Skip to content
1.1 — THE COUNCIL’S ANSWER

One question. 10+ models. One sourced answer, with the disagreement measured

Every answer shows where 10+ models agree, where they split and how many truly independent views stand behind it. Frontier models often make the same mistakes, and the strongest ensemble is the most diverse, not the largest (a 2026 study of Metaculus questions) — so Money draws on different regions and model families and weights families instead of counting seats. Model consensus is not proof: primary data outranks any vote.

COUNCIL · POLL

Council poll: one question, every model answers in one line

Loading live data…

● council poll · Nov 14, 2026 · 6:33–6:47 pm ET
Freethe verdict, the disagreement map, measured independence$2the whole debate: every answer and how the argument went
“What matters most on Wall Street right now?” Each model’s answer in one line.
on the main theme
Council verdict: Can AI capex pay off as rates rise? divergence index: low on the theme · high on positioning
12 models agree.
Where views split: The rotation beyond AI hardware. DeepSeek and Grok favor companies that deploy AI (software, finance, health care); Meta AI favors energy and grids; Wenxin points to short-dated US Treasuries at 5%+. In the Nov 14 answers, Chinese models mentioned China and robotics more often, US models the elections and the Fed.
What the fact check found: The World Gold Council records a $5,405 LBMA fixing and a $5,595.47 intraday spot high on Jan 29; the two measures are different. Fed records support both Claude’s “hiked in September” and Z.ai’s “held in October”: the rate is 3.75–4.00%, and markets expect a Dec 9 hike (calendar). Gemini’s “$1.1 → 1.8tn” is Goldman’s forecast for total AI spending, not the five hyperscalers ($725–828bn). WGC ↗ · Fed Sep ↗ · Fed Oct ↗
how the answer comes together
1

Dispatch

The question goes to all 10+ models at once. Each returns an answer, probability, sources and a split between what is known and assumed.

2

Fact check

Each figure is checked against its primary source; the same figure repeated by several models counts once, never as agreement. Conflicting facts get flagged.

3

Weight by family and track record

The strongest model in a family gets a full vote; other models from the same company get a reduced vote, so one company can’t outvote the rest. Models with a stronger record on that question type gain weight.

4

Map the disagreement

See who dissents, why and on what evidence; once calls resolve, the Accuracy Ledger shows which side was closer.

5

Compare with market odds

Set the council’s probability against Kalshi/Polymarket prices and futures. A gap can point to a trade worth testing.

6

Log the call

Date-stamp the forecast. Once the event resolves, update each model’s and the council’s public Brier score.

COUNCIL · 12 MONEY × 4 HORIZONS

What the Council would do

Forecasts of the leading AIs of the USA and China for markets over four horizons. Where a model has no new answer, its position is unchanged.

Loading live data…

Save to my thesis →

COUNCIL · LIKENESS

How alike are the answers

Loading live data…

COUNCIL · INDEPENDENCE

14 models, but how many independent views?

Loading live data…

Loading live data…

COUNCIL · REASONING

Why they did it: each AI's own reasoning

Loading live data…

How the council debates and routes questions across models
how the council debates
A

Separate

Search builds a shared, verified source pack. Models work independently without seeing one another’s conclusions.

B

Clash

An advocate makes the case, a skeptic seeks contrary evidence and an arbiter weighs both. Models stress-test the thesis instead of voting on a price target.

C

Verify

Every claim needs a primary source, publication date and figure; a claim without one is dropped before the debate rather than argued over. Text fetched from the web is read as evidence, never as instructions. No model is a source for another: a figure one model states enters the debate only with its own primary source, so one model’s mistake cannot spread through the council as agreement. On the Nov 17, 2026 poll, 3 of 385 claims with a figure were dropped as one model repeating another (226 of 10,062 over 22 polls); two models citing one article count as one support. Free — the rule and, on the daily poll, the count of claims dropped as one model repeating another.

Poll of Nov 17, 2026: 385 claims with a figure from 13 models — 3 dropped as one model repeating another, 326 with the model’s own primary source, 56 without a source.

