Multi-Model Intelligence

Three models analyse
every finding.
The system publishes
its mistakes.

LOFTREK's Cognitive System cross-references your internal procurement data with external intelligence across every jurisdiction where you operate. Disagreements between models are flagged, not hidden. That is how you know you can trust what remains.

Every brief includes a Verification Log

Other systems hide their errors. Ours publishes them. Findings removed for insufficient evidence are listed with the reason for removal. Corrections are visible because a visible correction workflow is evidence of a working quality system.

verification-log-week12.jsonl
06:00:12 Claude Finding SA-041: Linen contract 16% above market
06:00:14 GPT-5.4 CONFIRMED range 14-18%, ICE cotton data verified
06:00:15 Gemini CONFIRMED range 13-19%, adds TRY hedge note
06:01:33 Claude Finding SR-007: Supplier Y sanctions exposure
06:01:35 GPT-5.4 DIVERGENT subsidiary not on OFAC list, parent entity only
06:01:36 Gemini DIVERGENT agrees partial, flags for manual review
06:01:37 // Escalated to MANUAL REVIEW -- models disagree on scope
06:02:44 Claude Finding SA-012: HVAC parts 22% below Etimad avg
06:02:46 GPT-5.4 RETRACTED Etimad data is 2024 Q2, stale for comparison
06:02:47 Gemini RETRACTED confirms stale benchmark, insufficient evidence
06:02:48 // Finding REMOVED -- insufficient evidence. Listed in retraction log.

Most intelligence products tell you what they found. We also tell you what we got wrong.

  • Three independent AI models analyse every finding. Not one model checked by itself. Claude Opus, GPT-5.4, and Gemini 3.1 -- each produces its analysis without seeing the others.
  • Disagreements are disclosed, not resolved silently. When models diverge, the finding is flagged and the divergence is shown. You decide, not the algorithm.
  • Retractions are listed with reasons. Findings removed for insufficient evidence appear in the retraction log. A system that hides its mistakes is a system you cannot audit.
  • Every number has a source chain. Monetary figures link back to the commodity index, the tender database, or the financial filing they were derived from.

Fractal Diamond Refinement Protocol

A recursive quality framework applied across 87 independent expert analysis rounds. Every failure becomes a specification for the next improvement cycle.

Phase 01

Collect

11 data sources ingested simultaneously -- company registries, tender databases, commodity indices, sanctions lists, regulatory filings, and news across all operating jurisdictions.

Phase 02

Analyse

Three AI models run independently. Fishbone root cause analysis across 6 dimensions. Each finding rated by confidence, severity, and source quality. No single model bias.

Phase 03

Verify

Cross-model consensus scoring. Divergent findings escalated for human review. Insufficient-evidence findings retracted with documented reasoning. Verification log generated.

Phase 04

Deliver

Monday morning brief with severity-rated findings, evidence packages, and the full verification log. Every claim traceable. Every retraction visible. Ready for board-level review.


Production infrastructure analysis

Real numbers from a live multi-domain analysis. External attack surface assessment combined with structural, MEP, fire safety, materials, and seismic analysis -- completed across 87 expert rounds.

807
Threats identified from public data
external attack surface
451
CVEs catalogued and severity-rated
cross-referenced NVD
2,022
Findings across all analysis domains
87 expert rounds
3
Full audit cycles completed
convergence verified
Cyber Defence

External Attack Surface

Production infrastructure scanned from the outside -- the same view any adversary has. 807 threats identified, 451 CVEs catalogued, severity-rated, and cross-referenced against the National Vulnerability Database.

807 threats · 451 CVEs
Multi-Domain

Structural + MEP + Safety

Structural loads, thermal performance, radiation shielding, fire safety, seismic response, and materials compliance -- each analysed by domain-specific expert rounds with cross-verification.

2,022 findings · 53 experts R1
Verification

CERT Cooperation

Intelligence output shared with European national Computer Emergency Response Teams under active cooperation. Government security agencies considered the analytical output credible enough to act on.

Active EU CERT cooperation

Three models. Independent analysis. No single point of failure.

Every finding is produced by three AI models operating independently. This is not one model checking itself. It is three separate reasoning systems with different training, different architectures, and different failure modes.

C

Claude Opus

Anthropic · v4.6

Primary analysis engine. Strongest in nuanced reasoning, regulatory interpretation, and cross-jurisdictional analysis. Constitutional AI alignment.

G

GPT-5.4

OpenAI · Codex Pro

Independent verification layer. Different training data, different architecture. Catches failure modes that Claude's training may share blind spots on.

G

Gemini 3.1

Google DeepMind · Pro

Third independent opinion. Web-grounded, strong on factual verification and data freshness checks. Eliminates correlated failure between the other two.

When all three agree, confidence is high. When they disagree, you see why.

Disagreements are not failures -- they are the most valuable output the system produces. A finding where three models converge is reliable. A finding where they diverge tells you exactly where to focus your own judgment. No averaging, no silent resolution, no hidden doubt.


Start with zero risk. Scale with evidence.

Phase 1 runs entirely on public data. No NDA required. No internal system access. You commit nothing except 45 minutes to review what we find.

1

OSINT Mirror

Immediate -- public data only

We reconstruct your supply chain's external footprint as any outsider -- competitor, journalist, or adversary -- already sees it. Company registries, tender databases, sanctions lists, corporate intelligence, across every jurisdiction where you operate. The gap between what the mirror shows and what is actually true is itself the highest-value intelligence.

Zero internal data required
2

Procurement Intelligence Advisory

After Phase 1 delivers recognised value

Commodity timing signals, supplier risk monitoring, cross-property synergy identification, and competitive benchmarking from public tender databases. Weekly Monday briefs calibrated by your feedback on Phase 1 accuracy. Still no internal system access.

3

Cognitive System Node

Your infrastructure, your data

A dedicated analytical engine deployed inside your own environment. Cross-references your Oracle ERP procurement data with continuous external intelligence. Your data never leaves your infrastructure. LOFTREK provides the methodology and the protocol. You control the data.

Data never transits LOFTREK systems

Operational record

Twenty years of multi-site industrial operations. The AI capability was built by transforming our own operations first -- not as a research project, but as a survival imperative.

790+ Clients Across Europe

Multi-supplier operations across 6 European manufacturers. Industrial logistics, procurement, and compliance at scale for two decades.

First Anthropic Global Hackathon

Selected from 13,000 applicants. 500 accepted. Fewer than 200 shipped. FDRP methodology originated there and has since evolved into 184 documented methods.

Active EU CERT Cooperation

Working relationship with European national Computer Emergency Response Teams. Government security agencies considered analytical output credible enough to act on.

EU/NATO Jurisdiction

Romania -- European Union and NATO member state. GDPR-compliant by constitutional obligation. Genuinely neutral jurisdiction: no political baggage of intelligence-adjacent consultancies.


One meeting.
45 minutes.

We walk through what public data reveals about your supply chain. You tell us whether we found something real or something you already solved. Either way, 45 minutes well spent.