🎮 The Next Input — Issue #211

In partnership with

⚡ The Briefing — 60 sec

• Anthropic’s annualised revenue surges to $65B WTF. There’s really not much else to say. Watching frontier labs go from “interesting research company” to generating nation-state-sized numbers has been absolutely bonkers.

• Google wins $10 million Spirit Airlines data auction for AI training Honestly? It’s a steal. Ten million bucks for a unique corpus of real-world operational data at Google scale is basically somebody finding a rare dataset in the bargain bin.

• Australian social media ban report criticised over AI hallucinations There are few better ways to undermine an important policy discussion than confidently inserting made-up shit into the evidence base. AI can accelerate research. It can also accelerate you directly into embarrassment if nobody checks the receipts.

🛠️ The Playbook — Evidence Integrity Engine

Mission
Build an AI research workflow that traces every consequential claim back to verifiable evidence before it reaches executives, customers or policymakers.

Difficulty
Intermediate

Build time
3–5 hours

ROI
Accelerates research and reporting while dramatically reducing hallucinations, unsupported claims and reputationally expensive corrections.

0) Why This Matters

Data is becoming extraordinarily valuable.

Frontier labs are paying for proprietary corpora because differentiated information can produce differentiated capability.

At the same time, organisations are using AI to consume more information than any human team could reasonably verify manually.

That creates the catch:

More information does not automatically mean better evidence.

An AI-generated report can look immaculate while containing:

• fabricated citations

• distorted statistics

• outdated sources

• unsupported causal claims

• summaries that do not match the underlying document

So the workflow has to change.

Generate quickly.

Verify ruthlessly.

1) Architecture

2) Workflow

1. Ingest original documents, datasets and authoritative sources into a controlled repository.

2. Store source metadata including publisher, date, ownership and retrieval location.

3. Require AI-generated research to produce discrete claims rather than one untraceable narrative.

4. Match every material claim to one or more supporting source passages.

5. Flag unsupported, contradictory or low-confidence claims for human review.

6. Publish only after the evidence trail survives a final verification pass.

3) Example Prompts

Claim Extraction

You are an evidence-integrity analyst.

Review the following draft report.

Extract every material factual claim.

For each claim return:
1. exact claim
2. importance: LOW / MEDIUM / HIGH
3. evidence required
4. cited source if available
5. whether the current source actually supports the claim
6. confidence level

Do not repair unsupported claims yet.
Flag them explicitly.

Evidence Verification

Compare each supplied claim against the retrieved source passages.

Classify each as:
- DIRECTLY SUPPORTED
- PARTIALLY SUPPORTED
- CONTRADICTED
- NOT SUPPORTED
- SOURCE INSUFFICIENT

For every classification:
1. identify the supporting passage
2. explain the relationship between evidence and claim
3. note qualifications omitted from the claim
4. recommend corrected wording where necessary

Never invent supporting evidence.

Citation Audit

Audit the following report and citation set.

Check for:
- citations that do not support the associated claim
- fabricated sources
- statistics missing context
- outdated evidence
- secondary sources where primary evidence exists
- causal claims supported only by correlation
- quotations that alter the original meaning

Return:
1. blocking issues
2. corrections required
3. claims requiring expert review
4. publication recommendation

4) Guardrails

• Never allow the model to fabricate a missing citation.

• Preserve original source documents rather than summaries alone.

• Prefer primary sources for consequential claims.

• Require claim-level evidence rather than bibliography-level evidence.

• Clearly distinguish fact, inference and opinion.

• Flag conflicting evidence instead of silently selecting one version.

• Record the model and prompt used to generate material reports.

• Require human review for legal, regulatory, medical or policy-sensitive conclusions.

5) Pilot Rollout — 3 hours

1. Select one recurring research or executive-reporting workflow.

2. Collect ten trusted source documents and preserve their metadata.

3. Generate a draft report using retrieval from only those sources.

4. Extract every factual claim and automatically map it back to supporting passages.

5. Send unsupported and low-confidence claims to a named human reviewer.

6. Compare the verified report against your existing research process for speed and correction rate.

6) Metrics

• Percentage of material claims with verified support

• Unsupported claim rate

• Citation mismatch rate

• Human-review escalation rate

• Research turnaround time

• Corrections required after publication

• Primary-source utilisation

• Retrieval precision

• Average evidence confidence

• Cost per verified report

Pro Tip: Do not ask AI to “provide sources.” Build a system where it cannot make the claim unless the source already exists.

🎯 The Arsenal — Tools & Platforms

• Azure AI Search · retrieves source-level evidence from controlled knowledge stores · Link

• PostgreSQL · stores claims, source metadata and verification records · Link

• Microsoft SharePoint · maintains governed source documents and organisational evidence · Link

• Microsoft Teams Approvals · routes uncertain or consequential findings to accountable reviewers · Link

• OpenAI · performs evidence-grounded extraction, synthesis and structured analysis · Link

Copy-paste prompt block:

You are designing an AI evidence-integrity system for my organisation.

Research workflows:
[LIST]

Source types:
[LIST]

Authoritative sources:
[LIST]

High-risk outputs:
[LIST]

Existing document systems:
[LIST]

Human reviewers:
[LIST]

The system must:
- preserve original evidence
- maintain source metadata and provenance
- separate factual claims from inference and opinion
- map every consequential claim to supporting evidence
- detect unsupported or contradictory claims
- prevent fabricated citations
- prefer primary sources where available
- route uncertainty to appropriate human reviewers
- maintain an auditable verification trail

Return:
1. architecture
2. source-ingestion standard
3. claim-level data model
4. retrieval workflow
5. verification rules
6. human-review thresholds
7. publication gate
8. pilot rollout
9. operational metrics

💡 Free Office Hours

AI can make research dramatically faster. It can also make bad research dramatically faster. The advantage increasingly lies in building systems that combine machine-scale synthesis with evidence humans can actually inspect.

Book here: https://calendly.com

Reply to everything. Edit nothing.

Your inbox is full. Slack is piling up. Client messages need a response yesterday. Typing thoughtful replies to all of it takes hours you don't have.

Wispr Flow turns your voice into clean, professional text you can send the moment you stop talking. Speak like you would to a colleague — tangents and all — and get polished output. Emails, Slack, LinkedIn, WhatsApp, whatever's open.

89% of messages sent with zero edits. Used by teams at OpenAI, Vercel, and Clay. Works on Mac, Windows, and iPhone.

Try Wispr Flow free

🕹️ Game Over

Data is worth millions.

Frontier labs are worth billions.

And a hallucinated citation can still make the whole thing look stupid.

— Aaron Automating the boring. Amplifying the brilliant.

Subscribe: link