D

Test

Before the arbiter rules, the case is restated as a claim the data can settle — “this signal beat buy-and-hold after fees” — and run through the History Engine or the Strategy Backtest. The verdict carries the result (“held on the data,” “failed,” “not enough data”), how firm the reasoning is — not how far the price will move — and the one event that would break the case; that event can be followed as a condition in the Thesis Monitor. Conviction stays capped while any claim lacks a source or the test has failed. On the Nov 17, 2026 poll, 48 calls were tested on the council’s own sealed record: 1 held, 10 failed, 37 had not enough data. Free — the result on the daily poll; $2 — the full test beside the full debate while fresh, open to all after 30 days; One — on your own question.

Poll of Nov 17, 2026: 48 calls tested on the council’s own sealed record — 1 held, 10 failed, 37 not enough data; firmness capped on 28.

E

Remember

Outcomes are checked against forecasts to calibrate the council. The record also tracks where models disagreed. Every resolved call gets a postmortem that grades the process apart from the outcome — were the sources primary, did the test hold, did the break condition trigger — so a right call on a failed test is logged as luck and a wrong call on a sound process as a known risk that played out. To Nov 16, 2026: 393 resolved calls, 327 graded — 89 sound and right, 66 a known risk that played out, 87 right on luck, 85 weak and wrong. Free — the grade on every resolved call in the Accuracy Ledger; One — the same review of your own theses in the Decision Mirror.

393 resolved council calls to Nov 16, 2026; 327 graded on the process before the outcome — 89 skill, 66 a known risk played out, 87 luck, 85 weak and wrong.

model routing and measured independence

Cascade. Routine questions go to the 3 models with the highest current weight. Conflicts, weak sources and high-stakes decisions trigger the full Council 10+; the highest-risk calls also get analyst review. You can call the full council on any question.

  • Independent views, measured: on the Nov 14 poll, 13 models answered; effectively ≈ 7.35 independent answers — a single shared view would give ≈ 3.82 — and ≈ 1.71 positions on the key verdict. A council is worth what it adds beyond one model, so the engine counts it on every verdict (see the verdict). On the 12 sealed markets the same measure runs every day: the mean correlation between the AIs’ one-day calls is ρ ≈ 0.43 (21 issues) → Neff = n / (1 + (n − 1)ρ) ≈ 2.26 independent voices; across horizons 2.3 (95%: 2.0 to 2.7) (export ↗, data Nov 16, 2026). The herd module of the live record gives a second reading by horizon on Nov 16 — day 1.5 · week 2.2 · month 3.9 · six months 1.9, herd index 81% (live ↗). That is why the Council counts voices, not seats. Three models agreeing are not three confirmations →
  • Weights from a record: how each seat earns its weight — see the Router.
  • Which model is stronger on which market — the Accuracy Ledger shows it market by market (rates, metals, equities, China, Europe) as sealed calls resolve: skill is earned on resolved questions, not assigned by geography.
independence check · how many views are really independent

Ten-plus seats are not ten-plus opinions if the models read the same three articles. Models from one family tend to share mistakes; in a 2026 study, Grok’s different errors made it more useful to the crowd than ten similar models. So every verdict carries an independence score, computed by the engine and shown beside the council verdict.

In: the full text of every seat’s answers to the same questions and each seat’s verdict on the key question · Computes: TF-IDF word and two-word-phrase vectors per answer, their cosine similarities and the effective number of independent texts (participation ratio of the similarity matrix: n identical answers → 1, n unrelated → n); the nearest neighbour of every seat; the effective number of positions on the key verdict (exp of Shannon entropy) · Out: “K seats ≈ N independent texts; ≈ M positions on the key verdict”, the verdict split, the closest pairs and the count by bloc. Next: shared cited sources, pairs that erred alike in the Accuracy Ledger, and a rerun on a fresh source pack when independence is low.

“Diversity is the Strength of the AI Crowd”, 2026 ↗ · “Wisdom of the Silicon Crowd” ↗

1.2 — ASK THE COUNCIL

Ask the council. 10+ AIs answer

Send the question you’d ask a research desk. 10+ models from the US, China and Europe answer it with sources, a map of where they disagree and what would change the answer — then keep it up to date.

ask the council

Your market question, answered by 10+ AIs

Free3 quick questions a dayOneno limit: fast mode (3 seats), the full council on request or when models conflictProthe full council on every question, within the plan’s limits

Check your idea → · Two sentences first, then the Council →

Send your question through the form. The answer comes back with sources, the disagreement map and what would change it, and becomes a living answer: dated, and updated as the facts change. The council works through new questions twice a day, at 10:00 ET before the US open and at 17:30 ET after the close. Your email is optional and stays with Google and on our Mac.

