EdgeArbiter/2LEVEL_Data_Unverified

Original Level 2 captures and recovery

This manually gated dataset preserves complete original captures, closure/compression metadata and recovery tooling. Raw captures intrinsically contain their original identifiers. Book reconstruction quality and complete exchange coverage remain unverified.

Current research Parquet lives separately in 2LEVEL_Parquet_Data_Unverified. parquet/<contract>/<date>.events.parquet retains event-driven depth updates, trades and session/gap controls. The paired .snapshots.parquet retains every original captured snapshot observation. Files are grouped by CME trading day, including captured overnight hours. Derived viewer files are in replay/<contract>/; the laptop's Inspect action selects them automatically. Parts preserve distinct incompatible streams or stay within existing viewer capacity. The raw catalog pins the repository/revision and does not duplicate those files here.

Browse recordings, CSV, or exact metadata. Incoming market bounds refer to delivered depth/trade receipt times. They do not establish uninterrupted capture or exchange alignment. Control-only captures have no playable market rows. Original sequence boundaries, gaps and repeated trades remain meaningful.

Restore original journals on another computer

Install Python 3.11 or later, download archive.py and requirements.txt from this dataset, and use an owner account or an account with approved access:

python -m pip install -r requirements.txt
hf auth login
python archive.py restore --revision RAW_REVISION --destination "D:\Recovered\journals" --recording RECORDING_ID --report restore-report.json

Replace RAW_REVISION with a full pinned commit from dataset History or the owner metadata snapshot. Repeat --recording for several captures, or use --instrument CATALOG_LABEL --date YYYY-MM-DD. Omitting selection flags restores every catalog capture.

The destination is the journals folder; its immediate children are original recording IDs. Choose an empty destination outside a live recorder directory. Existing files are never overwritten. Restoration does not launch the recorder and does not need the original source location or private name mappings. Every restored physical file is checked against its recorded SHA-256 and length.

Bundles are ordinary TAR containers: already-compressed Zstandard sources stay compressed; other files may use gzip within the TAR. Existing source.zst, commit.json, closure markers, control files and dormant original files retain their layout and bytes. Writer locks are excluded. --deep additionally verifies decoded Zstandard bytes without expanding them onto disk.

To retrieve corresponding canonical daily Parquet, select a recording and add --format parquet. The utility follows the raw catalog's pinned current link and verifies the downloaded event, snapshot and viewer files. Open a file from replay/ in DOM Replay Viewer and use the paired parquet/ files for research. Older raw revisions retain their per-recording replay fallback. Review every returned part for complete selected coverage.

Earlier encrypted journal snapshots

Some earlier snapshots are not matched by newer complete captures. The migration retains them in legacy/earlier-journal-snapshots.tar. Its descriptor, checksum, size, source revision and snapshot count are in the raw catalog. The TAR contains an index, the old encrypted metadata and the original encrypted snapshots. Keep the existing owner recovery key; it is not included in the public dataset.

The owner workflow can restore one snapshot with its already protected credentials:

# From Market_Data_Manager/.support:
.\KCIT_Data_Updater\.venv\Scripts\python.exe manage.py restore-journal --archive archive/PREFIX/SNAPSHOT.zst.enc --output "D:\RecoveredEarlier"

Use an archive filename from the retained index. Before the consolidated bundle is published, supply --revision LEGACY_REPLAY_REVISION to recover from pinned old Hugging Face history instead.

For standalone recovery, load the existing key securely into the process environment variable LEVEL2_RECOVERY_KEY, then run:

python archive.py restore-legacy --revision RAW_REVISION --archive archive/PREFIX/SNAPSHOT.zst.enc --destination "D:\RecoveredEarlier" --cache "D:\RecoveryCache"

Do not put the key or token in a command line, script or shared log. This action downloads and checks the consolidated TAR, authenticates the encrypted metadata and selected snapshot, decodes the journal, verifies its source bytes, and restores its closure/run/control metadata. It refuses existing targets. This is recovery of retained earlier source material, not a claim that it is continuous with another recording.

Daily checks, full verification and local retention

The laptop's Preview → Inspect → Publish workflow prepares one managed candidate per collection. The one-time migration builds one day at a time and saves completed-day checkpoints; subsequent updates reuse unchanged history.

Normal publication checks that captures are closed, uploaded raw bundles match their sizes/checksums, source bytes have not changed, current replay coverage is accounted for, and private recovery mappings/pinned revisions are synced. Eligible local journals, redundant bundles and superseded local files then go to the Windows Recycle Bin. Already-backed-up captures are reconciled too. Locked, changed, unmatched or failed items remain protected/pending.

A full restore is a separate explicit action, for example:

python archive.py verify --revision RAW_REVISION --destination "D:\RestoreTest\journals" --deep --report verification.json

--compare-source optionally compares original physical sources. These checks can download the entire selected archive and use temporary space. The standalone utility never deletes original recordings. A successful upload/checksum check and a successful restore/open test have separate evidence labels; neither proves market-data accuracy.

Never delete or recycle history.db, in any case. Owner cleanup refuses that file and folders containing it. Source databases, application files, credentials, recovery keys and unmatched unique history remain protected. There is no permanent-delete fallback and the workflow never empties the Recycle Bin. Recycled files continue to consume SSD space until the owner empties it.

If interrupted, retain the candidate/checkpoints and retry the same operation. Failed restores can leave partial output: inspect it and choose a fresh empty restore destination. Recovery depends on retaining the hosted revision, account access and the owner key needed by earlier encrypted snapshots.