# BTC6H fixed entry-loss result — 2026-09-21

Reject the fixed 10% entry-loss policy and stop this experiment. It failed the unchanged 40% drawdown limit and incumbent Calmar floor. No additional percentages, seeds or thresholds were tried; nothing was promoted or deployed.

| IS metric | Frozen incumbent | Plus 10% entry-loss exit |
|---|---:|---:|
| Return | 2846.94% | 1522.29% |
| Close-equity drawdown magnitude | 50.50% | 55.92% |
| Calmar | 1.0045 | 0.7188 |
| Entries | 59 | 63 |
| Exposure | 40.52% | 37.19% |
| First subperiod Calmar | 0.6092 | 0.1952 |
| Second subperiod Calmar | 2.7671 | 2.6460 |

Both policies have zero trades and exposure in the already inspected OOS period. Their OOS values provide no positive validation. Both receive zero eligible P&L/DD because drawdown exceeds 40%; equality of those zero values does not rescue the candidate. Fragility, positive subperiod and existing eligible-fold WFO checks pass, but the complete promotion predicate fails. The candidate's low-trade 2022 fold loses 11.57%; it is retained in the evidence even though the existing WFO eligibility rule excludes it. This is a retrospective result, not untouched validation.

The candidate produced eight loss-only exits, three trailing exits and 52 score exits, versus seven trailing and 52 score exits for the baseline. No simultaneous exit flags occurred. Observed close overshoot beyond the 10% trigger ranged from 0.155 to 2.564 percentage points, before fees. Eight recovery events have complete 1/4/7-day horizons: zero, one and two respectively recovered their entry close. Loss exits changed subsequent entries as well as individual losses; the complete path performed worse. All event-level re-entry delays, fee amounts, completed-trade outcomes and both actual recovery paths remain in diagnostics.

## Verification and evidence

The preregistration is [the frozen protocol](mlp_entry_loss_protocol_2026_09_21.md). Mandatory parameter/Pine checks passed. The disabled policy exactly reproduced canonical actions, returns, fees, exposure and reported metrics. The runner evaluated exactly two policies under one contract, with distinct policy identities. Independent reconstruction from both 12,927-bar traces reproduced IS returns and drawdown and checked every loss trigger. All 941 protected files retained their hashes. The implementation's fast suite passed 236 tests with two slow cases skipped before execution.

Raw evidence is in `results/mlp_entry_loss_2026_09_21/`. A byte-verified archive is committed at `results/research_checkpoint_2026_09_21/mlp_entry_loss_2026_09_21.tar.gz`, with member hashes in `entry_loss_manifest.json`. Extract from the repository root with `tar -xzf results/research_checkpoint_2026_09_21/mlp_entry_loss_2026_09_21.tar.gz`.

Proceed with the isolated measurement fix for review. Keep research checkpoints on the experiment branch and leave main unchanged pending review. No candidate qualified for the conditional TradingView/forward-observation step; no new search follows automatically.