Track 1
Autonomous agents
The council, the receipts, and a watch loop that decides on a schedule, exits practice positions at stop or target, scores what has matured, and publishes to Telegram or Discord.
A trading call you can re-run.
A council of agents argues over live RYO market evidence, then stores the call as a receipt that reproduces exactly.
RECEIPT—
- Symbol
- —
- Decision
- —
- Chance it rises in 7 days (judge)
- —
- Evidence
- —
- Issued
- —
The decision above, reproduced from the evidence it was made on. Press “Verify a receipt now” to re-run it.
Nota audits the data it is given.
RYO publishes an ATR and an RSI. Nota recomputes both from public candles with Wilder's method and compares them, and it checks RYO's price against the median of up to five public price sources. When they disagree, the judge refuses to size a trade at all. The figures below are read from the newest receipt that holds all three checks, and every receipt in the ledger replays identically - including the one recorded on 10 September, when the builder key started answering 401 and all five RYO sections went to error at once. Nothing was patched for it: the council read what was left, the judge refused the trade, and the diff put the collapse at the top of the page.
Measure
Nota
RYO
Apart
ATR(14), USD
—
—
—
RSI(14)
—
—
—
Price, USD
—
—
—
reading the ledger…
And when it is not given at all.
A market tool is judged on the day its data source fails, not on the day it works. Below is the most recent receipt in this deployment where something did fail, read from the ledger as you load the page - not an illustration of what would happen. Each section carries the status its source returned, a value that could not be fetched stays null instead of becoming zero, and the judge is told what it does not have.
reading the ledger…
Want to see it fail on demand? The failure drill breaks RYO, the exchanges or the model on purpose and runs the real pipeline offline beside a healthy run.
It also keeps RYO's score.
Every deep_analysis answer comes with a verdict and a trade plan, and RYO keeps no
record of what became of either. Nota locks those plans for 25 majors once a day and settles each one on
OKX's hourly candles, so one question finally has an answer: does RYO's verdict change how often RYO's own
plan works?
reading the scorecard…
How one decision is made
Gather
Five RYO tools are read into one evidence pack. A tool that fails becomes a section marked with its error, and the pack is still built.
Argue
A macro, a technical and a narrative agent each return a stance, a probability and citations as dotted paths. A citation pointing at absent evidence is dropped in code, and the opinion is downgraded for it.
Judge
A fourth agent weighs the three by their own Brier score from past calls, then decides long, short, or no trade.
Size
Pure arithmetic, no model: stop at twice ATR, target at three times, size from one percent of the account. A missing price or ATR blocks the trade rather than assuming a number.
Record
Everything lands in a receipt with each number's RYO path, as_of, data_mode and trace id, so the decision can be replayed and later scored against what the price actually did.
Every receipt draws itself
These are the decisions in the shipped ledger. Every mark is generated from that receipt's own evidence hash, and its arc is the probability the judge stated, so two receipts can never draw the same figure and the same receipt can never draw a different one.
The dashboard opens on a thirty second summary: what was decided, what changed most since the previous receipt for that symbol, where the practice position stands, and what happens next. Every receipt is diffed against the one before it and ranked by impact, degraded evidence is announced at the top rather than buried, and the whole page is reachable from the keyboard. Open the dashboard.
Rules the code enforces
- null stays null
- A missing value is never turned into zero, anywhere in the pipeline.
- values come from evidence
- Each cited number is read back out of the evidence pack, never retyped by the model.
- simulated never trades
- If RYO marks the primary evidence as simulated, sizing is blocked and the receipt says so.
- no order is ever placed
- The RYO surface used here is read only. Positions are paper, and they are labelled as paper.
- the label is on the receipt
- Live, recorded and fixture evidence each carry their own source label, printed on the receipt.
- agents earn their weight
- Once a call resolves at its seven day horizon, each agent's Brier score changes how much the judge listens.
Three tracks, one system
Track 2
Dashboard
Diff first and keyboard first, with replay verification, Open Graph receipt cards, a public backing leaderboard scored when a call resolves, and the RYO Verdict Scorecard.
Track 3
Skills
Seven research skills served on RYO's own /api/skills paths in RYO's envelope, so adding them to RYO is a route registration rather than a port: narrative convergence, news verification, price and technicals cross-checks, a derivatives positioning check, a move base rate, and RYO's own verdict track record.
Public reasoning, publicly scored
A call only its author can see costs nothing to make. Every receipt has a permanent link and renders as a card when the link is shared, so the argument travels with the number.
Anyone can back a receipt or disagree with it under a handle, one stance each, before the outcome is known. At the seven day horizon the call is scored against a fresh price and every backer's accuracy moves with it. A no trade is never scored.
The agents are on the same hook: each carries its own Brier score from past calls, and that score is what decides how much the judge listens next time.
The public reasoning feed shows every call with each agent's stance, who dissented and why, and a share link. Its reputation board ranks backers and agents on the same hit rate, and stays unranked until a record is long enough to read. Handles can also keep a public watchlist.
/r/<id>.png.