research: validate SOXL volatility-delever hysteresis before any live change

[Pigbibi]

Trigger

The Longbridge SG history showed a volatility delever on 2026-07-09 followed by a rapid 70 percent SOXL re-entry on 2026-07-10. This points to missing hysteresis or cooling, not proof that a generic hard stop is beneficial.

Candidate set

• baseline current dynamic P95 volatility threshold with bounds 50 to 75 percent
• 5 percentage-point hysteresis with one-session confirmation
• 7.5 percentage-point hysteresis with two-session confirmation

Required evidence

• full available history from 2016 with realistic transaction costs
• frozen pre-change baseline and exact parameter manifest
• walk-forward and rolling out-of-sample results
• separate 2020, 2022 and 2026 event windows
• reject candidates with material out-of-sample drawdown or return regressions
• no live parameter, plugin authority or broker change until a valid P1 and P3 evidence chain exists

Current blocker

SOXL P1 data is deferred with ALPACA_SIP_ACCESS_FORBIDDEN, so P3 comparison data is unavailable. Restore compliant data access or an approved equivalent before forward validation.

[Pigbibi]

Preliminary historical sweep is now in PR #391. Using the archived 2010-09-30 to 2026-08-21 market-regime price input with an explicit 0% cash proxy and 5 bps turnover: 5pp/1d is rejected; 7.5pp/2d improves max drawdown from -49.55% to -48.17% and Calmar from 0.969 to 0.996, while CAGR decreases from 48.04% to 47.99%. This remains research-only pending BOXX/BIL-backed replication and a predeclared OOS split; no strategy parameter changed.

[Pigbibi]

Long-term compounding risk-budget sweep is now in PR #392 (research-only; no runtime change). Using the same 5 bps cost and zero-return cash proxy, from 2010-10-01 to 2026-08-21:

• current 70/65: CAGR 48.06%, MDD -49.55%, Sharpe 1.078, Calmar 0.970
• P95 re-entry hysteresis 7.5pp/2d, same cap: 48.01%, -48.17%, 1.080, 0.997
• 60/55 plus the re-entry control: 42.27%, -42.79%, 1.080, 0.988
• 50/45 plus control: 36.23%, -37.06%, 1.080, 0.978

60/55 plus control is the only cap that improved Calmar over baseline in both this history and the separate 2016-06-07 to 2026-08-21 replay. It remains a candidate only pending BOXX/BIL-backed and locked-OOS validation.

[Pigbibi]

Update: the BIL-filled / actual-BOXX repeat and fixed-spec annual stability diagnostic are in PR #393. This reduces the strength of the prior 60/55 recommendation:

• 2010-10-01 to 2026-08-21, current 70/65: CAGR 48.89%, MDD -49.45%, Calmar 0.988.
• 60/55 + 7.5pp/2d: CAGR 42.45%, MDD -42.65%, Calmar 0.995.
• The cash-backed replay therefore makes 60/55 a risk-preference trade-off, not an unconditional improvement.

Fixed-spec complete-calendar-year diagnostic for 2016-2025: 60% risk-adjusted win rate, median excess Calmar +0.007, but worst excess CAGR -29.21pp (2017), versus a hard -3pp limit. Gate fails. The artifact is explicitly research_only and not promotion eligible; Longbridge runtime must remain unchanged.

[Pigbibi]

Data-entitlement update: PR #394 is merged and the manual non-live SIP diagnostic run 32711627096 returned ALPACA_SIP_ACCESS_OK for SOXL, SOXX, and BOXX (as well as QQQ, TQQQ, QQQM) at the 2026-08-21 cutoff. No bars were persisted by the diagnostic.

This removes the authorization blocker for the existing source-frozen SOXL P1/P3 daily research lane, whose next scheduled run will continue to be data-only and research-only. It does not change the #393 conclusion: the 60/55 + cooldown candidate failed its annual stability gate and is not eligible for automatic promotion or Longbridge runtime changes.

[Pigbibi]

已完成研究控制面的跟进：\n\n- 已确认 2026-08-22 的 SOXL P1 终态为 DEFERRED / ALPACA_SIP_ACCESS_FORBIDDEN；该数据权限问题随后已由 SIP 诊断验证为恢复。\n- 同日组合就绪任务 32571482813 因缺少所需 TQQQ P1 终态工件而失败；这不影响 Longbridge 实盘，但原有监控没有把该类终态/工件异常形成明确告警。\n- 修复已随 PR #395 合并到 main（2e3bb514）：缺失终态工件会被收敛为显式状态；DEFERRED、工件缺失或就绪任务失败会创建/更新一条去重的 research-only 告警；只有两路脱敏 P1/P3 终态恢复到就绪状态才自动关闭。\n- 严格边界：不重试数据采集，不调整策略，不触碰 paper/shadow/live/订单权限。\n\n下一次计划任务会在下一交易日收盘后的 2026-08-25 02:45 UTC 运行，用于验证恢复后的完整 P1/P3 链路。