austin-starks/Public-Portfolio-Challenge

Sweep Re-Optimization & Provenance

The three standing methodology rules that exist because a deploy candidate was once certified on knobs inherited from a sweep on a different structure and presented as "optimal."

The three standing rules (binding on all future episodes)

1. Re-optimize on ANY structural change — never inherit a genome across structures. Knobs (TP, total budget, per-name size, rank window, VIX gate, entry cooldown, DTE bracket) are structure-dependent. Whenever the structure changes — affordability rungs, strike depths, universe membership (e.g. adding a name), DTE family, entry/exit shape, or adding an alt-data rank signal — re-run the engine_kind:"sweep" walk-forward on the NEW book before certifying. A backtest_only cert on inherited knobs only confirms the book generalizes; it does not establish the knobs are good for that structure.
2. Separate "explore designs" from "optimize the chosen design." Screen candidate designs → pick one → then sweep that exact design's knobs → then backtest_only-certify. If the screen picks design X, the sweep base must be X itself, not its neighbours. Never sweep the losers' variants while leaving the chosen book hand-tuned.
3. Label every deployed parameter's provenance in the deliverable — swept-on-this-book / inherited-from <study-id> / hand-set. "Inherited" is amber, not green, until re-swept on the current structure.

Self-check trigger words: "carry over the knobs," "reuse the winner's params," "same config on the new ladder," "already certified, no need to re-sweep," "just raise the allocation." Each must prompt: did the structure change since those knobs were searched? If yes → re-sweep before certifying.

Running the sweep

1. get_sweep_surface on the chosen book (or a clean clone) FIRST to get the real sweepable field names. Re-sweep from a CLEAN seed — re-sweeping an already-swept book can stack/duplicate genes; field-audit for duplicated conditions.
2. run_walk_forward_study with engine_kind:"sweep", inner_mode:"optimize", certification:true, mode:"validation", anchored, 5 folds (4 acceptable), 2022→today, oos_width_days:252. See walk-forward-oos for the full param block. GA overfits and is rejected for deploy certification — sweep is the certified path. preview_only:true first to compile genes + cost-check.
3. Each sweep must be non-trivial: ≥3 gene_intents axes × ≥3 values each (e.g. BuyingPowerPct, TakeProfitPct, RollTriggerDte, rank depth, delta band, DTE). A 1-axis or 2-value sweep does not qualify.

Exact sweep field mappings (get these wrong and the sweep is meaningless)

• Take-profit = TakeProfitPct on Action scope. Never sweep it as ExitCondition/OptionPositionPercentChange.
• Roll DTE = RollTriggerDte on Action scope (roll the CloseOption only).
• CloseOption pnl triggers: minPnlPercent:100 = +100% TP; maxPnlPercent:-50 = −50% stop; DTE roll = {type:"dte","maxDte":45}.

Selecting the winner (robust, not argmax)

Read aggregate.crossFoldRobustSelection — metric:"minSortino", candidateKey, score, perFoldSortino: the grid cell with the highest minimum validation Sortino across folds. Compare it against each fold's selectedCandidateKey, aggregate.dominantCandidateKey, and winnerStableAcrossFolds. Deploy the cross-fold-robust config, never the per-fold argmax — the per-fold winner is high-variance noise.

Then run the full walk-forward-oos Stage-C checks 1–7 on that winner before it may be the finalist, and attach provenance labels (rule 3).

Custom indicators are NOT sweepable

The sweep engine's allowedIndicatorTypes does not include CustomIndicator, so alt-data indicator-shape exploration is a hand-built variant grid (create_portfolio_variant: deep-copy base + JSON-Pointer patch, dry_run first), each certified backtest_only and labeled hand-set (amber). The classic knobs (alloc / budget / TP / SelectTop / DTE) remain sweepable if a winner warrants a knob re-sweep. See alt-data-indicators.