Repository navigation
REST API
Choose the API that matches your integration. The V2 ASGI app and the V3 V7 service are separate applications; starting one does not start the other.
| Interface | Purpose |
|---|---|
V2 /calculate
|
Stateless field calculation using the supplied history. |
V2 /simulate
|
Audit a supplied mark sheet under assumed performance variation. |
V2 ledger and /v1/shadow routes |
Authenticated stored prediction/settlement evidence. |
V3 /v3/* service |
Authenticated competition lifecycle under the frozen V7 contract. |
| Linux V3 subprocess | Local competition runtime used by STRATHEX; not the V2 HTTP app. |
Install the API extra, then choose an explicit database path.
Linux shell:
mkdir -p api-data
export STRATHMARK_DB_PATH="$PWD/api-data/results.db"
python -m uvicorn strathmark.api:app --host 127.0.0.1 --port 8000Windows PowerShell:
New-Item -ItemType Directory -Force api-data | Out-Null
$env:STRATHMARK_DB_PATH = Join-Path (Get-Location) "api-data/results.db"
python -m uvicorn strathmark.api:app --host 127.0.0.1 --port 8000Open http://127.0.0.1:8000/docs for request examples and /health for model status.
Public /calculate is stateless and unauthenticated; keep this example on loopback.
The stored ledger/shadow routes require their own credentials and configuration.
V7 has 18 routes and an exact OpenAPI checksum. Pin its contract and source identity,
verify signed receipts, and use a bearer service credential plus Idempotency-Key
for trusted POSTs. Loopback is the default; non-loopback service operation additionally
requires pinned mutual TLS. Actor headers are audit metadata, not human permissions.
Pre-field forecasts supply seed times and say issued_mark=false. Exact-field
assembly produces marks. /v3/approvals/decide records review decisions;
/v3/issues/acknowledge is the separate issue action.
Use the V3 route table and examples and consumer migration guide. The frozen OpenAPI and checksum are the integration reference. Windows production qualification remains incomplete.
Home · Quick start · FAQ · Repository
Start here
Understand the predictions
- How handicaps work
- V2
- V3
- Accuracy Preview
- Wood and diameter
- Recency
- Uncertainty and simulation
- Fairness
- LLM integration
- Rulebooks
- Legacy prediction keys
Operate and develop