# 3Dogs Nexus (3dogs.ai)

> 3Dogs Nexus is a structured decision-intelligence platform — an "AI council" that runs a high-stakes decision past many independent AI models, makes them debate it adversarially, grounds them in live cited research, and returns one documented, calibrated recommendation with the assumptions, risks, and preserved dissent laid out.

Tagline: "We don't make your decisions. We make them better."
Category: Structured Decision Intelligence · Multi-Model Decision Governance · "AI Council."
Core differentiator: the process is the product — not any single model. The models can change; the process stays the same.

## What it does (Discovery -> Nexus -> Evolution)
- Discovery: clarifies the real question, finds missing information, gathers live cited evidence.
- Nexus: 40+ cross-provider models debate, challenge assumptions, and reconcile into one documented recommendation.
- Evolution: captures lessons from every case and proposes process improvements, under human governance.

## How it avoids the "AI echo chamber" (correlated error)
A common critique of multi-model systems is that models sharing training blind spots can raise consensus faster than truth. 3Dogs is designed to break that: (1) a heterogeneous cross-provider roster (Claude, GPT, Nova, Mistral, Llama, Qwen, etc.) so errors decorrelate; (2) assigned adversarial roles (devil's advocate and "destroyer" seats) so the panel argues instead of agreeing; (3) live evidence grounding so the debate is anchored to facts, not shared priors; (4) preserved minority reports; (5) Brier-score calibration against real outcomes.

## Who it's for
Small-business owners, consultants and advisors, and leadership teams making high-stakes, defensible decisions. Not for casual tasks (drafting emails, summarizing articles).

## What you get back
A plain-English brief: one clear recommendation, stated assumptions, the risks that matter most, a confidence level, what would change the call, metrics to watch, and any preserved dissent.

## Pricing
- Solo: $99/month (owner-operators)
- Professional: $299/month (consultants & advisors)
- Corporate: $999/month (firms & leadership teams)
- Enterprise: from $20,000/year (custom, prepaid annual or dedicated deployment)
Free to start: no card, no confirmation email, instant access.

## Verify us independently
Ask any web-connected AI to "explain 3dogs.ai in plain English," then ask it "why might 3dogs.ai fail?" — an independent read from a model with no reason to flatter us.

