Microstructure/docs/DATA_CONTRACT.md

Data contract and lineage

Scope

The normalized layer separates exchange-specific acquisition from research logic. An adapter may add a new venue, but it must produce the same versioned event contracts, preserve original bytes, and declare how observation time is approximated. No normalized table is evidence that an event was available to a real colocated strategy unless local receipt time was actually captured.

Current schema version: 1.0.0.

Clocks and ordering

All timestamps are signed UTC epoch nanoseconds:

• event_ts_ns: timestamp supplied by the market-data source;
• received_ts_ns: local wall-clock receipt when captured live, otherwise null;
• available_ts_ns: earliest time the pipeline permits the row to enter an information set;
• availability_basis: explicit reason, such as local_receive_time, exchange_event_time_proxy, or synthetic_receipt;
• capture_seq: local arrival ordering when a collector supplies it;
• continuity_id: a feed epoch that cannot be crossed by rolling features, labels, open orders, or markouts.

Within a continuous book epoch, sequence IDs—not exchange timestamps—are the authoritative reconstruction order. Research ordering is stable on availability time, sequence/capture order, and event identity. Separate trade and book streams with equal timestamps have no assumed common ordering, so cross-stream joins are strictly prior unless a future source proves a shared sequence.

Exact numerical representation

Adapters retain integer price_ticks/quantity_lots and the corresponding tick_size/lot_size. Floating price and quantity columns are convenience units and must agree with the exact representation. Binance symbol scales come from public exchangeInfo filters for each download; fixed 1e-8 scale defaults exist only as explicit low-level fallbacks and are not the configured sample path.

Derived ratios, log returns, volatility, probabilities, and P&L use Float64 with their units named or documented. Execution quantities are rounded down to the observable lot size; slippage-adjusted prices round adversely to the tick.

Normalized tables

Trades

Identity is (venue, symbol, trade_id). Required economic fields include exact and floating price/quantity, quote quantity, first/last constituent trade IDs, buyer-maker flag, normalized aggressor side, timestamps, and source artifact ID. A buy aggressor lifts the ask; a sell aggressor hits the bid.

Depth deltas

Each event contains first_update_id, last_update_id, optional previous update ID, and bid/ask lists of exact (price_ticks, quantity_lots) changes. Quantity zero is a delete instruction; a negative quantity is invalid.

Book snapshots

A snapshot carries its request/receipt/availability times, last update ID, depth limit, exact levels, scale metadata, source artifact ID, and a new continuity ID. Binance REST snapshots do not provide an exchange event timestamp; the local receipt time is the anchor availability time.

Book observations

Reconstruction emits best bid/ask, L1 quantities, cumulative depth at 1/5/10 levels, spread, mid, microprice, queue imbalance, sequence range, validity flag, and full lineage. A crossed/locked or emptied book terminates the epoch rather than being silently repaired.

Sequence gaps

A gap row records expected and observed ranges, the missing inclusive range, detection time, continuity ID, source artifact, and reason. After a forward gap, the book is not live again until a new snapshot starts a new epoch.

Binance acquisition semantics

The configured historical adapter uses only credential-free public market-data REST endpoints. It starts aggregate-trade pagination with a fixed UTC interval, then advances by aggregate trade ID so trades sharing a timestamp are not lost. HTTP 408/418/429/5xx, connection failures, and recoverable interruptions while streaming an HTTP 200 body use bounded exponential backoff; Retry-After is honored for rate-limit responses. Response bodies have a byte ceiling. Every accepted exact body is content-addressed and manifested before normalization; an interrupted or oversized response preserves a bounded rejected prefix and an explicit rejection sidecar before retry/failure.

The optional live collector uses the market-data-only WebSocket endpoint and diff-depth U/u updates. A correct local book buffers updates, fetches a public snapshot, discards stale events, requires the first usable event to cover lastUpdateId + 1, and then validates continuity for every event. Reconnection starts a new continuity epoch. These rules follow Binance's official Spot WebSocket stream guide. Public REST behavior and rate-limit headers are described in the official Spot REST documentation.

