Does RYO's verdict change how often RYO's own plan works?

RYO attaches a trade plan (stop and target) to every verdict and never reports what became of it. Nota asks for 25 tokens daily, locks every plan RYO returns (6 a day since 2 Oct; the other 19 answer partial without technical indicators, so without a plan, and are counted as no_plan) and settles them on OKX candles. The answer so far:

Reading the ledger…

RYO's plans are scored below by which level they touched first; RYO states no probability. Nota's council does, so its 7-day forecasts are scored against always-0.5, persistence, RYO's verdict, the token's base rate and each single LLM on the same resolved questions: the table, JSON.

Horizon

Settling next

Every deep_analysis answer carries a verdict and a bracket: a stop 1.5 ATR away and a target at +1R. These are the plans still open, nearest to settling first. A red mark means the verdict leans against the direction of its own plan.

Settled

Which level OKX's candles touched first: 1-minute candles from the lock to the first full hour, hourly candles after. An hour that touched both levels is split on its 1-minute candles (marked 1m). Neither means the window closed between the two; ambiguous means both fell inside one minute.

Does the verdict matter?

Each pair of verdicts, of confluence states, and RYO's momentum gate passed against failed. Plans are compared within the same lock day and the days are combined with Mantel-Haenszel weights, so neither the market's mood that day nor a lopsided day can fake a difference. The p-value shuffles the labels within each cluster of overlapping windows; nothing is called distinguishable before 10 shared days and p below 0.10.

Nota's forecasts against baselines

Every number below is read from the ledger; no model is called to make a baseline. Hover or focus a column header for its definition.

Coverage

Every lock is a row, including the ones that failed. Nothing leaves this count quietly.

RYO lane reliability

How often each part of RYO's deep_analysis answer came back available, partial or missing, in RYO's own words, across every lock.

Method

  1. Lock. Once a day, deep_analysis is called for 25 majors and the whole answer is stored with its trace id: verdict, confluence state and score, the three gates, and the trade plan. A token locked in the last 20 hours is not locked again, so two locks either side of midnight never count as two days. A missing key, or three failures in a row, stops the run with one aborted row.
  2. Re-anchor. Settlement uses OKX, not RYO, so RYO does not grade itself. The stop and target keep their distance from RYO's entry in percent, applied to OKX's price at the moment of the lock. When the two prices differ by more than 2%, the lock is kept as basis mismatch and never settled.
  3. Settle. OKX's 1-minute candles from the lock to the first full hour, then its closed hourly candles, are walked forward until the stop or the target is touched, or the horizon ends. A window with an hourly candle OKX never published is retried for a day, then counted as unsettleable.
  4. Compare. Day-matched Mantel-Haenszel differences for the target-first rate and for the return at the horizon, a within-cluster permutation p-value, and from 10 shared days a 90% interval bootstrapped over clusters, all with a fixed seed, so this page reads the same every time it loads.

Source: nota/scorecard.py (private repository; judges have access). Raw numbers: /api/scorecard. Not financial advice, and no order is ever placed.