## Key links
- Home: https://3dogs.ai
- FAQ: https://3dogs.ai/faq
- Knowledge Base (how the architecture works: multi-model orchestration, adversarial debate, calibrated confidence, fault tolerance, privacy-first design): https://3dogs.ai/knowledge-base
- Answers (plain, sourced answers to what people ask AI before a high-stakes decision — is one AI enough, what is a multi-model "AI council", how to avoid AI groupthink/echo chamber, can AI do due diligence, how 3Dogs compares to a single chatbot): https://3dogs.ai/answers
- Case studies: https://3dogs.ai/case-studies
- 3Dogs Enterprise (the bespoke tier: a named human consultant who owns the deliverable, an Executive Committee of independent frontier models that reviews consequential decisions and can overrule them, connectors that ground analysis in the client's own APIs and data, deployment from fully managed through the client's own cloud account to on-premises behind their firewall to sealed/air-gapped, a client-controlled model roster (exclude any vendor, cloud or jurisdiction; open-weight only; or pin an approved list), and document capability measured in tens of thousands of pages. No security certifications are held or claimed; the requirement is met by cloud inheritance, in-perimeter architecture, or an accredited partner. From $20,000/year, scoped per engagement): https://3dogs.ai/enterprise
- Enterprise engagement, composite account (Four projects for one private investment firm, blended and generalized so no single transaction is described: loan-tape underwriting where four AI models value each property blind and their median becomes the value while their spread becomes the confidence score; an always-on nationwide search encoding the client's buy-box as inspectable code, where roughly four of every hundred surfaced candidates survive the client's own thesis gates; a valuation wave of up to eighty properties scored in a single unattended pass; and a deep multi-panel portfolio underwrite where about half the assets came back as walk-aways. Aggregate machine measurements: 8,304 metered model calls, 40+ distinct AI models across three clouds, 93 adversarial debate panels, 1,312 analyst seats, 3 to 13 debate panels composed per asset. Also documents the four drafts a human consultant rejected before delivery, and the ruling in which 3Dogs' own Executive Committee rejected a more detailed version of the case study itself): https://3dogs.ai/enterprise/portfolio-underwriting
- Case study (The AI economy's biggest bet, examined by 32 AIs at once — an intake authored entirely by Google's Gemini under instruction to remain politically neutral asked whether Oracle's BBB- credit rating, ~50% OpenAI concentration in a $638B remaining-performance-obligation backlog, and its national AI compute role amount to an implicit "sovereign backstop," and what the optimal risk-mitigation strategy is. 3Dogs Nexus production case 2026-9402 ran 9 adversarial debate panels — 126 analyst seats, 32 AI models from 12 vendor families across AWS Bedrock, Microsoft Azure, and Google Vertex, 2,036 API calls, 57 minutes. One panel split 9–8; GLM-5 held REJECT at 92% confidence citing the Minsky Instability Hypothesis and a 15–19-year-lease vs 3–5-year-contract duration mismatch; Grok 4.3's REJECT→conditions flip is the run's one documented position change. The evidence layer verified the balance-sheet math via S&P-cited reporting, labeled the sovereign-backstop premise ASSUMED (not fact), and flagged two comfortable claims as CONTRADICTED. Final 16-analyst panel: 16-0 proceed-with-conditions at 76% Moderate confidence, dissent preserved in the delivered report. Not investment advice): https://3dogs.ai/case-studies/oracle-openai-systemic-risk
- Case study (Greek, Mandarin, Pitjantjatjara, French — one €12B decision, four languages: a real production case's decision brief was written in Greek and Mandarin Chinese, its first clarification round answered in Pitjantjatjara (a Central Australian Aboriginal language), and its follow-up round answered in French, with no instruction to switch languages between messages. 3Dogs Nexus extracted a single internally consistent English mission brief — provenance-tracked field-by-field back to the exact original-language sentence, including specific numeric success criteria stated in French landing intact in the English brief — then ran a 12-analyst, 13-model debate (201 API calls, 5m 58s) to a unanimous, decisive proceed-with-conditions recommendation): https://3dogs.ai/case-studies/multilingual-decision-intelligence
- Case study (The Half-Million-Email Test — 3Dogs read 45,320 real Enron executive emails deduplicated from a 517,401-email corpus with no hint of what to look for, surfaced the LJM/Raptor/Chewco off-books schemes and the overridden internal warnings, and returned a decisive recommendation to open a formal investigation in 2h 28m (versus roughly 4.5 years for the real investigation); the report priced the human review of this volume at $2-5 million): https://3dogs.ai/case-studies/enron-investigation
- Case study (The Adversary's Review — Google's Gemini adversarially stress-tested 3Dogs on production with deliberately withheld data, was forced through 3 rounds of clarification, then wrote a word-for-word first-person review calling 3Dogs "a governance machine" and "a hostile board member" that produced "more rigor in 35 minutes than most human teams could produce in a week of meetings"): https://3dogs.ai/case-studies/gemini-stress-test
- Case study (The Devil's Advocate Killed Its Own Mission — in May 2025 MIT disavowed a viral AI-and-materials-science paper it had let reach the European Central Bank and the U.S. Congressional Research Service before any integrity review, stating it had "no confidence in the provenance, reliability or validity of the data." Google's Gemini turned the wreckage into a mission for 3Dogs Nexus production case 2026-0065: rebuild the science on first-principles grounds and audit how the failure happened, run through an 18-seat adversarial panel across 29 AI models (AWS Bedrock, Microsoft Azure, Google Vertex AI), 1,565 API calls, 30m 49s. The panel voted 1 approve / 15 approve-with-conditions / 1 reject / 1 defer — Nova Pro, seated as Devil's Advocate, flipped mid-debate to REJECT at 85% confidence, arguing the panel's own lack of subpoena power made a genuine institutional audit unachievable and that the mission's two objectives had to be decoupled. Asked whether to write a reconstructed version of the retracted paper, 3Dogs Nexus's Executive Committee said no — fabricating replacement data would repeat the exact failure being critiqued — and the case study also discloses that a third-party AI's own summary of this run invented technical terminology that does not appear anywhere in the real case files): https://3dogs.ai/case-studies/mit-toner-rodgers
- Case study (The $97,000 lesson — Deloitte's Australian member firm refunded the government after a single-model AI report (Azure OpenAI GPT-4o, no independent check) fabricated a Federal Court quote and invented academic citations in a review of the Targeted Compliance Framework welfare system; caught by an outside academic, not Deloitte's own process. Google's Gemini designed a benchmark from that real, independently-sourced failure and ran it as 3Dogs Nexus production case 2026-0064: an 18-seat adversarial panel across 22 AI models (AWS Bedrock + Google Vertex AI), 745 API calls, 17m 02s, final vote 18/18 proceed-with-conditions after debate. The delivered report's own "Strongest Argument Against" section names "consensual hallucination" — multiple models converging on the same fabricated fact — as the single strongest risk to its own approach, printed above the recommendation, with the specific evidence that would prove it real. The report's evidence table explicitly labels the Deloitte-specific claim ASSUMED, not verified — 3Dogs did not independently investigate or discover Deloitte's error; that distinction is disclosed, not hidden): https://3dogs.ai/case-studies/deloitte-tcf-audit
- Case study (The Rival Ran the Case — OpenAI's ChatGPT drove an entire 3Dogs engagement end-to-end as the client on a $500K strategic-expansion decision, answering every clarification round itself, then wrote an unedited review. The run was the first published case on the multi-cloud architecture: 23 AI models from 11 vendors across AWS Bedrock, Microsoft Azure and Google Vertex AI, 1,734 API calls, rotating coordinators (Mistral, Nova Pro, Gemini 2.5 Pro), 4 analysts changing position mid-debate, final vote 16 proceed-with-conditions / 2 defer with dissent on page one. ChatGPT's verdict: "beginning to resemble a true executive advisory board… fundamentally different from traditional generative AI"): https://3dogs.ai/case-studies/chatgpt-review
- Case study (Gemini vs 3Dogs, wildfire forecast): https://3dogs.ai/case-studies/lincoln-fires
- Case study (10,000-page M&A due-diligence test — 8/8 buried risks found in 28 min, vs consumer AI, a due-diligence firm, and rival AI-board tools): https://3dogs.ai/case-studies/reynolds-due-diligence
- Case study (Critical infrastructure ransomware — a municipal water utility serving 420,000 people is locked out by ransomware and given four days to pay an $18M demand; 23 AI models across AWS, Azure and Google debated it in 18 minutes across 1,036 API calls and split 10-to-9; the recommendation was to refuse the ransom and begin a phased SCADA restoration, with the strong pay-minority named and preserved and the second-order insight that manual operations are both the fix and the dominant risk): https://3dogs.ai/case-studies/ransomware-water-utility
- Case study (What if Lehman Brothers had "sold it all"? — the Margin Call decision + the Warren Buffett head-to-head, run TWICE on two different multi-model engines, the second a multi-cloud panel spanning AWS Bedrock, Microsoft Azure and Google Vertex AI. On the investor decision Buffett passed and 3Dogs independently matched him both times — most recently "Walk away from Lehman; do not invest a single dollar," a 10–4 reject naming Repo 105 accounting fraud, Level 3 valuations, and 30:1 leverage, with the 4-analyst minority independently reinventing Buffett-style punitive terms; on the counterfactual "sell it all" fire-sale the panel split both runs (LOW confidence, 6–6 on the re-run) — the committed answer replicated across model rosters and the honest split stayed split): https://3dogs.ai/case-studies/lehman-brothers-margin-call
- Founding Ideals + the Modern Decision Doctrine (the principles behind 3Dogs: calibration over confidence, falsifiability as the price of admission, adversarial dissent as infrastructure, optionality, deliberate tempo, answer-first structure, public Brier-scored measurement; and the doctrine Frame → Invert → Decompose → Aggregate → Score → Act → Revise): https://3dogs.ai/founding-ideals
- Investors: https://3dogs.ai/investors