Live capture persists each WebSocket frame before UTF-8/JSON parsing in a typed base64 journal containing exact bytes, local receipt time, capture sequence and continuity epoch. Each reconnect snapshot's raw body and sidecar are journaled as an anchor. A frame above 1 MiB fails only after its evidence is preserved; normalized Arrow spools flush before an estimated 16 MiB batch ceiling. A capture-ID-scoped normalized root produces one streaming Parquet descriptor per nonempty table, and a capture-ID completion summary is atomically published last. The fixed summary name is only a latest-pointer, never the sole completion record. Historical REST output retains the date-partition layout described below.

Every historical download requires a finite per-symbol event cap, but that cap is not treated as a RAM budget. The default path lazily yields at most one REST page per Arrow batch, updates exact disk-backed quality state, and feeds the Parquet writer once. Eager compatibility materialization has a separate hard row guard. A ConfiguredDataAdapter protocol/registry is the normalized extension boundary for later venues or institutional sources; its mode must match the resolved configuration before dispatch.

Daily-archive acquisition boundary

A daily-archive acquisition object authenticates raw transport evidence; it is not a normalized-data object. Acquisition may hash exact bodies, authenticate the official CHECKSUM, and inspect bounded ZIP end-of-central-directory/central-directory metadata for member name, count, and declared compressed/expanded sizes. It must not open or extract the CSV member, read decompressed member bytes, parse a header or row, expose economic fields, or compute row/ID/timestamp coverage. The first decompressed member byte is the economic-data-open boundary, and the archive API must expose a fail-closed guard immediately before that byte can be read.

For the prospective M8 study, all declared ZIP, CHECKSUM, and exchange-metadata responses are acquired and placed in an immutable raw acquisition manifest before any member is opened. Only train and validation members are then stream-normalized and quality-checked. Their immutable development normalized manifest is an input to selection. One per-symbol analysis lock and an aggregate lock committing both child-lock hashes are closed and fsynced, along with their digest files and containing directories. Each child lock binds the selected specification, development-frame identity, feature order, imputer and scaler parameters, selected-estimator state, independent historical-prior state, calibration state, fit cutoffs, and the canonical numeric fitted-state SHA-256. Those states are fit on development data and durably closed before held-out member access. The aggregate lock binds the protocol/config, raw acquisition manifest, development normalized manifest, and clean real Git revision. Every primary/replication member-open guard must re-read and re-hash that exact durable lock before allowing decompression. Test normalization may construct held-out features, but prediction restores the locked numeric state; evaluation cannot reselect, fit, refit, recalibrate, or update it.

Declared-object 404/410 responses, invalid exchange-metadata semantics, official CHECKSUM or ZIP structural violations, frozen per-response size violations, and retained-evidence budget exhaustion are typed deterministic insufficiency and publish an immutable raw-only INSUFFICIENT_DATA authority. Retry exhaustion, connection interruption, permission/local-I/O errors, collisions, and program faults remain nonterminal system failures so they cannot masquerade as a market or data result.

The M8 byte limits have distinct meanings. The compressed ceiling applies to each archive response while streaming. The expanded ceiling applies to both the declared and actually streamed bytes of each CSV member. The total-download ceiling is the hard sum of immutable raw-response evidence retained for the study: ZIP bodies, official CHECKSUM bodies, every response sidecar, exchangeInfo bodies and sidecars, and all retained rejected-response prefixes and rejection sidecars. Retried attempts still consume the total. A physically retained content-addressed file is counted once even if referenced more than once; separate retained copies are counted separately. Normalized Parquet, quality outputs, and derived aggregate indexes are outside this raw-evidence sum. Type-specific bounded CHECKSUM, metadata, and rejection responses remain subject to the same total. Accounting is enforced while bytes are retained, not after all downloads finish, and the raw acquisition manifest records every accepted/rejected artifact and the final exact total.

If held-out normalization, continuity/completeness validation, or quality fails after lock durability, the run publishes an atomic, checksum-protected, immutable INSUFFICIENT_DATA terminal artifact under the same run identity. It binds the raw and development manifests, per-symbol and aggregate locks, clean Git identity, typed failure, partial-evidence hashes, and unopened remainder; it contains no endpoint result. That identity may subsequently verify/reuse the failure but may not overwrite it, return to selection, substitute inputs, or publish a partial successful bundle. A source change creates a new clean Git identity rather than rewriting prior evidence.

