A single model can sound confident and still be guessing. MarketXED does not publish a gut-feel ticker from one score. A multi-agent committee scores the same setup from different lenses, then the app shows one directional verdict plus conviction and agreement you can read before you act. This is decision support, not financial advice.
What each specialist looks at
The roster is a set of named agents, not a marketing slogan. Typical seats include Wyckoff structure, Bayesian updating, momentum and RSI, mean reversion, volume-flow (OBV), Sentiment-X, a postmortem helper that remembers failure patterns, ATR volatility, trend moving averages, microstructure (wick imbalance), overnight context, and relative strength versus a liquid benchmark. One seat (Risk/Kelly) can stay muted so it does not invent a vote. Weights are signed: an agent that has been wrong in a given universe can be flipped instead of blindly trusted.
- Vote: direction from about -1 to +1.
- Confidence: how loud that agent is allowed to be in the mix.
- Thesis: a short reason string you can audit.
The chair step counts how many seats lean long, short, or neutral and states conviction as a percent. High conviction with fractured votes is a different tape than a quiet unanimous lean. That split is the point of a committee.
How votes become one number
The ensemble takes a confidence-weighted average of votes, then measures pairwise agreement among seats that actually leaned. Agreement can amplify or damp the thesis. Projected probability starts from a conservative universe prior (the names you scan are not a random S&P draw) and is then squashed into a humble band so the UI cannot print heroic win rates from noisy agents. The learning loop and isotonic calibration can still lift or cut those odds after labeled outcomes arrive.
If coverage is thin (few agents with real confidence), the projector pulls back toward the prior. That is why a half-asleep tape should look quieter than a full-agreement session. You are supposed to see the disagreement, not have it averaged away.
How to read conviction versus agreement
Use the two meters as a pair. Strong conviction plus high agreement means the specialists mostly told the same story. Strong conviction plus low agreement means one loud cluster overrode the rest; treat size and time stops as first-class, the way the risk-based playbook already does. Weak conviction is a pass more often than a hunt.
The Wyckoff chart phases sit in the same workflow so you can ask whether the vote is happening in accumulation, markup, distribution, or markdown, not in a vacuum. Sentiment is one seat (Sentiment-X), not the chair. Scanners only decide who is eligible to be scored.
What you still decide
The committee does not place the order. You still choose whether the name fits a cash account or a pattern-day-trader constraint, whether the session is regular hours, and whether a 24-hour pass is even unlocked. Live tickers stay behind the subscription gate. The useful habit is to treat the vote tape like a desk: read the split, read the playbook, then pass or take risk on purpose.
A worked read of a split tape
Suppose Trend-MA and Momentum/RSI are long, Mean-Reversion is short, and Wyckoff is marking time. The chair can still print a modest long if the long seats are louder. That is a “lean,” not a mandate. Cut size or pass. The useful artifact is the thesis strings: if Mean-Reversion cites an extreme z-score and Volume-Flow still says distribution, you are looking at a late gainer, which is exactly the universe prior the projector is trying not to over-celebrate.
Write one sentence before you act: direction, agreement, and whether the playbook clock fits the next hour. If you cannot write that sentence, you are not using a committee. You are using a mascot. Re-read the meters after the next poll; do not average two conflicting cycles into a story. The scoreboard only teaches if each print was a single, dated decision.
Worked example: gainer already up 8 percent
A name hits the scanner because it is already up hard. Momentum/RSI is long, Trend-MA is long, Mean-Reversion is short, Volume-Flow is flat, Wyckoff confidence is low. The chair can still print a long with middling conviction. The correct read is late continuation in a gainer universe, which is exactly why the projector stays humble. Pass, or cut to a token size if you are collecting a label and the account can afford the day-trade count.
Write the sentence: long, agreement fractured, 90-minute clock, late markup. If that sentence feels ugly, you used the committee correctly. If you still click because the color is green, you used the mascot.
Keep a paper log for ten prints: date, agreement, whether you passed, and the label. That log is more valuable than a tenth reread of the agent list. The committee is a desk process. Desks write tickets or they write “no.” Both are data. If you only write the longs you liked, the learning loop and your memory will agree on a fiction.