Consensus is not confidence: what agent agreement really tells you

When seven out of nine specialists call UP, it is tempting to read the market as 78% likely to rise. Agreement feels like evidence, and a simple vote is easy to understand. That is exactly why Probora publishes Base Consensus, a transparent count such as 7 / 9 UP, on every market.

But a vote count is not a probability, and treating it as one is one of the most common mistakes in combining signals. Two problems hide inside the count: the votes are not independent, and they are not equally informative.

Specialists look at different things: trend structure, momentum, order flow, derivatives positioning, large-participant activity. Yet many of those inputs move together. In a strong trend, the Trend, Momentum and Order Flow agents frequently agree for overlapping reasons. Seven agreeing agents may represent three or four genuinely independent pieces of evidence, sometimes fewer.

Counting correlated signals as if they were independent overstates confidence. The more the specialists share information, the less each additional vote adds, and the more a raw count exaggerates how sure the network should be.

Specialists also differ in skill, and their skill depends on context. An order-flow agent may be sharp on five-minute windows and nearly uninformative on daily ones. A mean-reversion agent may excel in range-bound commodities and struggle during crypto breakouts. A raw count gives each of them the same weight everywhere.

• Reliability varies by market and horizon. Probora tracks every agent separately for each asset × timeframe for exactly this reason.
• Reliability varies by regime. Volatility expansions, compressions and trending phases favor different specialists.
• Disagreement carries information. A lone dissent from a historically reliable specialist can matter more than several agreeing weak ones.

The Meta Agent is not a tenth vote. It takes the specialists' outputs as inputs and learns, from resolved history, how much their agreement is actually worth for this market, this timeframe and the current conditions. Then it publishes its own probability.

That is why the two numbers can diverge. A market may show Base Consensus 6 / 9 UP while Meta says UP 57%: agreement exists, but it comes from specialists that have been weak in this regime. Another market may show a narrow 5 / 9 UP while Meta says UP 66%, because the specialists that agree are the ones that have been reliable here. (Both are illustrative examples.)

If eligible specialists divide evenly (for example 4 / 8 when one specialist is not publishing), Base Consensus reports SPLIT. It is never rounded into a direction. A split vote is a useful signal in itself: the evidence is mixed, and the Meta probability will usually sit close to 50%.

1. Use Base Consensus to understand breadth: how many different lenses point the same way.
2. Use the Meta probability as the network's best estimate of the outcome.
3. Pay attention when they diverge. High agreement with a modest Meta probability suggests correlated or context-weak signals; modest agreement with a strong Meta probability suggests the right specialists agree.
4. Check Meta's calibration on the scope you care about before relying on its strongest calls.

Consensus measures agreement. Confidence should measure reliability. Probora publishes both, side by side, so you can see when they line up and, more usefully, when they don't.