Deep Mode: our deepest multi-model AI analysis, forced onto a real capital decision.

By Alan Finney — Founder, 3Dogs Nexus

Most of what 3Dogs does happens behind one PDF. This time we ran it in deep mode — the setting that forces multiple independently-composed panels and won't skip the hard questions — and logged every stage: the research, the follow-ups, the disagreement, and a moment where the system caught a flaw in its own output.

What happens in the deepest mode of a multi-model AI decision analysis?

Deep Mode convenes several panels rather than one, re-researches between rounds, and reconciles the panels into a single recommendation. This case shows the whole thing end to end — two full ensembles, the research passes between them, and the decisive call that came out, with nothing hidden.

What actually happened, stage by stage

Deep mode doesn't just run one debate for longer — it re-runs research when the debate is unresolved, forces real clarifying questions instead of guessing, and composes a fresh panel each time. Here's the actual log.

Adaptive research ran four separate times — each one informed by what came before

Discovery didn't just gather facts once. It built a research plan (classifying the decision type and its stakes), ran targeted web-grounded research against that plan, then repeated the whole pass three more times — once after each round of clarifying answers, and once more after the debate itself flagged what was still unresolved. That last pass didn't just re-ask; it fed the panel's own live disagreements about reimbursement risk and physician retention back into fresh research questions before anyone answered them.

The clarification loop asked real questions — and didn't repeat itself

Deep mode won't let a thin brief through. It asked two rounds before Nexus even started, then one more after the first debate, each time targeting the specific number the panel needed and had not yet gotten.

Two complete 3-panel ensembles ran — six panels, no two alike

Deep mode doesn't re-run the same panel for confidence theater. Each panel was freshly composed — different size, different bench models pulled in, different seat framing — and each one argued the case from scratch before a fourth pass (Wisdom-of-Crowds) reconciled all three.

Real infrastructure limits showed up — and the system worked around them live

Large panels push against real-world API rate limits. Three different models (Llama 4 Maverick, Pixtral Large, Palmyra X5) got throttled mid-debate across the run. Each time, the system detected the failure, ejected that model for the remainder of the query, and automatically rerouted its seat to Haiku 4.5 rather than dropping the voice or crashing the debate. On the largest panel (15 seats, three models rerouted to Haiku simultaneously) Haiku itself briefly got congested and a handful of individual challenge exchanges failed outright — the debate absorbed the loss and kept going instead of stalling.

The final call — and an honest confidence number, not an averaged one

The panels' own average confidence across the second ensemble was 82% / 79% / 82%. The meta-synthesis didn't just report that average — it identified that all three panels converged on the same structural fault line (whether cardiologist retention can actually be secured before capital is spent) and discounted the final confidence to 74% to reflect that unresolved risk.

Bonus: the system audited itself — and found a real gap

After the PDF was delivered, 3Dogs' own Evolution process reviewed the case and flagged something genuinely useful: the pipeline's modification-list dedupe only catches exact-string duplicates, so two panelists issuing literally opposite instructions ("require binding retention agreements first" vs. "don't wait for binding contracts") could both survive to the client-facing report unreconciled. It filed that as a HIGH-severity fix proposal — propose-only, nothing changed automatically — and this exact demo query is what surfaced it.

What this is NOT

A single model answering once, a canned demo script, or a case where dissent got quietly averaged into a tidy number. Every stage above came out of the actual run logs — including the parts that show real strain (the 429 throttling) and a real, previously-uncaught gap (the modifications dedupe).

What deep mode is for

Not every decision needs this. Deep mode is the setting for capital commitments, board-level calls, and anything where "the model sounded confident" isn't a good enough reason — where you want the disagreement on the record, the confidence explicitly justified (or discounted), and a documented basis for the number you're about to defend.

It didn't converge because we asked it to. It converged because the evidence did.

Two independent ensembles, six differently-composed panels, real infrastructure hiccups mid-run, and a system that caught its own defect afterward — and the recommendation still landed in the same place both times, with an honestly discounted confidence number attached. That's the standard we hold deep mode to.

Questions this case answers directly

Plain answers on multi-panel deliberation, and where wisdom of crowds stops working. Every figure below comes from the delivered report for this case. These are clearly-labeled panel estimates from a multi-model adversarial analysis — not investment, legal or professional advice.

What is multi-panel deliberation and how does it work?

Instead of one panel debating a question once, several independent panels — each with a different model roster — run the same question separately, and their conclusions are then reconciled. The value is in the comparison: agreement across panels built from different models is meaningfully stronger evidence than agreement inside a single panel, because it is far less likely to be an artefact of one particular roster's shared assumptions.

What are the challenges of combining multiple AI reasoning paths?

Reconciliation is the hard part. Panels can agree on a direction while disagreeing entirely about why, which looks like consensus and is not. They can also converge for the wrong reason — shared training data producing correlated error rather than independent confirmation. The practical defences are roster diversity across vendors and clouds, and reporting directional consensus separately from within-panel convergence so a decisive answer sitting on top of genuine disagreement cannot be presented as settled.

Can wisdom of crowds fail? Examples.

Routinely, and in predictable ways. It requires genuinely independent judgments; the moment participants see each other's answers, you get an information cascade rather than aggregation. It fails when everyone shares a bias — averaging a thousand people who all learned the same wrong thing yields a confident wrong answer. Applied to AI, models trained on overlapping corpora are not independent estimators, which is why roster diversity matters more than roster size, and why a unanimous panel warrants more scrutiny than a split one, not less.

What are the limitations of multi-panel deliberation?

It costs more and takes longer, which is a real constraint rather than a footnote. It cannot manufacture evidence that does not exist — a thin brief debated by fifty models is still a thin brief. And it does not eliminate correlated error, only reduces it. The honest claim is narrower than it sounds: multi-panel deliberation makes disagreement visible and makes confident errors less likely. It does not make the output correct.

See the actual delivered report

The full PDF this run produced — case 2026-0078, generated on 3Dogs' AWS development environment.

Try this on your own question.

Free, no card. Bring a real decision — ideally one where you already know the answer — and see what the panel does with it.

Questions this case answers

How is Deep Mode different from a normal case?

A normal case runs one panel. Deep Mode runs several with fresh research between them, then reconciles across panels — used when the decision is large enough to justify the extra time and cost.

Can I watch it run?

Yes. The War Room streams the research and the debate live, including a running API-call counter, so you can see the work rather than wait for a verdict.

Try this on your own question.

Free, no card. Bring a real decision — ideally one where you already know the answer — and see what the panel does with it.

Related decision case studies

• AI That Refuses to Fabricate: The Deloitte Re-Run

  Most AI invents something when the evidence is thin. We ran a well-documented public AI failure through a multi-model panel to see whet

• Which Road-Safety Projects Should a State Fund?

  A 12-model AI panel ranked state DOT road-safety projects and produced the audit trail behind the ranking, including a Bayesian crash-m

• Can AI Read 500,000 Emails and Find Fraud?

  We read 45,320 real Enron executive emails blind - deduplicated from 517,401 - in about 2.5 hours for roughly $69, and surfaced LJM, Ra

• AI Platforms Reviewed by Rival AI Models

  We let Google's Gemini act as a hostile client, drive a full decision case, and then write its own review of the output. We published i