Manifest order books, 11–12 September 2026

The files are assets of the dataset-v1 release. Each table comes in both formats with the same rows in the same order. Take Parquet if your tools read it; it keeps types, including exact 38-digit prices.

SHA256SUMS lists a checksum for every file. The export is deterministic: the same rows always give the same bytes.

No field is ever empty. Timestamps are UTC with microsecond precision. Prices and sizes are integers in the smallest on-chain units, exactly as the market account stores them; the next section converts them.

levels

One row per book state × side × level. Join to states on state_id.

states

One row per distinct book state. A new state is recorded when the market account changes, so the time between rows is the time the book stood still.

arrivals

One row per state per channel that delivered it. A state missing a row for a channel was not delivered by that channel.

markets

channels

coverage

With b and q the base and quote decimals from markets:

price in quote per base = price / 10^18 × 10^(b − q)
size in base tokens     = size / 10^b

For cbBTC/USDC (b = 8, q = 6) a price of 777037910000000000000 is 77,703.791 USDC per cbBTC. Keep the raw integers for any sum or difference and convert at the end: that is how the terminal stays exact.

Deep levels can sit far from the market: some cbBTC asks rest at 10⁸ USDC and above, and their price runs to 27 digits. In CSV, price is a plain integer. A reader that guesses types turns it into a float and loses the last digits; read it as a string or a decimal (in pandas, dtype={"price": str}).

DuckDB reads the files where they lie. The best bid and ask of cbBTC over time, in USDC:

SELECT s.first_seen_at, l.side,
       l.price / 1e18 * 10 ^ (m.base_decimals - m.quote_decimals) AS price_usdc,
       l.size / 10 ^ m.base_decimals AS size_base
FROM 'quantedge-dataset-v1-levels.parquet' l
JOIN 'quantedge-dataset-v1-states.parquet' s USING (state_id)
JOIN 'quantedge-dataset-v1-markets.parquet' m ON m.market_id = l.market_id
WHERE m.base_symbol = 'cbBTC' AND l.level = 0
ORDER BY s.first_seen_at, l.side DESC;

How far each channel lagged behind the first arrival, at the 95th percentile:

SELECT c.name, count(*) AS states,
       quantile_cont(epoch_us(a.received_at) - epoch_us(s.first_seen_at), 0.95) / 1000 AS p95_lag_ms
FROM 'quantedge-dataset-v1-arrivals.parquet' a
JOIN 'quantedge-dataset-v1-states.parquet' s USING (state_id)
JOIN 'quantedge-dataset-v1-channels.parquet' c USING (channel_id)
GROUP BY c.name;

• One collector held an accountSubscribe WebSocket subscription per market at each of two RPC providers, Helius and Alchemy, at confirmed commitment with base64 encoding.
• Each update was keyed by (slot, SHA-256 of the account bytes) before decoding, so the same state from both providers is one row in states and two rows in arrivals.
• The book was decoded from the Manifest market account, aggregated by price and cut to the best 15 levels per side.
• received_at comes from the clock of the single host that ran the collector. Differences between channels are measurements; the absolute time since the on-chain event is not, and is not in the data.
• The collector also records the coverage it had: the window is one segment per market with no gap, which is why it starts at an odd microsecond.

The dataset is licensed under Creative Commons Attribution 4.0 International (CC BY 4.0): use it for anything, including commercial work, and credit the source. The code that recorded and exported it is under the Apache License 2.0.

QuantEdge Terminal. Manifest order books, 11–12 September 2026 (dataset-v1).
https://quantedgeterminal.github.io/QuantEdge-Terminal/dataset/. CC BY 4.0.

The export is apps/dataset in the repository. It refuses a window with a gap in coverage and any channel that is not real.