btcpredictor.io // cognitive architecture

The Council of AI

Six artificial minds. One deliberation.
No single point of failure.

Scroll to explore
Chapter I

In the beginning
was the prompt

There was a time when artificial intelligence answered like a bored clerk. Neutral. Flat. Without personality. Then someone wrote two words that changed everything.

> "Act as a senior financial trader with 20 years of experience in derivatives markets."
// The response changed. Not the words — the structure of thought.
>

“Act as” — two words that opened a door. The discovery was simple and profound: a language model doesn't just change its tone when you assign it an identity. It changes the structure of its reasoning. An “institutional trader” doesn't use the same heuristics as a “generic assistant”. It weighs risks differently. It notices different patterns. It has different biases.

It was the beginning of prompt engineering as a discipline. Not programming. Not configuration. Something closer to directing an actor: give it a role, and the intelligence reorganizes itself around that role.

An LLM is like a rough diamond with infinite facets. The prompt doesn't create the intelligence — it chooses which facet to illuminate. — Lesson learned, session 1
Chapter II

The limitation of the
single expert

So we built our first predictor. A single AI model — Claude Sonnet — with a 3,000-word prompt. Fifteen rules. Twelve data sources. We christened it the Trader.

🤖
A single model. A single mind. A single bias.

And for a while, it worked. Then the data spoke: across 613 predictions, the Trader called “DOWN” 74% of the time. Like an analyst who lived through the 2008 crash and has seen bear markets everywhere since. Not because the data suggested it — because its “personality” led it there.

The problem wasn't the model. The problem was the single cognitive point of failure. A brain, however brilliant, has blind spots. It has biases it cannot see because they are part of its very perceptual structure.

No serious hedge fund lets a single trader decide. No board of directors has a single member. No court has a single judge. Why should we do it with AI?

Chapter III

From the single
to the collective

The inspiration doesn't come from technology. It comes from history.

In 1950, the RAND Corporation invented the Delphi Method: gather independent experts, have them vote anonymously, share reasoning, repeat. Collective wisdom systematically outperformed the single expert. Not because the group was smarter — but because informed disagreement produces better decisions than blind consensus.

This is the principle behind the AI Council. Not an “ensemble” — that's just voting. Not a “mixture of experts” — that's routing. The Council is something different:

Six artificial intelligences, each with a specific expertise, each with different inputs, that argue their position before converging on a decision. Or on no decision — because sometimes courage means not acting.
🔧Technician
🧠Sentiment
On-Chain
📊Quant
😈Contrarian
🔮The Oracle
Chapter IV

The six minds

Every member of the Council has a precise role, a dedicated AI model and — crucially — different inputs. They don't all see the same data. Just like in a real board of directors, the CFO doesn't read the same reports as the CTO.

30%
🔧
The Technician
Claude Sonnet 4.6
RSI, EMA, MACD, Bollinger, candlestick, order book. Reads the chart like a surgeon reads a CT scan.
15%
🧠
The Sentiment
Gemini 2.0 Flash
Fear & Greed, funding rate, long/short ratio, news. Feels the pulse of the crowd and knows when it's lying.
15%
The On-Chain
GPT-4o-mini
SOPR, MVRV, whale alerts, exchange netflow. Sees what the crowd cannot see: the movements beneath the surface.
25%
📊
The Quant
XGBoost ×3
No language, only numbers. Three models: trend-follower, mean-reversion, regime-classifier. Cold, statistical, relentless.
5%
😈
The Contrarian
Claude Haiku 4.5
The most uncomfortable and most important role. When everyone says UP, it must explain why it could be DOWN. The devil's advocate.
10%
🔮
The Oracle
Claude Opus 4.6
Not a dictator. A synthesizer. It intervenes only when there is strong disagreement and weighs the arguments, not the votes. Sees everything. Decides everything.

The weights aren't fixed. They self-calibrate based on each member's track record. In a trending market, the Technician weighs more. In a fear-dominated market, the Sentiment leads. The system learns who is right, and when.

Chapter V

The deliberation

The Council doesn't just vote. It deliberates. The difference is enormous: voting is counting raised hands. Deliberating is listening to reasoning, weighing arguments, embracing dissent.

Round 1 — Independent Vote
Each member speaks without knowing what the others say
Technician, Sentiment, On-Chain and Quant analyze their own data in parallel. Each produces: direction, confidence, reasoning. No one sees the others' votes. Total independence.
Round 2 — The devil's advocate
The Contrarian challenges the consensus
If disagreement is low (>70% agreement), it proceeds. If there is tension, the Contrarian enters the scene. It receives all the votes and presents the strongest counter-argument against the majority position. Not out of spite — for stress-testing.
Round 3 — Synthesis
The Oracle synthesizes or declares SKIP
The Oracle sees everything: votes, reasoning, counter-argument. It has three options: confirm the majority, overturn if the Contrarian is right, or SKIP — no trade. Because sometimes the bravest decision is not to make one.
Deliberation simulation — cycle #644
🔧
Technician
UP
0.68
🧠
Sentiment
UP
0.58
On-Chain
DOWN
0.54
📊
Quant
UP
0.71
😈
Contrarian
DOWN
0.47
🔮
The Oracle
UP
0.62
Council Decision
↑ UP
Weighted confidence
0.643 — agreement 78%
Chapter VI

Evolution

The Council is not static. It learns.

After every trade, the system records who was right and who was wrong. The weights shift. The Technician who correctly predicted 65% of trends gains influence during trending phases. The Sentiment who read the panic better than the others gains weight during fear phases.

Technician
Sentiment
On-Chain
Quant
Contrarian
The Oracle

But the real evolution isn't in the weights. It's in the self-improvement loop: every week the system analyzes its own mistakes, identifies recurring patterns and updates the councillors' prompts. It's not a human who decides what to change. It's the Council itself reflecting on itself.

The difference between an intelligent system and a system that learns is the ability to look in the mirror and not like what it sees.

And when disagreement is too high? When the Council can't converge? The answer is the wisest of all: SKIP. Don't trade. Wait. Because the market will still be there tomorrow, but lost capital doesn't come back.

The future isn't predicted.
It's deliberated.

Six minds. A war table. A collective decision better than any individual decision — not because the individuals are perfect, but because informed disagreement is the engine of wisdom.

Watch the Council in action →
btcpredictor.io — the ai council — cognitive architecture v1 — 2026