Before I re-codify the strategy population under engine v2, I need policy answers — not "what to build" but "how faithful should the v2 codification be to the original source." Each section settles one piece of the policy: the headline tier (strict / source-bounded / permissive), per-component rules, the tuning protocol, the v1-vs-v2 verdict logic, and edge-case handling.
Each task carries a default recommendation. YES accepts the default. MODIFY if you want one of the alternatives (or your own approach). NO means "don't apply this rule at all." Decisions auto-save. Submit when done.
The headline choice in Section 1. Determines how "faithful to source" v2 codification is:
| Policy | Behavior | Trade-off |
|---|---|---|
| Strict | v2 matches source's literal description. RSI(14) stays RSI(14). No tuning at all. | Most defensible. Most likely to under-perform on your specific data. |
| Source-bounded | Source defines the kind of strategy; engine tunes the specifics. RSI period tuned within reasonable bounds, validated on OOS. | Default. Most pragmatic. Source structure locked, parameters tuned. |
| Permissive | Source is a direction; engine substitutes its own choices when it has reason to. May replace RSI with a related indicator if better empirically. | Highest upside, highest drift risk. Hardest to claim "this is the source's strategy." |
signal_locked=true strategies (SMA-4, Turtle S1, Donchian breakout) override this — they are always strict regardless of the section 1 answer.
The version bump that justifies a graveyard re-trial:
| Component | v1 (April harness) | v2 (today) |
|---|---|---|
| Timeframe stack | Single TF | Multi-TF mandatory · explicit roles per TF |
| Indicator set | ~8 indicators | 33 indicators · cross-TF interactions |
| Gate layer | Mostly absent | 8 categories · first-class predicates |
| Position mgmt | Hardcoded constants | Empirical per (strategy, instrument) |
| Cost model | Flat fee + slippage | Path-dependent · funding accrued per interval |
| Regime | After-the-fact label | Vector · first-class gate input · bucketed reporting |
5 of 6 components are major version bumps. This is the trigger event for a one-time graveyard re-trial.
signal_locked: true strategies have their signal envelope frozen forever — protected from automatic "improvement."Once submitted, Claude reads your responses and starts executing. You can still come back and edit, but rerunning the build manually after edits is on you.