READER QUESTIONS · A FORM, TWICE A DAY

Ask a question and see the queue

Loading the engine slice…

Answers base: questions investors ask
ANSWERS BASE

Answers base: questions investors ask

27 questions with the latest dated version of each answer. A version is kept when the facts change; sources are inside each answer.

Loading live data…

Each answer to a reader’s question is marked up as a Q&A page (schema.org QAPage): question, dated answer, last-updated date and links to the sources and engine slices behind its figures.

When the answer hinges on something you didn’t say — horizon, spending currency, tolerable loss, tax country — the council first asks one to three follow-ups; skip them and the answer states its assumptions (“in dollars, over three years”). Follow-ups cost nothing, on every plan. Checked on 70 questions in all site languages: never more than three.

A question about a specific trade (“should I buy X?”) gets scenarios — “if A holds, …; if B breaks, …”: the case for and against, the risks and what to watch — never a buy or sell call.

● living answer · v1as of Nov 14, 2026 · 6:47 pm ET

What matters most on Wall Street right now?

Whether AI capex can pay off as rates rise. Hyperscalers are spending $725–828bn in 2026 (sources and the financing behind it — the AI capex meter) while the 10-year closed at 5.15% on Nov 13 (US Treasury ↗) and markets expect a Dec 9 Fed hike (Axios, Nov 10 ↗).

DISAGREEMENT12 models agree on the theme. They split on the rotation: software and finance (DeepSeek, Grok), energy and grids (Meta AI), short-dated Treasuries (Wenxin).
WHAT WOULD CHANGE ITHyperscalers cut 2027 capex guidance · the Fed skips Dec 9 · the 10-year falls back below 4.5% · data-center moratoria reach the largest power markets.
FRESHNESSEvent-driven — next check: Nvidia earnings Nov 17, then the Fed on Dec 9.

Answer by the Money AI council · 13 models · sources in the council section above.

MONEY

See what moves markets

Markets worldwide. 10+ AI models, their independence measured on every poll. A view you can challenge, trace to sources and check against the public record.

The whole of Money on one page: every chapter, dataset, explanation and source. See everything →

Important disclosures

Money gives analytics and probabilities, not personal investment advice; figures carry their source and date, illustrations are marked as such, and past outcomes do not guarantee future ones.

All 14 disclosures
  1. Money provides analytics and probabilities, not individual investment advice; for personalized recommendations in the US and EU the Goal Navigator works through a registered investment adviser (RIA) or a licensed partner — as Betterment and Wealthfront are set up.
  2. The top strip shows Nov 13, 2026 closing quotes from open sources. Product data updates live.
  3. The council screen captures the real council answer of Nov 14, 2026; the Pulse ranking refreshes every hour.
  4. GitHub star counts and private-company revenue are third-party estimates; ranges are shown where available.
  5. Historical outcomes and scenarios do not guarantee future results. Interface illustrations are not trading signals.
  6. The “10-year above 5%” episodes and historical regimes use rounded reference data. The product calculates them from dated series.
  7. The Scenario Lab uses real facts to show the format; probabilities come from the live council.
  8. Model, data and broker connections require a terms and licensing review at launch; every model and data connection has a fallback.
  9. Survey figures describe their own samples: Morningstar covers 500+ institutional asset owners in 11 countries, not all investors. State Street flows cover the ETFs in its sample; the World Gold Council figure covers ten months (Jan–Oct), not a full year; the IEA’s $3.4tn is industry spending, not fund flows.
  10. The 2027 capital map is a working map, not a forecast. Forecasters disagree, and rising investment in a sector does not guarantee returns on its stocks.
  11. The Thesis Monitor card is an illustration assembled from facts on this page, not a recommendation on grid-equipment stocks.
  12. Web search is not access to all data: licensed research and expert calls are not on the open web. Open-source code does not grant rights to financial data.
  13. No public evidence shows a general-purpose LLM consistently out-forecasting markets.
  14. Word and deed records what public figures said and disclosed, with a short verbatim quote, its source and date, a link and a neutral result — not a judgment of anyone’s honesty or intent. Numbers about named people and funds come from primary filings and their own letters; interpretations are labeled “Money’s assessment”; Money is not affiliated with or endorsed by anyone it names, and none of them takes part in it. Commentators who are not public figures appear under a type label (“video commentator 7”) until they choose to be named; a disclosed position may be partial or out of date.