dataset-v1

Manifest order books, 2026-09-11 17:31 → 2026-09-12 09:05 UTC: three markets (cbBTC, RENDER, USDT vs USDC), 106,440 book states, top 15 levels per side, per-channel arrival times from two RPC providers. Parquet and gzipped CSV; fields and units: https://quantedgeterminal.github.io/QuantEdge-Terminal/dataset/. Checksums in SHA256SUMS. Licensed CC BY 4.0.

v0.2.0 — the result says what the data can and cannot resolve

Since v0.1.0, what a user sees changed in four ways.

A delay grid finer than the data now says so. Book states arrive about 1.5 s
apart, so a shorter delay looks at the same state as the faster level and
returns the same row — by construction, not because the delay is free. Each
row reports the share of steps on which its view actually moved, "same view"
where it never did, and the run reports the median gap that caused it. The
cost of 100 ms carries its grid: it is a secant between the two end levels, and
the result now says that, warns when the far end is a level the data cannot
see, and states that the 0 ms row is the execution model rather than a market
measurement.

A channel that stops delivering is named on the compare screen instead of
leaving it empty, and it is resubscribed on its own.

A channel can be retired without losing what it measured: its row and arrivals
stay, it stops being offered as a lane.

The subscription names its encoding instead of relying on the provider's
default, which silently delivered zero-byte accounts on a provider that follows
the specification.

Unchanged: SC-003 is still 0 of 3. Whether the cost of latency rises above
noise on this dataset is not settled by this release.

v0.1.0 — first release.

A terminal that shows how much market-data latency costs a strategy, measured on real Manifest order books recorded from two independent RPC channels. Run screen: five delay levels, "delay → P&L" and the cost of 100 ms; Terminal: live book with measured channel lag, data age and staleness; Compare: two real channels side by side, an emulated lane only on a typed profile and always labelled as such. Runs over incomplete data refuse and name the gap; every run is byte-for-byte reproducible. Nothing is sent to any market.