How the research works.

Where the numbers come from

• Daily cron calls /api/paper-run at 00:30 UTC, pinned to fra1 (the US region cannot reach the funding endpoint).
• One point per UTC day, idempotent — a second call is a no-op.
• Same five majors — BTC, ETH, SOL, BNB, XRP — same parameters as the backtest.
• Appended points merge over the committed public/paper.json; the committed file wins on collision.
• “Generated at” is the newest point on record, never the current time.

When the data isn’t available

• Nothing is written. No interpolation, forward-fill or estimate. The curve stops where the data stopped.
• It fails in the status code: runner exits non-zero, endpoint answers HTTP 503 naming the symbol. 200 only when a point was appended or already existed.
• An earlier version answered 200 with a “skipped” body; monitoring read it as success and the record sat frozen for 22 days. Silent success on an outage is the one failure a transparency page cannot afford.
• Venue fallback. Binance first; on failure or an eight-second timeout, Bybit. Never blended — one exchange per number, recorded per symbol per day; /paper names fallback days; older points show no source. Removes one cause of a frozen record; does not prove it was the cause.
• A point computed but not saved is also HTTP 503, and the outcome record notes whether it reached the store. With no store configured, /paper says so.

How staleness is surfaced

Age is computed at render time from the data, never hardcoded or baked in at build. Past 48 hours:

• /paper switches to “not currently reporting” with the last date and age. An unreadable date fails safe to the warning.
• /funding drops “live” and labels the table a committed snapshot with date and age. An unparseable timestamp fails safe to stale.
• Both self-correct when the runner stops or resumes.

Each run also writes its outcome — appended, no-op or skipped, with reason — to a rolling 30-day record /paper reads back: a missing point cannot distinguish an honest skip from a job that never fired. Never inferred from the curve; with nothing recorded the page says it cannot tell.

What is deliberately not here

• No live track record yet. Zero real fills; /track-record shows an explicit empty state, not a sample curve.
• No third-party capital, no custody. The app never places an order, touches an exchange key or moves money.
• No signals for sale. Nothing here is a buy or sell recommendation.
• Sequencing in order: backtest, paper run, then — only then — Jukka Blomberg trading his own money, disclosed as a personal record.