Run Episode
The orchestrator entry point for the Public Portfolio Challenge. Every episode is a specific job over specific real artifacts, but the discipline is the same and lives in the functional skills. This skill picks the right runbook, pins its artifacts, and runs the stages in order, delegating each to the skill that owns it. It does not re-implement the discipline — it sequences it.
Invocation
/run-episode <episode>[ <attempt>]
Examples: /run-episode 10 · /run-episode 10 addendum · "run the episode-10 bakeoff".
First: load the actual runbook (it is the source of truth for THIS run)
Read the target runbook and treat it as authoritative for the job + artifacts (this skill only encodes the ordering and delegation):
| Episode | Runbook | Shape |
|---|---|---|
| 10 | episode-10/BAKEOFF_RUNBOOK.md | Multi-family bakeoff (search→certify→lockbox→deploy) |
| 10 / addendum | episode-10/addendum/RUNBOOK.md | Redesign entries + exits, prove sell behavior and retain the incumbent edge |
From the runbook, pin before doing anything: the SUBJECT (live book / build) IDs, the incumbent bar (cert study id + per-fold numbers), the fixed universe + capital, and any carried-over artifacts. The runbook's own Stage list wins if it disagrees with the generic ordering below.
The bookend rules (fixed for every episode)
- Stage A first — confirm the SUBJECT by FIELD (
get_portfolio→conditionFieldAudit), re-verify the incumbent bar, spot-check one fresh backtest, and (if live) check pending orders +automaticOrderApproval. Deliver a one-paragraph before state. - Deploy last, and GATED — no clone, no reconcile, no orders until the human says the episode's explicit go phrase ("deploy + clean up"). See deploy-gate.
- Bugs are a deliverable at every stage — bug-protocol on any failure; quarantine tainted results.
- Verify, don't assert — every engine-behavior claim checked against events/fields/repro (via portfolio-certification's first principles) before it lands in the log.
Episode 10 — the bakeoff sequence
Delegates, in order:
- engine-sanity — S0 contract checks. STOP on any failure. (These don't count toward the cert body.)
- lockbox-holdout — load A/B/C baselines as OOS bars (S1); run the S1.5 gate-coherence auto-relax.
- strategy-bakeoff — the SEARCH→CERTIFY funnel: ≥3 families, per-family ledger, verdict-integrity. Each promoted seed is certified through walk-forward-oos + breadth-audit, held to options-structure-rules (hard gates) and sweep-reoptimization (re-sweep from a clean seed).
- lockbox-holdout — the single-touch lockbox confirmation on the frozen finalist (S2).
- deploy-gate — GATED. Discover the live target via
fetch_portfolios, owner-confirm, clone, field-verify, re-attach monitoring.
Episode 10 addendum — entry and exit redesign
The addendum follows the same certification spine with a redesigned middle:
- Stage A baseline + inventory (bookend rule above).
- Entry + exit redesign: current-engine authoring/sweep inventory → matched entry/exit mechanism families → event-level exit-binding audit → sweep-reoptimization on the promoted design → fixed-config walk-forward-oos vs a fresh control.
- options-structure-rules hard-gate + breadth-audit gate on every finalist.
- Present, then stop — verdict up top, per-fold OOS table, side-by-side vs incumbent, provenance labels, honest caveats, bug ledger. No deploy.
- deploy-gate — GATED on the explicit go, for the named finalist only.
What this skill guarantees
- The bookends are never skipped (Stage-A field-verify first; deploy gated last).
- Each stage runs through the skill that owns its checks, so the verdict is holistic OOS, breadth is at fixed $25k, structural changes force a re-sweep, and every parameter carries provenance.
- A null result is a valid outcome — the orchestrator does not torture variants to manufacture a win.
If the target runbook adds, renames, or reorders stages, follow the runbook — this skill is the default spine, not an override.