Files
Aura/docs/specs
Brummel 1e2ac82473 spec: 0097 IS-refit walk-forward over a loaded blueprint (boss-signed)
Add `aura walkforward <blueprint.json> --axis <name>=<csv> [--axis …]`: a true
in-sample-refit walk-forward over a loaded blueprint (Arm A, #173). Per IS
window it re-optimizes the blueprint's params over the user's --axis grid,
selects the winner by sqn_normalized, runs it out-of-sample, and aggregates the
per-window OOS reports into a FamilyKind::WalkForward family that is
content-addressed (put_blueprint) and aura-reproduce-aware. Closes the World/C21
milestone's last piece (the walk-forward half of #170).

Design (Arm A settled by the user; sub-forks derived and recorded on #173): the
loaded-blueprint WF mirrors the hard-wired stage1-r walk-forward arm one
substitution deep — the generic engine walk_forward driver + select_winner +
the 24/12/12 synthetic schedule are reused verbatim; only the per-window IS-sweep
and OOS-run source the loaded blueprint (wrap_stage1r via cycle-0096's
blueprint_axis_probe) over the user --axis grid. Persist + reproduce ship
together (coupled — a persisted family with no store hook is unreproducible, the
#170 hazard); the reproduce WalkForward branch reconstructs each OOS window's
windowed source + winner params from the member manifest, mirroring the 0095
MonteCarlo branch. Loaded members are reduce-mode (R-metrics, no raw pip curve),
so the WF is measured in R (oos_r), stitched pip-equity empty — consistent with
the C10 R-first yardstick.

Boss-signed on grounding-check PASS (independent, fresh-context; all 8
load-bearing current-behaviour assumptions ratified by named green tests —
incl. an empirically-confirmed one: a loaded-blueprint member carries and ranks
by sqn_normalized. One spec-internal tension flagged + fixed before sign).
Engine/registry untouched (invariants 1/2/8/9).

refs #173
2026-07-01 15:16:35 +02:00
..