# MLB Cross-Venue Matched Book — Schema and Methodology

One row anchors a captured Polymarket sample for `team` and pairs it with the
nearest captured Kalshi sample within **±30 seconds**. A later Kalshi sample can
be chosen. This is retrospective nearest-neighbour alignment, **not a causal,
backward-only as-of join** and not an execution backtest. The dated game key is
`(date, game)`; coverage and absent dates are explicit in `MANIFEST.json`.

| Column | Type | Meaning |
| --- | --- | --- |
| `ts` | double | Polymarket capture/sample timestamp, Unix seconds UTC; not an exchange freshness timestamp |
| `date`, `game`, `team` | string | dated matchup key and represented outcome team |
| `poly_bid`, `poly_ask`, `poly_mid` | double | captured Polymarket quote and midpoint on the 0–1 scale |
| `kalshi_yes_bid`, `kalshi_yes_ask`, `kalshi_mid` | double | nearest captured Kalshi YES quote and midpoint on the 0–1 scale |
| `xvenue_spread` | double | `poly_mid - kalshi_mid`; not executable profit |
| `winner` | string | carried finalized Kalshi label; null when unresolved |
| `won` | double | carried 1/0 label for the represented team; null when unresolved |
| `settled` | bool | finalized resolution evidence was present in the source build |

The capture timestamp says when our sampler recorded the value. It does not say
when the exchange last confirmed the quote. There are no exchange timestamp or
depth-validity fields here. Retained prices can therefore be older than their
sample times. A ±30-second sample match does not guarantee fresh prices, valid
size, a fill, simultaneity, or executable arbitrage. Sampled cadence is roughly
30–60 seconds; it is not event-level or tick-complete capture.

The frozen v2 observation window is 00:00 UTC on the game's dated key through
strictly before 12:00 UTC the following day, subject to an available Kalshi
sample within the join tolerance. This includes evening play after UTC midnight,
which the original one-file loader dropped. It does **not** stop at the final
out: post-final and postponed-game quotes may persist in the window, and a
rescheduled listing can retain its old date. Do not interpret the window as
game finality or complete in-play coverage.

Doubleheaders are excluded rather than ambiguously joined: the pair key cannot
distinguish games with multiple Kalshi times or an explicit G1/G2 marker. Only
games discovered and captured on both venues appear. One Polymarket outcome is
independently captured per game; an opposite-side complementary price is a
derived identity, **not an independently observed opposite-side book**.

This packer copies the supplied matched Parquet bytes without changing quote
fields or carried labels. It does not fetch outcomes or settle games. The source
builder must enforce the stated dated-key and doubleheader exclusion policy;
packing a file does not independently verify a discarded G1/G2 marker. Unresolved
labels remain null, never losses. Check exact source-byte preservation and the
current per-file hashes in `MANIFEST.json` before use.