What Pyth's sponsored feeds actually did.

PER FEED · LAST 72 HOURS

USABLE — share of our Pyth reads where this feed passed every check (owner is a Pyth program, discriminator is PriceUpdateV2, verification level is Full, feed id matches, publish age under 120s). TELEMETRY — reads that carried publish_time, confidence and age, which is what every distribution to the right is computed from. It is lower than SAMPLES because price sampling predates this page; below 30 the distributions are withheld and the cell shows a dash. ADVANCES — publish_time moves we could attribute to the feed. BLIND (OURS) — pairs of looks more than 90s apart, where we stopped sampling; thrown out of every figure on this row rather than charged to the feed, because our outage is not their stall. GAP — seconds between consecutive publish_times. This is an upper bound: we only ever see a feed's latest publish, so an interval we measure may contain publishes we never saw. AGE — how old the price was when we read it, truncated by our own 120s filter (anything older is a miss, not an age). CONF — the publishers' own confidence band, in basis points of price. DIV — absolute distance from Coinbase spot, in basis points, sampled every 10 minutes, withheld below 6 cross-checks; DIV n is how many there have been.

WHAT THE TIMING COST · LIFETIME

This is the part only a game can measure. A shot expires; we look for the first oracle sample at or after that instant. If none has landed yet the shot is deferred and we look again. If the 15-minute grace window closes with nothing, the shot is voided and the stake goes back — nobody wins, nobody loses, and the feed's timing is the only reason.

SINCE WE STARTED · 2026-08-20 → 2026-08-21

The table above is a 24-hour window, because the raw per-minute record is kept for four days and then expires — it exists for settlement, not for posterity. Every completed day is folded into a summary before that happens and the summary is kept for a year. This is the part that cannot be built by trying harder later. It can only be built by having started earlier.

CONF · DAILY MED is the median of each day's median confidence band — a robust summary, but not the median of every individual reading, and the page says which it is rather than calling it "typical" and leaving you to guess. DAYS is how many days it rests on; below 3 the figure is withheld, because one day's median wearing the word typical, in a row whose sample count is summed across every day, is the same trick this page exists to avoid. REWINDS counts publish_time moving backwards, which should never happen. It is here because a number that is always zero is worth publishing precisely so that the day it is not, somebody notices.

METHOD

• One getMultipleAccounts read of each sponsored price account per minute, over plain Solana JSON-RPC. No Pyth API, no key, no plan — these are ordinary accounts and reading an account is not something anyone can bill for.
• PriceUpdateV2 is decoded locally from the raw bytes. A number is kept only if the account owner is one of Pyth's own programs, the Anchor discriminator matches sha256("account:PriceUpdateV2")[..8], the verification level is Full, the 32-byte feed id equals the one we expect, and publish_time is under 120s old. Anything else is dropped and counted as unusable rather than quietly used.
• Settlement uses the first recorded sample at or after a shot's expiry — the same first-crossing rule (prev_publish_time < expiry <= publish_time) the on-chain program enforces — so these samples are not a side record. They are the evidence.
• Completed days are folded into a kept summary before their raw samples expire, so the all-time view above keeps growing while the per-minute record stays a four-day rolling window. The fold uses exactly the same code as the live report — a day summarised months ago was computed the same way as the last hour.
• Every sample behind every statistic is served back raw: /api/game?action=path&feed=SOL&from=<ms>&to=<ms>. Recompute anything here, or disagree with it, without asking us.

LIMITS

• Price sampling predates this telemetry. A feed can report many usable reads while its gap, age and confidence figures rest on far fewer — `telemetry` is the count those distributions actually use, and below 30 of them the distributions are withheld rather than published thin.
• Gap figures are an UPPER BOUND. We sample once a minute and only ever see a feed’s latest publish_time, so an interval we measure may contain publishes we never saw.
• Gaps are only counted between observations less than 90s apart. Anything wider means we stopped looking, and is counted as a blind window of ours instead of a stall of theirs.
• Observed age is truncated by our own filter: a read older than 120s is discarded, so it lands in misses rather than in the age distribution.
• ourDutyPct is OUR sampling coverage, not the feeds’ uptime — a serverless instance that never woke up records nothing.
• misses counts a feed unusable in a sample where Pyth itself read fine (stale beyond 120s, wrong owner, wrong feed id, or account missing).
• Divergence is measured against Coinbase spot every 10 minutes. It is disagreement between two sources, not error in either.
• settle counters are lifetime totals from the moment this measurement shipped, not windowed.