# Claude Validator: Feb->Jul Retune + Rolling Walk-Forward Verdict

Reviewer: `l_ClaudeBTCPhaseLiquidMoneyOps1`
Date: 2026-07-16
Task from Taras: make the liquid-money strategies as profitable on
2026-02-14 -> 2026-07-16 as the old (pre-Feb) champions were on their window.

## What Was Produced (in-sample, the literal ask)

Codex tune `mature_reparam_tune_20260716\run_20260716_172624` (Feb14->Jul16,
1x/1x, fee 0.0005, slip 0.000938, min step 2.5) gave live sleeves:
PYTH +165.8%, HYPE +142.6%, SEI +119.0%, RENDER +108.2%, ROSE +97.7%,
SOL +83.6% (buy-hold negative for 4 of 6).

Portfolio on those curves (my run of `research_per_wheel_suspension_allocator.py`,
out `per_wheel_suspension_mature_retuned_20260716`, curve map
`curve_map_run_20260716_172624.json`):

- best tuned: `core3 / per_wheel_0035` **+177.79%, MDD -4.70%, stag 14d**
- identity (no allocator tuning): +156.95% / -7.17%
- equal-weight 6 sleeves (no allocator at all): **+134.37%, MDD -2.86%**
- half-split balanced: ~+55..67% each half -> portfolio layer is not the
  overfit source; the sleeves are.

## Honest Validation (rolling walk-forward, params never see the fold)

Tunes: train_feb_apr (run_20260716_180340), train_feb_may (run_20260716_174950),
train_feb_jun (run_20260716_181050); same costs; EW-6 portfolio; two models:
fresh = fold starts with reset equity; carry = replay from Feb 14, score fold
segment only (state/cushion carried).

| fold | fresh total | fresh MDD | carry total | carry MDD |
|------|------------|-----------|-------------|-----------|
| May  | +2.72%     | -8.55%    | +3.26%      | -5.16%    |
| Jun  | -20.82%    | -23.74%   | +1.70%      | -5.40%    |
| Jul (to 16th) | -5.07% | -6.90% | +2.31%    | -0.82%    |
| **chained May->Jul** | **-22.79%** | | **+7.44%** | |

Key structural finding: the strategy family (stubborn adds, loss_cut -65%,
long holds) is strongly path-dependent. Full-window champions fresh-started
on Jun 1 lose -8..-72% per sleeve on the very segment their own backtest
shows as profitable. The good "second half" of any tuned curve rides on the
profit cushion built earlier. Start-date risk is first-order.

Side test: causal SMA10/20/30 price gate (sleeve to cash below own SMA)
cuts train profit ~3x and improves holdout only marginally -- not worth it.

## Verdict

1. The +177.8% Feb->today chart is REAL arithmetic but IN-SAMPLE: sleeve
   params were tuned on the full window. It must not be sold as forward
   expectation -- this is exactly how the old +725% was born and died.
2. The defensible process is: monthly retune + CONTINUOUS operation +
   equal-weight sleeves (no allocator tuning needed). Its honest OOS
   estimate on May->mid-Jul: **+7.4% per 2.5 months (~+2.9%/mo), MDD ~-5%**,
   with no fold blowing up.
3. Fresh-starting capital at an arbitrary date is the dangerous mode
   (-22.8% chained). Any paper/live launch should expect this asymmetry.
4. Jun-Jul 2026 is a hostile regime for this family regardless of tuning
   window; no parameter choice found by the tuner would have earned there.

## Recommendation To Codex Lane / Taras

- Treat `per_wheel_0035` (+177.8%) as an upper bound, not a forecast.
- If deploying paper: EW-6 on the retuned params, continuous, small size,
  monthly retune cadence, and a portfolio-level MDD kill at ~-10%.
- Report metrics as combined MTM (they are: replay_live_like total_pct).
- Related: see REVIEW_btc_phase_negative_equity_leak_20260716.md -- the
  +1518% BTC-phase champion is an artifact (negative-equity leak), and
  `replay_live_like` still needs a bankruptcy floor; that fix does not
  affect the runs above (all sleeves stayed positive) but blocks reuse of
  older ENA/UNI-template artifacts.

## Artifacts

- Curve map: `_reports\liquid_lunar_solar_overnight_20260712\mature_reparam_tune_20260716\curve_map_run_20260716_172624.json`
- Portfolio run: `_reports\liquid_lunar_solar_overnight_20260712\per_wheel_suspension_mature_retuned_20260716\`
- WF tunes: `mature_reparam_tune_20260716_train_feb_apr|_train_feb_may|_train_feb_jun`
- Holdout replay CSV: `mature_reparam_tune_20260716_train_feb_may\holdout_jun1_jul16_replay.csv`