checkable facts, forecasts, guarantees, comparisons, recommendations, or outcomes.

3
Evidence boundary

What the document itself can support, cite, quote, anchor, or show.

4
Labels

Supported, Weak, Unsupported, or Needs External Verification.

5
Checks

Registry records, filings, dates, market values, math, or configured lookups.

6
Human review

Reviewer accepts, revises, escalates, narrows, or rejects the claim.

7
Receipt

What was reviewed, what was flagged, and what remains open.

03 · Rule layer

The product is the boundary between wording and evidence.

IQAI Risk compares claim type against the memo evidence boundary. It asks whether the wording stays within the evidence, partially outruns it, fully outruns it, or needs an external check.

Question 1

What does the document claim?

Facts, forecasts, guarantees, comparisons, operational outcomes, identifiers, figures, and other reliance-bearing statements.

Question 2

What does the document itself support?

Memo-bound evidence is separate from outside truth. A claim can sound polished and still outrun the file.

Question 3

What needs outside verification or human judgment?

Registry identifiers, filings, market values, date checks, math reconciliation, and broad judgment calls move into separate lanes.

This is the control layer: a green label means the wording fits the provided evidence; it does not mean the outside world has been proven.
04 · Review labels

Four labels. One purpose: make reliance reviewable.

The labels describe evidence fit. They do not decide legal liability, certify truth, or replace human sign-off.

Label Plain-English meaning Review consequence
Supported The memo evidence is enough to stand behind the claim as stated. Supported by the provided file. Not a claim that the outside world has been proven.
Weak / Needs Review Evidence is thin, indirect, partial, ambiguous, inferential, or softer than the wording. A human should review the linked evidence and decide whether the wording is acceptable.
Unsupported The claim outruns what the memo actually shows. Stop signal for reliance from the memo alone. More proof, narrower wording, or escalation is needed.
Needs External Verification The claim points to an outside fact: registry number, filing reference, market rate, date-based value, or similar fact. A lookup may be required. The result informs review but does not automatically make broad claims Supported.
05 · Review surface

The file shows what needs attention.

Risk gives reviewers a quick picture of support posture, weak claims, unsupported exposure, high-impact claims, and external-check load.

Support rate
50%
Illustrative memo example.
Weak claim rate
42%
Claims needing review.
High-impact rate
83%
Claims with material review relevance.
Outside checks
67%
Claims routed toward verification.
Queues

Revision, escalation, and external checks.

Claims can be routed into clear reviewer queues instead of disappearing inside a polished draft.

Pattern flags

Overstated, causal, forward-looking, not proven.

The type of support failure is visible before sign-off.

Receipt

Review record for the file.

The run records what was checked, flagged, reviewed, and left open.

06 · Worked example

Guarantee language can outrun memo evidence.

The rule layer is easiest to see when a sentence promises a future operational outcome, but the memo only proves a process update.

Claim
“The updated cyber process ensures incidents will be escalated quickly.”

Claim type: forward-looking operational assurance. Triggers: “ensures” and “will be escalated quickly.”

Evidence fit

The memo shows the checklist was updated and circulated. That is evidence of process change. It is not proof that future incidents will be escalated quickly.

Checklist updated Circulated No tabletop exercise No SLA measurement No audit report
Result: the wording needs more proof, narrower language, or human review.
07 · Verification surfaces

Tiny facts. Big consequences.

Some claims are memo-evidence questions. Others point outside the file and need separate verification.

Registry identifiers

Entity names, NEQ numbers, corporate records, and official registry references.

Filing references

SEC filings, exhibit references, dated disclosures, and public-record anchors.

Price and date checks

Market rates, exchange rates, time-sensitive figures, and on-date values.

Math reconciliation

Arithmetic, ratios, percentage changes, subtotals, and memo-to-table consistency.

External verification informs review. It does not automatically convert a broad, overstated, or forward-looking memo claim into Supported.
08 · Early pilot

Human / LLM / IQAI Risk comparison.

A small human-in-the-loop pilot compared five human reviewers, multiple LLMs, and IQAI Risk on the same ten claim/evidence questions.

Human reviewers
5
MTurk participants, anonymized.
Questions
10
Claim/evidence pairs.
Support κ
0.664
Human support-label agreement.
Issue κ
0.402
Human issue-label agreement.
Study conclusion

The useful result is not “Risk beats AI.”

Humans and LLMs aligned strongly on obvious evidence gaps. IQAI Risk aligned directionally while adding governed scoring, repeatability, verification routing, and receipts.

Governance gap

Judgment is not the same as a control record.

Human and LLM answers can be useful, but they do not automatically produce queues, policy boundaries, verification routing, or receipt-level traceability.

Calibration question

Weak vs Unsupported

The pilot exposed the boundary that matters commercially: when a forward-looking claim has no direct evidence, should it remain Weak / Needs Review, or be promoted to Unsupported?

Directional early evidence, not a publication-grade validation study. The value shown is repeatable review infrastructure: claim support, verification surfaces, calibration policy, and receipts.
09 · Buyer deliverables

A review package, not just a pass/fail label.

The deliverable is designed for people who need to sign, publish, file, rely on, or defend a document.

Claim register

The document’s checkable claims extracted into a structured review list.

Evidence map

Linked memo evidence, figures, dated actions, quotes, schedules, and support anchors.

Support labels

Supported, Weak, Unsupported, and Needs External Verification labels.

Reviewer queues

Unsupported, weak, overstated, external-check, and revision queues.

Verification surface

Registry, filing, price/date, math, and configured external checks.

Human sign-off state

What a reviewer accepted, modified, escalated, rejected, or left unresolved.

Review receipt

What was checked, when, under which scope, and what remained open.

Export package

Structured evidence for governance, audit, quality, legal, or compliance review.

10 · Best-fit uses

Where documents become accountable artifacts.

The sector changes, but the reliance problem repeats: claims, gaps, checks, human judgment, and receipts.

Legal memos

Claims, authorities, citations, assumptions, and review state.

Consulting reports

Client deliverables, references, evidence support, and unsupported recommendations.

Investor / board materials

Diligence summaries, risk statements, forecasts, causal claims, and management assertions.

Regulatory responses

Compliance conclusions, remediation claims, control language, and external verification needs.

11 · What Risk does not claim

Decision support, not an automatic truth engine.

Risk supports review. It does not replace the people responsible for sign-off.

No legal determination

Risk does not decide legal liability, regulatory breach, or final release approval.

No automatic truth certification

Supported means supported by the provided file, not proven true in the outside world.

No replacement for experts

Legal, compliance, audit, domain, and editorial reviewers retain authority.

Risk conclusion

Before a document is trusted, its claims should be reviewable.

IQAI Risk turns AI-assisted documents into claim/evidence records, review queues, human checkpoints, and receipts before reliance.