SCHEMA.md

SCHEMA — data dictionary

Two layers:

1. Event tables (data/**) — one row per captured message, all sharing a single 56-column envelope schema. The venue's own frame is carried verbatim and unparsed in payload_raw.
2. Derived tables (labels/, metadata/) — small, typed, ready to join.

1. The envelope (every row of every file in data/)

One schema across the whole data/ tree, so pyarrow.dataset.dataset("data") reads it as one table. Columns not applicable to a given stream are null; null columns cost essentially nothing in Parquet.

1.1 Always present

column	type	meaning
schema_version	int32	Collector record-format version (1 throughout).
record_type	string	feed_event (websocket / internal) or rest_response.
venue	string	Origin: polymarket_clob, polymarket_rtds, polymarket_gamma, polymarket_data_api, polymarket_crypto_price, coinbase, binance_spot, binance_futures, host.
stream	string	Logical stream. Per-window streams are <sym>/<name> (e.g. btc/poly_a); run-level venue streams are bare (e.g. binance_spot).
channel	string	Sub-channel within the stream — see §2.
sym	string	btc, eth, or null for run-level venue streams that carry both.
window_ts	int64	Unix seconds of the 5-minute window open; null for run-level streams. Window closes at window_ts + 300.
stream_seq	int64	1-based monotonic counter per stream file. Contiguity is the drop-detection invariant: verified gap-free and duplicate-free for every stream (see metadata/quality_report.json).
t_arrival_ns	int64	Host CLOCK_REALTIME at message arrival, nanoseconds. Chrony-disciplined (§4). Use for wall-clock alignment.
t_arrival_mono_ns	int64	Host monotonic clock, nanoseconds. The only valid clock for durations — immune to NTP steps.
n	int64	Payload size in bytes as received.
payload_raw	string	The venue frame, verbatim. Not parsed, not reformatted, not re-serialised. Empty string for records that carry no payload (e.g. connected).

1.2 Websocket transport

column	type	meaning
conn	string	Connection identity, <stream>#<attempt>. Increments on reconnect, so it partitions the stream into connection epochs.
conn_msg_seq	int64	Message counter within one connection. Resets on reconnect, unlike stream_seq.
connections	int64	Cumulative connection count at that point.
url	string	Full endpoint URL including the subscription query string.
frame	string	Outbound frame we sent (subscribe payloads), verbatim.
error	string	Transport error text on read_error.
detail	string	Close-frame detail on server_close.
frames_this_conn, idle_ms, limit_ms	int64	Silence-watchdog state when a watchdog record fires.

1.3 REST transport

column	type	meaning
http_status	int32	HTTP status. 429 rows are retained deliberately — throttling is part of the observable record, and payload_raw then holds the error body, not JSON.
role	string	Why the call was made: strike, close_slow, close_fast, live_poll, final_drain, pre_open.
limit, offset, page, rows, attempt	int32/int64	Pagination and retry state.
leg	string	up / down — which binary leg a CLOB book snapshot refers to.
token_id	string	ERC-1155 token id of that leg.
slug	string	Polymarket market slug, e.g. btc-updown-5m-1785357600.
secs_from_open	float64	Signed seconds from window open (negative = before open).
secs_from_close	float64	Signed seconds from window close.
stable_reads	int32	Consecutive identical strike reads at that point.
taker_only	bool	Trades query filter.
accepted	bool	Whether the collector accepted this read as authoritative.
start, end	int64	Requested time range (unix seconds).
granularity_secs	int32	Candle granularity for backfill.
requested_start_ts, requested_end_ts	int64	Backfill range.
row_order	string	Column order of returned candle arrays, e.g. [time,low,high,open,close,volume].
interval	string	Kline interval, e.g. 1m.

1.4 Collector control / annotation records

Emitted by the collector itself, not by a venue. They document decisions in-band.

column	type	meaning
value	float64	Confirmed strike on a strike_confirmed record.
polls, http_429	int32	Poll and 429 counts behind that confirmation.
confirmations	int32	Repeat reads agreeing on a close price.
next_window_open_ts	int64	The next boundary, recorded when a close is confirmed.
got_close_slow, got_close_fast	bool	Which close paths succeeded.
t_request_ns, t_response_ns, rtt_ns	int64	Clock-probe timings against venue /time endpoints.
chrony_present	bool	Whether chrony was running on the capture host.
extra_fields_json	string	Escape hatch for envelope keys outside this schema. Null in every row of this release — verified, see conversion.no_unknown_envelope_keys.

2. Streams and their channels

2.1 data/clob_ws/{sym}/{window_ts}/poly_{a,b}.parquet

Polymarket CLOB market websocket, wss://ws-subscriptions-clob.polymarket.com/ws/market, subscribed to both legs of one window's market. poly_a and poly_b are two independent simultaneous connections to the same channel — a redundant pair, so that gaps in one can be detected against the other rather than assumed absent. Agreement is measured per window in metadata/quality_report.json → dual_capture_report.

Channels: market (the order-book stream), pong, connected, subscribe, read_error, task_exit.

payload_raw for market is a JSON array of CLOB events, each with an event_type: book (full snapshot: bids/asks as {price,size} arrays, hash, timestamp), price_change, tick_size_change, last_trade_price. The timestamp inside the payload is the venue's millisecond clock; t_arrival_ns is ours. Their difference is the observable transport latency.

Directory names mirror the sym and window_ts columns. They are plain directories, not hive key=value partitions, deliberately: hive keys of those names would collide with the identically-named data columns and break schema merging.

