Skip to main content

Reproducibility

Reproducibility across profile versions: standard@2026-11 for backtest and live, standard@2027-05 published later; research re-run and shadowed live on the new version, then a switch; the old version served at least 24 months after its successor

PromiseWhat it means for you
The same bars and profile version give the same segments, bit for bitA backtest's segments are exactly the segments live trading would have had
A profile version never changesA result stored with its version can be reproduced at any time
Old versions are served for at least 24 months after a new oneYou choose when to move, and can re-create any result in that time
A new version is announced in the changelogNothing changes under you

Pinning​

Name the version in every request: standard@2026-11, never standard, in research and live. Record it beside every stored result and in the model registry with the model it fed. An edge is sealed with its version, and is refused with another (409 edge_mismatch).

Moving to a new version​

  1. Re-run research on the new version as jobs, into a separate path of your storage.
  2. Re-validate the models that use the segments, as your model risk policy asks.
  3. Shadow live: run the new version beside the old for a period, from its own edges (a finished series of the session so far, with return_edge), and compare.
  4. Switch the instruments to the new version's edges, and keep the old results with their version for audit.

For model risk​

For each profile version, Patternode provides under NDA a description of its inputs, outputs, known limits and changes from the previous version, for your model inventory and validation (SR 11-7, PRA SS1/23). The version string is the identifier to cite.

Next: Operations and compliance.