archive: v0.3.0 snapshot (MTF + 5-pair portfolio, clean Sharpe 1.07)

[luokerenx4]

Summary

v0.3.0 run completed. 39 rounds on autoresearch/apr24, tagged v0.3.0 at commit 8d007d1. Archives meta.yaml + results.tsv + 3 alive strategies + retrospective.md + run.log. Index in versions/README.md updated.

Headline

First project-wide clean Sharpe > 1.0. BTCLeaderBreakX reached Sharpe 1.07 / DD -8.8% / all 5 pairs positive, with zero Goodhart attempts. v0.2.0's clean peak was 0.67 — a +60% improvement in ~half the rounds.

Three-version progression:

• v0.1.0 true-edge Sharpe: 0.19 (headline 1.44 was exit_profit_only in bull regime, agent's own sanity check confirmed)
• v0.2.0 clean Sharpe: 0.67
• v0.3.0 clean Sharpe: 1.07

Behavioral milestones in this run

• First fork event in project history. v0.1.0 + v0.2.0 = 180 rounds, zero forks. v0.3.0 r13: agent forked BTCLeaderBreak to test a compound change (Donchian 20→15 + exit SMA50→SMA20). Compound change failed, but instead of abandoning, agent used r14 to isolate by reverting ONE of the two changes. Result: isolated the Donchian-15 entry as the real improvement. Textbook A/B decomposition executed autonomously.
• v0.2.0 MACDMomentum 0.67 peak reinterpreted as BTC/ETH-specific overfit. Same stack in 5-pair universe tops at 0.41. Motivates an errata.md addition to versions/0.2.0/ in a separate follow-up PR.
• Cross-pair volume asymmetry finding: for cross-pair strategies, volume confirmation on the TRADE pair >> SIGNAL-source pair. Impossible to produce without both cross-pair affordance AND per-pair reporting.
• Paradigm-theory-level exit asymmetry: "ride the move" paradigms (breakout, volatility) benefit from patient exits; "manage the trend" paradigms need responsive exits. Agent's words.
• Zero Goodhart attempts for the third consecutive run. Multi-axis comparison + per-pair reporting + prior-retrospective awareness continue to serve as defense.

Scope

Archive-only. Touches versions/0.3.0/ and versions/README.md index only. No scaffold changes.

Test plan

• Skim retrospective.md for factual correctness — peak numbers match results.tsv, aha moments match commits
• meta.yaml numbers match results.tsv
• strategies/*.py in archive match user_data/strategies/*.py at commit 8d007d1
• Tag v0.3.0 resolves on GitHub to commit 8d007d1
• Consider: follow-up PR adding versions/0.2.0/errata.md (per retrospective recommendation — out of scope for this archive-only PR per immutability rule)

🤖 Generated with Claude Code