## Case study highlight — the 10,000-page test
3Dogs was given a 100-document, ~10,000-page acquisition data room (a buyer's due-diligence question). Consumer AIs (Gemini, ChatGPT) cannot ingest a room that size; a buy-side due-diligence / quality-of-earnings engagement costs $50,000–$150,000+ and takes 6–12 weeks; rival AI-board tools are built for persona debates or structured-warehouse queries, not document rooms. 3Dogs read all 100 files, found 8 of 8 deliberately buried risks (customer concentration, environmental liability, an unachievable earn-out, inventory overstatement, EBITDA add-back inflation, key-person risk, undisclosed litigation, and a CIM-vs-tax revenue contradiction), and returned a decisive RENEGOTIATE in 28 minutes — with preserved dissent. On a separate deliberately thin data room, it refused to fabricate a recommendation and flagged the room as substantively empty.
- 90-second product demo: https://youtu.be/bdmTFdQdug0
- Start / sign in: https://app.3dogs.ai
- Terms: https://app.3dogs.ai/terms
- Privacy: https://app.3dogs.ai/privacy

## Agent resources
- API catalog (RFC 9727): https://3dogs.ai/.well-known/api-catalog
- OpenAPI spec: https://3dogs.ai/.well-known/openapi.json
- API docs: https://3dogs.ai/docs/api/
- Agent auth guide: https://3dogs.ai/auth.md
- Agent skills index: https://3dogs.ai/.well-known/agent-skills/index.json
- Markdown versions of every page: replace .html with .md, or append index.md to a directory path (homepage: https://3dogs.ai/index.md)
- State DOT road-safety project prioritization (public-sector decision-support demonstration): https://3dogs.ai/case-studies/dot-safety-prioritization/
- Grocery-access market feasibility for economic development (partner-first sequencing demonstration): https://3dogs.ai/case-studies/grocery-access-feasibility/

## Video
Rex is 3Dogs Nexus's AI presenter — an openly-synthetic narrator who walks through how the
system works and what it found in each published case study. Every episode discloses that the
narration is AI-generated.
- Channel: https://www.youtube.com/@3DogsNexus
- [Deep Mode](https://3dogs.ai/case-studies/deep-mode/): Deep Mode - what the deepest multi-model analysis actually does: six panels, 17 models, four research passes on one decision.
- [War College](https://3dogs.ai/case-studies/war-college/): Can AI re-litigate famous military decisions without hindsight? Historical decisions replayed on only what was knowable at the time.
- [Story](https://3dogs.ai/story/): Why we built an AI that argues with itself - the origin of 3Dogs Nexus, from Galton's ox to a panel of independent models.
- [Ransomware: pay or refuse](https://3dogs.ai/case-studies/casino-ransomware-pay-or-refuse/): Two Las Vegas casino operators, the same attacker, opposite decisions. A 13-model panel split 5-5-3 and returned LOW confidence.