oak-domain-betting

The betting domain for the OakQuant platform. Right now it is one thing only: a point-in-time odds capture that runs from day one.

pip install -e .
oak-odds snapshot      # every open market on every venue, timestamped
oak-odds status        # what the corpus holds

Signing in

The domain is capability-gated, fail-closed: a tenant whose enterprise does not carry betting sees none of it, and never another domain's content instead.

# inside grove, after the package is installed
python -m oak_domain_betting.provisioning.seed_betting_owner --dry-run
python -m oak_domain_betting.provisioning.seed_betting_owner --password-stdin

Creates the oakquant-betting enterprise, grants it the betting capability, and provisions owner.betting@oakquant.ai as its owner. Idempotent — re-running never duplicates a row and never deletes one.

⚠️ Never pass a password as an argument. It lands in the process table, in shell history, and in any transcript. --password-stdin reads from a pipe or a TTY prompt and never renders it. Omit the flag and the owner stays on the invite flow.

Why capture comes before the model

Everyone has results. Almost nobody has timestamped odds — and without them there is no point-in-time price, no closing line value, and no honest backtest. An hour of prices not captured cannot be reconstructed at any price. This is the same irreversibility that already cost this platform its macro board history, and it is the only part of the programme that has a deadline of now.

Everything else can be written later against the history this is accumulating today: the de-vig, the hierarchy, the features, the gates.

What it does not do

It does not convert a price into a probability. Removing the overround is Phase B0, and proportional, Shin and power methods disagree materially on longshots. Get it wrong and every number downstream is wrong — silently, and in the flattering direction. So nothing here pretends to have done it.

Quote.mid exists and is documented as not a probability. Polymarket's outcomePrices already sum to 1 and are deliberately not promoted to the quote, because a field that looks finished is more dangerous than one that obviously is not.

The venues, and why these two

Both are public and key-free. Nothing here has been paid for, and nothing will be until the retention and use-class terms have been read — a monthly price reads like a purchase and is not one, and a personal-use tier does not cover a billing product.

⭐ Three-way soccer arrives as three binary markets sharing one event ticker. KXEPLGAME-26SEP20FULMUN has a -FUL, a -MUN and a -TIE, and their three YES mids sum to more than 1. That excess is the exchange's overround, sitting in the data, which makes Gate B0 testable the moment the instrument is written: a harness that cannot reproduce the venue's own cut from quoted prices cannot be trusted to find an edge.

What lands on disk

$OAK_ODDS_DIR/                       # default ~/oakquant-odds
  raw/<provider>/<date>/<ts>.json.gz    # verbatim. THIS is the asset.
  canonical/<date>.jsonl                # flat rows. Derived, regenerable.

Raw is written before anything is parsed. If normalisation throws, the irreplaceable half is already safe. Normalisation can be rewritten a year from now and re-run over raw; a field dropped at capture time is simply gone.

⚠️ A skip is not a pass

Five silent losses shipped on this platform in one week, every one the same shape: a loop that skipped records and reported success. So the capture counts what was written, not what was attempted; every skipped market carries a reason; truncation from hitting a page limit is recorded differently from a feed that simply ended; and a run that captures nothing exits non-zero and says so — "NOTHING CAPTURED. This hour is lost and cannot be refetched."

Apache-2.0.