Limitations
The model card. Each entry below is a thing this system cannot do or does not know, stated flatly rather than framed as a feature. Where a limit has a source in the code or in the API, the source is printed with it.
It never trades, and it could not if it tried
The market credential is read-only
Ambang reaches RYO with a credential whose scope is
tools:read. There is no order endpoint in the tool catalogue it can reach, and no write scope on the key to reach one with. This is not a policy the product enforces on itself; it is the shape of the access it was given.scope recorded from GET /api/mcp/whoami on 2026-09-02, six tools confirmed live at GET /api/mcp/tools
The wallet signs a sentence, never a transaction
Connecting a wallet proves you control an address and nothing else. The server issues a single-use nonce that expires in ten minutes, you sign a plain text message with it, and the server recovers the signer from the signature. The message you are asked to sign ends with the line
This signature proves ownership only. No transaction is requested.
No code path in this product constructs a transaction, asks a wallet to send one, holds a private key, or talks to a chain RPC. Signing in claims authorship of a strategy. It does not authorise spending, and there is nothing here for it to authorise.
apps/server/app/auth.py, SIGN_IN_TEMPLATE, recover_signer(), verify_sign_in()
Every position is on paper
An entry price on a receipt is the price the evidence reported at the moment the decision was made, and an exit price is the price the evidence reported at the horizon. Nothing was bought and nothing was sold. The hackathon rules this project was built under prohibit moving real funds, and the product was built to match.
It holds no price history, so it draws no trends
RYO has no history endpoint
Every RYO tool returns a point-in-time read. There is no way to ask any of them what a price, an RSI or a funding rate was an hour ago. One tool looks backwards at all, the sentiment-shift monitor, and its window is fixed at seven days.
Every piece of history in this product is therefore history the product itself accumulated, one stored snapshot at a time. A first-time visitor has nothing to compare anything against, and the interface says so rather than filling the gap.
No trend is drawn anywhere
There is no sparkline, no trend line, no moving average and no ghost plot in this interface. That is not a stylistic preference. With a database this small, a line through the points would be a shape invented by the renderer rather than a shape present in the data, and a reader would have no way to tell the difference. The only chart form in the product is a row that shows one current value against one fixed threshold.
The record is tiny, and none of it has been graded
Reading the size of the record from the server…
Calibration is the only reputation signal this product recognises, and calibration needs resolved outcomes. Until receipts reach their horizon and outcomes resolve, the Brier score is undefined and every comparison group is empty. A confidence of 70 currently means this strategy declares 70
, not this strategy is right about seven times in ten
.
The source is rate-limited, and the key expires
Sixty calls per minute
The RYO credential is limited to 60 calls per minute, with no published daily or monthly cap. One full cycle across three symbols costs roughly eight calls, so ordinary operation sits well inside the budget, but a burst of manual cycles can exhaust a minute. When it does, the affected tool returns unavailable and the decision degrades rather than waiting or guessing.
limit recorded from GET /api/mcp/whoami on 2026-09-02, per-cycle call count measured against the live catalogue the same day
The credential expires on 2026-12-17
After that date the live path stops working. A failed RYO call returns an envelope marked
unavailable; it never returns a stored fixture dressed as a live read, so an expired key produces visible degradation rather than a quietly stale product. Receipts and snapshots already stored stay readable and replayable, and each keeps the source label it was recorded with. Nothing captured before the expiry becomes less verifiable, but nothing new can be fetched.expiry recorded from GET /api/mcp/whoami on 2026-09-02, apps/server/app/ryo_client.py, call_with_retry()
Degraded evidence produces a degraded decision, and says so
When a field a rule reads did not resolve, the rule cannot be evaluated. The side falls to NO_TRADE, the decision is flagged degraded, the confidence loses 15 points, and the receipt records the reason in words. Any flip condition that depends on the missing field is reported as not evaluable, not as a distance of zero.
Two words that are easy to confuse, and that this product keeps apart: evidence is partial when any field is missing from the snapshot at all, and a decision is degraded only when a field that this strategy actually reads is missing. A receipt can honestly carry partial evidence and an undegraded decision, and several in this database do.
Degradation is not an error state to be cleared. It is the correct output for incomplete evidence. A version of this product that filled the gap to avoid the label would be less useful and less true, and it would break the one thing the whole design rests on: that the threshold which would flip a decision can be computed from what is actually known.
It is not advice, and it is not a service
This is a hackathon entry: a research instrument built to be inspected, not a product with users, uptime or support. It may be unavailable, it may be wrong, and it may stop entirely without notice.
Nothing on it is financial, investment, legal or tax advice, and nothing on it is a signal to act on. It has no user research behind it, no benchmarks, no track record and no price. A decision printed here is a deterministic reading of a rule against a snapshot, and the honest use of it is to disagree with it in the open, using the same evidence it used.