Storage and manifests

Normalized Parquet is content-addressed and partitioned beneath:

<root>/<dataset>/schema-<version>/venue-<venue>/symbol-<symbol>/date-YYYY-MM-DD/

Writers reject oversized input batches before row conversion and consume bounded Arrow record batches; they do not require all event rows in memory. Each part has an immutable sidecar with source URI, download/creation time, requested and observed ranges, row/byte counts, schema version, transformations, input checksum, write ordinal, and Parquet checksum. A dataset manifest lists all parts, sidecars, hashes, row counts, observed bounds and write order. Validation checks those bytes before scanning rows. Run bundles snapshot these manifest hashes and protect every included file with checksums.sha256.

Each config-driven ingestion also writes a content-named immutable ingestion manifest linking only the exact raw responses used by that invocation to the normalized dataset manifests. It records the row cap and each symbol's complete_range state, so capped coverage cannot survive merely as terminal output or be mistaken for the entire requested interval. Quality reports and full findings JSONL are atomically published under per-run names and their SHA-256/byte counts are bound into the same ingestion manifest.

The public research reader verifies metadata, URI/symbol semantics, exact raw record membership, inverse raw-to-normalized coverage, scales, parts and sidecars before exposing rows. Its batch API performs physical-order incremental quality checks while a single upstream Arrow stream is staged through a memory-limited, spill-capable DuckDB canonical sort. The legacy eager reader has an independent finite materialization guard checked before any Parquet row read.

For M8 specifically, the raw acquisition manifest and normalized manifests are different immutable stages and must not be collapsed into one post hoc record. A self-contained run preserves the acquisition authority as an exact raw-only copy below data/input; normalization, Parquet, and DQ evidence live separately below data/normalized_input. The final all-date manifest is rooted at data so it may bind both trees without making derived artifacts members of the raw authority. A failed run inventories every pre-terminal regular file, binds any completed and failed normalization evidence, and applies the same verifier both before and after atomic publication. A successful final bundle binds the raw acquisition manifest, final normalized manifest for all eight declared symbol/dates, official CHECKSUM and metadata bodies plus their sidecars, per-symbol and aggregate lock hashes, and the clean real Git revision. Verification rejects a dirty, unborn, synthetic, or changed source identity and rechecks all bound artifact bytes before reuse.

External raw and normalized data are ignored by Git. Small deterministic test fixtures may be committed only under tests and must state whether they are synthetic or sampled public observations.

Frozen dual-symbol live-L2 session authority

Each prospective L2 date uses one absolute UTC start/end barrier and one cross-symbol authority for BTCUSDT and ETHUSDT. The two single-symbol producers run concurrently and preserve raw frames before parsing. Coverage is computed only from intervals backed by consecutive OBSERVED reconstructed book states; raw socket first/last span, stale updates, excluded messages, gaps, and silent holes do not create coverage. The session gate uses the exact union/intersection of those intervals, requires one sufficiently long valid continuity epoch, reconciles raw/normalized/reconstructed/excluded rows and snapshot anchors, and enforces the frozen gap, error, warning, frame-byte, and Arrow-batch limits.

All per-symbol raw journals, snapshots, normalized data, manifests, quality reports, and capture summaries must form an exhaustive regular-file inventory. A passing dual-symbol session is published atomically with checksums and exact _SUCCESS bytes; a typed capture/gate failure publishes an immutable INSUFFICIENT_DATA authority. Permission, local-I/O, source/config/protocol drift, or program faults retain nonterminal raw evidence but may not publish a research terminal. Neither terminal authorizes live trading.

Quality policy

Validators never sort, deduplicate, clip, interpolate, or rewrite observations. They emit typed findings for duplicate trades, timestamp/order reversals, availability/receipt contradictions, sequence gaps or stale updates, crossed books, invalid price/quantity, scale mismatches, abnormal spread, nonmonotone depth, long silence, and receipt-clock reversal. A downstream research view may exclude a row, but it must preserve the normalized source and report the exclusion count/reason.

Incremental validators retain only a bounded in-memory finding preview; exact duplicate/clock/sequence state spills to SQLite and complete findings stream to JSONL. A failed validation closes and removes only its temporary sink, leaving any previously published evidence untouched.