If you trade, you already know the pattern. You pull a filing, run a screen, write up an event, then do a version of it again in ChatGPT, again in Claude, maybe again in Perplexity. The notes come out close enough that you could have swapped them. The labs still bill you separately.
Collective Intelligence is a seat on a small research desk. The repeated pull is run once and split. Your own list is not. You get 15 private runs a month, on a desk that matches what you actually trade, and the write-up keeps the argument instead of sanding it down to one view.
Founding group. We send a sample before we bill. This is research sharing, not advice and not a signal.
Groups stay small. Pick the desks you want. We write back with a sample page before there is an invoice.
Most of the AI spend on a trading desk is not a secret. It is the same screen, the same monitor, the same first pass on a print. Buying another subscription does not make that memo different. Paying each lab so you can line the models up and compare them is a second bill on top.
Plus or Pro on two or three models, a data feed, a screener, and a prompt that is basically the one the person next to you used. OpenRouter, Poe, or a short script can ask several models cheaply. They will not keep a red-team pass, a source list, or a run on your private names. You still have to stitch the answers together.
The labs, for tokens you and someone else both spent on the same pull. We run that pull once, on the desk you joined, and leave your own question off the shared page.
A note that the whole room trades is not much of an advantage. The seat is set up so the common work is common, and the book is not.
Fifteen a month, on your prompt or your list. The result is not added to the shared binder. The point is that you should not need a Claude seat and a ChatGPT seat just to ask about your own names.
You do it in one of two places. A private page on the member desk, or a command in your own Telegram or Discord thread: /run and the name. /flash if it cannot wait. Not a shared channel, and not an email you paste a prompt into. The reply is the page.
Semis, energy and commodities, event-driven, FX, small-cap. You sit where the repeated work actually is. A vote on mega-cap tech does not set the menu for an FX book.
What gets shared is the ordinary layer: a filing extract, a screen, an event log. That is not a trade ticket. A private run is not circulated, so nobody in the room can trade a question they were never sent.
The desk does not take the other side, and it does not tell you the size. It shares the research. It is not a signal service, and it is not acting as your adviser.
Models trained on the same web often land on the same cautious line, and they miss the same thing. If we made them agree, both of those problems would disappear into the prose. The page is the argument, with one model whose only job is to pick at it.
What has to be true, and which filing, table, or official print it depends on.
The other reading, including the one a single model would have smoothed out.
The variable they actually disagreed on. That is usually the only part worth your time.
One agent checks liquidity, a misread filing, a regulatory tail, a valuation trap. It does not get a vote that averages the others.
There are two speeds. You pick. The morning note is not the only way in.
Under a minute. Two quick models, one pass, for a print, an earnings line, or a move on the tape. The disagreement is marked. It is not argued out. Same door as the private run: the desk, or /flash in your thread.
Several models, a red team, citations. For a sector note or a screen you were going to run on Sunday anyway.
A threshold or a screen hit goes out on Telegram, Discord, or a webhook. You should not have to log in to find out the 7am note exists.
CSV or JSON next to the note, so a systematic book can take the extract without retyping it.
This is a sample of the energy desk, written so you can judge the prose. It is not an old call, not a backtest, and not a claim that anyone made money from it. The figures are the shape of the weekly table a real note would cite. We don’t bill a founding seat until we’ve sent a current sample.
Sample · Energy desk · overnight · private run, not posted
Models on the run: a fast reader, a long reader, a red team. Sources under each line. No trade and no size.
Long. Commercial crude stocks fell 3.1 million barrels on the week, against a build the tape had already priced, and refinery runs held above 16.5 million barrels a day. For that to be bullish, the draw has to be demand, not a gap in the timing of imports. It hangs on the EIA Weekly Petroleum Status Report, Table 1, week ending 18 September 2026, not on a recap.
Short. The same table shows gasoline stocks up 1.4 million barrels and distillate stocks up 0.8 million. A crude draw with product builds is a refinery story. The short reading is that Gulf spare capacity is being over-read off a single crude line.
Where they split. Not on the headline draw. They split on whether the product builds reverse it inside two weeks. The rest of the first answers were the same paragraph.
Red team. Cushing stocks are not the Gulf, and a Cushing draw does not travel into a Brent crack. Liquidity in the front WTI contract is fine. Liquidity in the single-name refiners someone might hang this note on is not. The note only says that if the red team is allowed to refuse the average.
| Line | Week | Source |
|---|---|---|
| Commercial crude | −3.1 mb | EIA WPSR Table 1 |
| Gasoline | +1.4 mb | EIA WPSR Table 1 |
| Distillate | +0.8 mb | EIA WPSR Table 1 |
| Refinery runs | 16.6 mb/d | EIA WPSR Table 2 |
Sample only. The stock changes above show the table, they are not a quote of the live print. A note we bill for links the primary table. This page does not.
A month, pooled against APIs, filings, and compute. That is the invoice.
The standing work for that book, plus your private allowance. Not twelve people asking the same 7am question in public.
Shared extracts in the room. Private runs only to you, on the desk or in your thread. You can vote the shared menu and drop a prompt that is noise.
For people who want the repeated work done once, and don’t want the room to see the question. Sample page first. The trade is still yours.
Pooled against the live stack. The private runs are part of the seat, not an extra you run on your own keys. The public site is collectiveiq.kumijo.org.
Not investment advice and not a signal. Not a licensed adviser and not an MPF intermediary. No track record is claimed here. You remain responsible for the decision.
They should not be reading your question. The shared output is the ordinary layer: a filing extract, a screen, an event log. The 15 private runs stay off the binder. The room cannot trade a prompt it never received.
Two doors, same engine. A private page on the member desk, or a command in your own Telegram or Discord thread: /run and the name, or /flash if it cannot wait. Not a shared channel. Not an email you paste a prompt into. The reply is the page.
We don’t average them into a safe line. The page is the long side, the short side, the variable they split on, and a red team that does not get a vote. Agreement is written down only where it actually happened.
The overnight note is overnight on purpose. A tape move uses the fast check: two quick models, one pass, under a minute, with the split marked. Alerts go to Telegram, Discord, or a webhook.
Not for the shared pull, and not for the 15 private runs. A question past that allowance is on your own keys. The seat is meant to replace the overlapping briefs, not every chat you will ever have.
You are not in a general pool. Desks are semis, energy, event-driven, FX, and small-cap. The menu you vote on is the desk you joined.
Those are meters. They are useful, and cheaper for a one-off. They don’t hold a red team, a cited page, a private watchlist allowance, or a desk that already ran Sunday’s screen. Use them for the odd prompt. The seat is the work that gets repeated.
No. Nothing on this page should be read as one. The sample is a format sample. Before you are billed you get a current sample, the source list, and the model names. We are not inventing advisers.
No. It is research sharing. You still place the trade.