01
Framing
The question is disambiguated first
Before any engine sees it, the question is structured and its ambiguities resolved — so every engine that follows is answering the exact same question, not five slightly different ones.
LEX Routing · Inside EX MULTIS
EX MULTIS puts a single question through a sequential, adversarial pipeline of independent AI engines — framed once, answered apart, cross-examined, verified, and only then synthesized.
Running engines side by side is a comparison. Running them in sequence is a deliberation.
UNA QUAESTIO · MULTAE MENTES · RESPONSUM FIRMIUSOne Question · Many Minds · A Stronger Answer
The Problem
A
sk five different AI engines the same question and you get five confident answers, with no way to know which one is quietly fabricating. Averaging them together doesn't fix that — it just launders the disagreement into something that sounds settled. The disagreement between engines was never the problem. Having no way to see it was.
The case for a deliberation, not a comparison
The Sequence
Five stages. One deliberation.
Not five windows open side by side — a pipeline, run once, where each stage exists to make the next one harder to fool.
02
Independent Response
Genuinely different engines answer apart
Different architectures, different vendors, no visibility into one another's work. Agreement that's too cheap to earn — models that saw each other's answer first — is worthless as evidence.
03
Cross-Examination
Each response is turned against the others
Every claim is tested for assertions the others didn't make, reasoning they didn't follow, conclusions nothing supports. One engine's invented detail is rarely invented identically by another.
04
Verification
Surviving claims are checked, not assumed
What's left after cross-examination gets verified rather than trusted on confidence alone. Genuine disagreement is preserved as disagreement — never averaged into mush.
05
Synthesis
One answer, with its record attached
A coherent answer is assembled from what held up — delivered alongside the record of what was contested, what was corrected, and what remains genuinely unresolved.
Three Commitments01 02 03
What the pipeline is built to protect
Independence
Genuinely different engines — not the same model, asked twice.
Visibility
Where engines disagreed, you see it — an honest dispute beats a smooth answer that hides one.
Traceability
Every answer carries its lineage — you can audit exactly how it was reached.
See It In ActionQuestion Under Review“Is this market-sizing claim accurate?”4 engines·3 valid·1 dropped·confidence: mediumDEGRADED
The claim ledger
The output of a run isn't a paragraph — it's a ledger. Every claim is entered, labeled, and left open to audit. Click a row to see why it was labeled that way.
Illustrative example — not a live model callWhy this label: the input cited a source for this figure. The pipeline did not independently re-run that source this cycle.
Action: flag for the user
Why this label: asserted as fact by the input; no engine could locate a supporting source for the 80% figure or the four-vendor count during verification.
Action: remove or qualify
Position: one engine found TCO dominating procurement conversations in the sources it checked.Counter-position: a second engine found security and compliance ranked higher in the sources it checked.Resolution: flagged as unresolved — both positions are cited rather than one being silently preferred.
Action: flag for the user
Why this label: cross-examination surfaced an existing competitor serving this niche, contradicting the claim directly.
Action: remove or qualify
What It Catches
The failure modes a single engine won't tell you about
Confident fabrication
A single engine states an invented detail as flatly as a real one. Cross-examination catches it because a second engine rarely invents the same detail.
Single-model blind spots
What one architecture consistently misses, another was trained to catch. Independence exists so those gaps don't line up.
Unstated uncertainty
Ask one model how sure it is and it tells you what you want to hear. Verification checks the claim instead of trusting its confidence.
Framing errors
A misread question produces a well-argued answer to the wrong thing. Framing catches the mismatch before it reaches any engine.
What It Is Not
- Not a chatbot with a wrapper — and not a side-by-side comparison tool. The pipeline is sequential and adversarial by design.
- Not a claim of infallibility — it reduces error meaningfully, it doesn't eliminate it, and it's built to make remaining uncertainty visible rather than pretend it away.
- Not a replacement for judgment, and not a model of its own — it's the architecture and discipline that make independent engines useful together, while keeping a person meaningfully in the decision.