2.2 data/clob_rest/book_rest.parquet

clob/book REST snapshots, both legs, taken around window open. Use these to anchor the websocket stream to a known-good book state.

2.3 data/trades/

file	role
trades_live.parquet	data-api/trades polled during the window (role=live_poll).
trades_final.parquet	Post-close paginated drain (role=final_drain), the authoritative trade set. page/offset/rows describe pagination; secs_from_close how late the read was.

trades_final is the one to use for realised trade flow; trades_live shows what was observable during the window, which is what a live system would have had.

2.4 data/reference/

file	contents
discovery.parquet	gamma/markets responses that resolved slug → condition id → token ids.
strike.parquet	Every crypto-price/open poll, including all 429s. This is the raw evidence behind each strike, and the record of the throttling that forced the backfill.
resolution.parquet	crypto-price/close_slow and close_fast polls plus close_poll_done control records. Two independent close paths per window.
holders.parquet	data-api/holders snapshots, pre-open.

2.5 data/venue/ — run-level, continuous across all 96 windows

file	source	contents
rtds.parquet	wss://ws-live-data.polymarket.com	Channel crypto_prices_chainlink. The settlement oracle. payload_raw = {"payload":{"data":[{"timestamp":<ms>,"value":<price>},…]}}, a batched ~1 Hz Chainlink series.
coinbase_tape.parquet	wss://advanced-trade-ws.coinbase.com	Channel trades+ticker+candles+hb.
binance_spot.parquet	wss://stream.binance.com:9443	btcusdt+ethusdt × aggTrade, trade, bookTicker, kline_1m. Channel spot_combined.
binance_fut_market.parquet	wss://fstream.binance.com	btcusdt × aggTrade, markPrice@1s, forceOrder. Channel fut_market_combined.
binance_depth_snap.parquet	Binance REST	depth snapshots at limit=5000.
backfill.parquet	Coinbase / Binance REST	Historical candles fetched at run start to pre-warm history. row_order gives array column order.
clock.parquet	host + venue /time	Clock-offset probes against CLOB, Binance and Coinbase, with rtt_ns.

3. labels/labels.parquet — 192 rows (96 windows × 2 symbols)

The supervised target, with the full provenance of how each label was obtained.

column	type	meaning
window_ts, window_open_ts, window_close_ts	int64	Window bounds, unix seconds.
sym	string	btc / eth.
slug, condition_id, token_up, token_down	string	Market and leg identifiers.
resolution_source	string	Chainlink stream URL the market settles on.
tick_size, min_order_size	float64	Market microstructure params.
neg_risk	bool	Polymarket neg-risk flag.
strike	float64	The price to beat = the window's openPrice.
close_slow, close_fast	float64	Two independently polled settlement closes.
close_used	float64	The close the label was computed from.
terminal_logret_bps	float64	1e4·ln(close_used/strike) — signed terminal log-moneyness.
outcome	string	Up / Down.
outcome_rule	string	Up iff close >= strike (verbatim from the market text).
label_source	string	live (captured in-run) or backfill_rest+chain (recovered, two-source verified).
labels_agree_live	bool	Whether the live run's two label paths agreed.
outcome_from_slow, outcome_from_fast	string	Outcome under each close path.
close_paths_agree_1usd	bool	Whether the two closes agree within $1.
close_slow_secs_after_close, close_fast_secs_after_close	float64	Read latency after window close.
strike_first_value, strike_first_seen_secs, strike_stable_secs	float64	Strike settling behaviour.
strike_polls, strike_http_429	int32	Poll count and throttling behind the strike.
strike_changed_before_settling	bool	Whether the reported strike moved before stabilising.
backfill_chain_prev_close	float64	Previous window's close, used as the independent chain check.
backfill_chain_exact_match	bool	Whether REST strike matched the chain value.
window_clean_live_capture	bool	The live run's cleanliness verdict, preserved unmodified. Note 17 symbol-windows were unclean live and are clean after backfill; this column records the original state, not the current one.

The boundary-chain invariant

Consecutive 5-minute windows share one Chainlink print, so

strike(window N) == close(window N-1)

for the same symbol. This holds across every consecutive pair in the release and is used two ways: as the second independent source for backfilled strikes, and as a global consistency check on all 192 labels (labels.boundary_chain_consistent).

4. Time and clock discipline

• t_arrival_ns is CLOCK_REALTIME on the capture host, chrony-disciplined against the EC2 link-local NTP service. At run start: system time 161 ns fast of NTP, last offset −536 ns, RMS offset 754 ns, root delay 129 µs. Full chronyc tracking output is in metadata/clock.json.
• t_arrival_mono_ns is CLOCK_MONOTONIC. Use it for any duration or latency — it cannot step backwards.
• Venue-side timestamps live inside payload_raw and are on venue clocks. Cross-venue latency work must treat the two clocks separately; data/venue/clock.parquet carries round-trip probes to quantify the offset.

5. metadata/

path	contents
run_manifest.json	The collector's own end-of-run manifest, verbatim: per-window per-symbol detail, stream counters, run totals.
schedule.json	The window schedule decided at start, verbatim.
clock.json	Host clock state at start, verbatim.
label_backfill_summary.json	Every backfill decision with both sources and the accept/reject verdict, verbatim.
integrity/<window_ts>.json	Per-window integrity record written at finalisation, verbatim (96 files).
window_meta.parquet	Per-symbol-window market metadata, including the raw Gamma and CLOB discovery responses as JSON strings.
conversion_manifest.json	Every source NDJSON file with sha256, byte count, line count, parse failures, and the Parquet file it became with its own sha256 and row count.
quality_report.json	All verification checks with results and supporting detail.