Free workflow

AI research you'd stake a decision on.

A research system with confidence ratings on every claim, independent verification, and an adversarial review that attacks its own conclusions before you see them. See what it catches below.

The difference

Same question. Same AI. One has trust mechanisms. One doesn't.

"Evaluate VectorShield for competitive positioning potential in the AI security market."

Raw AI output

VectorShield is a leading AI security company with approximately $12M in funding. The AI security market is growing rapidly, with an estimated CAGR of 34.2% through 2028. The company has strong analyst coverage from Gartner and has been recognized with multiple industry awards. Three of five direct competitors were acquired in 2025, validating the market and creating significant opportunity. VectorShield's offensive security approach differentiates them from competitors, and their university research partnership gives them a technical edge. The company appears well-positioned for Series B fundraising.

Looks thorough. Reads well. Would you stake a decision on it?

After the workflow

[Strong — direct, consistent, multi-source]

VectorShield has $11.9M in total funding ($3.95M seed + $8M Series A). Revenue estimated ~$556K.

3 sources: Crunchbase, PitchBook, press release (DOI verified)

⚠ Precision concern

"34.2% CAGR through 2028" — no primary source found. This figure appears in training data only. Market growth estimates from verified analyst reports range from 22-40% depending on category definition.

[From training data — not verified] • Emerging confidence ceiling

[Moderate — consistent evidence, directness concern]

3 of 5 direct competitors acquired in 2025 (Protect AI → Palo Alto $634M, CalypsoAI → F5 $180M, Lakera → Check Point). Market validation is real, but VectorShield is the smallest independent player by every metric ($11.9M vs. next-smallest at $56M).

4 independent sources • ⚠ Missing counter-evidence: acquisition wave may indicate market consolidation that disadvantages smallest players

[Emerging — single dimension, synthesis]

"Well-positioned for Series B" is author synthesis. No named investor interest found. 3 PR agencies in 14 months is a volatility signal. Revenue ($556K) is thin for Series B at current valuations.

[Author synthesis] • Premortem flag: if Series B doesn't close by Q3 2027, competitive gap widens irreversibly

The raw output would have been shipped as "research." The system caught 4 issues in 4 claims — unsourced precision, missing counter-evidence, confidence inflation, and a synthesis claim presented as fact.

Who uses this

PR agency researching a prospect

"We're pitching a Series A cybersecurity startup next week. What's their media footprint, who's covering them, where are the gaps we can fill?"

Full media coverage audit with publication tiers, coverage gap analysis across trade and business press, competitor share-of-voice comparison, and LLM visibility assessment. Every finding confidence-rated so you know which gaps are verified opportunities and which are assumptions. Walk into the pitch knowing more about their PR landscape than they do.

Consultant scoping a new engagement

"A logistics company wants AI implementation. What's actually working in their industry vs. what's just vendor hype?"

Evidence-tiered analysis separating peer-reviewed outcomes from vendor case studies from blog posts. Triangulated ROI claims with precision flags on suspiciously round numbers. The adversarial review surfaces the 3 most likely ways the engagement could fail — before you scope it.

Founder prepping for investor meetings

"I need market sizing for my pitch deck. Every number I find online contradicts the last one."

Market size estimates triangulated from 3+ independent sources, each with source tier and methodology notes. Claims that can't be independently verified are flagged — not presented as facts. Your investors will ask where the numbers come from. You'll have the answer.

Agency owner evaluating a new service line

"Should we add AI implementation services? What does the evidence actually say about demand vs. what LinkedIn influencers are claiming?"

Dedicated contrarian search that separates practitioner evidence from hype. The 70–85% AI implementation failure rate — is that real, or is it one survey cited everywhere? The system finds out, rates the confidence, and tells you what changes in your decision if the bullish case is wrong.

Get the full system.

The workflow, the confidence framework, and installation instructions. Enter your email and the system reveals below.

01 — The process

Convergence-driven spiral.

Most AI research runs once: gather, write, done. Errors in early phases survive to the final output. This system runs a spiral — gather, synthesize, verify, audit, then loop back if quality triggers are still active.

Five triggers drive the loop: confidence gaps, coverage gaps, triangulation failures, missing counter-evidence, and stale sources. The spiral stops when fewer than two triggers remain active — or after 4 passes maximum. Quality-based termination, not arbitrary iteration.

Evidence: spiral processing catches 90% more issues than sequential single-pass (Saeidmehr, Steel & Samavati, 2024, Systematic Reviews 13:32).

02 — Claim verification

Evidence-first, not self-review.

Self-audit — asking AI to check its own work — confirms errors instead of catching them. This system retrieves evidence independently per claim, without re-reading the draft. Three layers of citation verification:

Existence — does this paper or source actually exist? (Crossref DOI lookup)
Metadata — do the authors, year, and journal match? (Crossref + Semantic Scholar)
Semantic support — does the source actually say what the claim says it says?

Claims that can't be verified aren't discarded — they're classified by why: no source found, source doesn't support, author synthesis, or unverified training-data recall. You always know what you're trusting and why.

03 — Confidence framework

Four domains, not one number.

Star ratings destroy decision-relevant information. "Strong evidence about the wrong question" and "weak evidence about the right question" both get 3 stars — but require completely different responses. This system assesses 4 independent domains per claim:

Evidence strength — source tier, sample size, replication status
Consistency — do independent sources agree?
Directness — does the evidence answer the question you actually asked?
Synthesis integrity — are the inferences sound, or is this a leap?

Plus two optional flags: precision concern (suspiciously round numbers, small samples) and completeness concern (suspected missing studies). You see why something is uncertain, not just how much.

04 — Adversarial review

Premortem, then targeted critique.

After the final spiral pass, the system attacks its own conclusions. Step one: "Assume this research has been proven wrong in 12 months. Write the 3 most likely reasons why." Step two: examine the specific claims each vulnerability touches. Step three: strengthen the evidence, weaken the claim, or document the limitation.

Premortem via prospective hindsight produces ~30% more identified risks than forward-looking analysis (Klein, 1998). If the premortem changes nothing, the system explains why — genuinely robust, or was the critique shallow?

Get the full system

Install in 30 seconds.

The complete system — spiral process, source discovery engine, claim verification pipeline, GRADE confidence framework, adversarial review, output templates, and source hierarchy reference — is a Claude Code skill you can install and run on any research question.

Works with Claude Code and Cowork. Installation is one command. The skill includes source tier definitions, MCP tool reference, evaluation rubrics, and 10 modes from quick briefs to full deep dives.

Install from GitHub

Want this as part of a broader system?

The research workflow is one piece. I build AI-native systems for operators — research, prospect intelligence, content pipelines, signal detection — each with the same structural guarantees. If you want help implementing this for your team, let's talk.

hello@thinkmetis.co