# BTC 6H bounded threshold comparison — September 21, 2026

**No configuration qualifies. The fixed-budget search is complete and stopped.** No model, winner, production parameter, Pine preset or deployment was changed. No candidate advances to forward observation, so candidate parity exports and new-observation evaluation are not triggered by this result. The incumbent remains installed but still fails the current drawdown requirement.

## Comparison and stop

Ten models: the incumbent dense seed4942, plus grouped, random-sparse and hybrid seeds 9090/9191/9292. Each received exactly 1,000 identical additional configurations and one saved baseline: 10,010 completed evaluations. RNG seed 20260921; wide generator ranges (entry −600 to <600 in steps of 5; exit/confirmation −750 to <750 in steps of 10; trailing stop 0 to <100 in steps of 5), with the existing shared MVRV and smoothing proposals. Both confirmation booleans were balanced: 250 configurations per pair, per model. The ordinary generator's extra plain artifact was excluded before evaluation.

The corrected evaluator, shared historical data, accounting and every gate were retained. All rows were preserved. Runtime was approximately 383 seconds with four workers. There are no adaptive rounds, additional seeds or promotions. A preliminary attempt was interrupted before results were inspected because the generator pinned confirmation booleans. Its workers were confirmed stopped; only `final/` is authoritative. This disclosed aborted work is outside the final equal budget and did not inform selection.

| Model | Best-ranked drawdown | Calmar | Trades | Qualifiers among 1,001 |
|---|---:|---:|---:|---:|
| incumbent | 49.82% | 1.019 | 43 | 0 |
| grouped_v1_seed9090 | 58.82% | 0.754 | 34 | 0 |
| grouped_v1_seed9191 | 63.10% | 0.559 | 35 | 0 |
| grouped_v1_seed9292 | 74.16% | 0.556 | 64 | 0 |
| random_sparse_v1_seed9090 | 48.72% | 0.961 | 98 | 0 |
| random_sparse_v1_seed9191 | 56.24% | 0.741 | 49 | 0 |
| random_sparse_v1_seed9292 | 0.00% | -10.000 | 0 | 0 |
| hybrid_grouped_v1_seed9090 | 31.87% | 1.076 | 94 | 0 |
| hybrid_grouped_v1_seed9191 | 62.38% | 0.581 | 63 | 0 |
| hybrid_grouped_v1_seed9292 | 43.61% | 0.850 | 58 | 0 |

These are the existing evaluator's best-ranked rows, not each model's minimum drawdown. Ranking prioritizes positive Composite before positive P&L/DD. The cash-like random-sparse9292 row has no trades and a −10 Calmar sentinel; it is not a successful low-risk outcome.

The hybrid9090 near miss has 31.8726% drawdown, Calmar 1.0760 and 94 trades, versus the frozen incumbent's 50.4976% and 1.00449. It clears all preserved conditions except `Fragile_0.05=2`. Its confirmation threshold −730 lies within 0.05 score units of two historical scores (2018-01-26 and 2019-02-09). This is the existing proximity diagnostic, not demonstrated sensitivity of actual fills. We did not move the threshold, weaken the diagnostic or rerun its neighborhood.

A different grouped9292 row reaches 10.09% drawdown and Calmar 1.384 but fails fragility and WFO coverage. Drawdown alone therefore does not explain every rejection. Across the full result set, 298 rows have positive eligible P&L/DD, 426 clear the WFO condition, and six meet the incumbent Calmar floor (including its frozen baseline); their intersection with all remaining gates is empty. WFO_Neg_Folds=6 can be the insufficient-coverage sentinel, not six observed negative folds. Full period measurements are retained.

## Loss diagnosis

Fixed-parameter replays of the frozen incumbent, its best-ranked tuned row and hybrid9090 reconstruct their reported drawdowns to within 1e-8 percentage points. These are diagnostic replays, not additional threshold trials. Bar traces include exposure, executed actions, raw exit opportunities and canonical net returns; monthly and trade summaries are retained.

- **Frozen incumbent:** peak 2017-12-07 18:00 UTC; trough 2019-12-18 06:00; recovery 2020-04-30 00:00. Peak-to-trough exposure was 14.76%, with six entries and six exits during the interval. Drawdown includes giving back an early unrealized peak plus subsequent losses; it is not evidence of constant bear-market exposure. A 2019-08-20 entry held 35.75 days and lost 21.58% net, exiting 22.12% below its trade high with no score-exit opportunity at closure. The 2024-01-11 entry lost 16.62% over 12.25 days, also without a score-exit opportunity. These are consistent with the existing 20% trailing stop being the eventual protection. Sixteen of 21 losing trades did have score-exit opportunities, so this is a concentrated tail problem, not universal absence of score exits.
- **Best-ranked incumbent tuning:** drawdown remains 49.82%, from 2019-06-26 to 2020-01-23, with 56.04% exposure during that interval. Its 35% trail allowed a 138.5-day 2022 hold to lose 25.66%, and a 20.5-day 2017–18 hold to lose 26.02%. Threshold tuning did not establish acceptable downside behavior.
- **Hybrid near miss:** its 31.87% drawdown spans 2021-01-29 to 2021-09-24, with 13.87% exposure, 11 entries and 12 exits (including a carry-in position). September 2021 lost 19.37% with four entries and four exits. Its worst completed trade lost 9.05% in 18 hours despite a 5% trailing threshold. Repeated losses, fees and close-based stop overshoot remain relevant even with lower exposure. This configuration remains ineligible because of fragility.

## One selected follow-up experiment — not executed

Test **one fixed 10% entry-loss exit on the original frozen incumbent**, retaining its existing 20% trailing stop, weights, entry/score-exit thresholds and all other settings. This addresses the observed tail losses without shortening every winning trade's trailing allowance. The 10% value is one predeclared hypothesis, not an estimated optimum or a guaranteed realized-loss bound.

Compare exactly two policies: unchanged incumbent and incumbent plus exit when a held bar closes at or below 90% of entry signal close. Preserve current entry/fill-bar handling, next-interval exposure and additive fees. A close-based trigger must use the observed close, never assume a fill at the threshold. Any implementation must establish control equivalence first and record simultaneous score/trailing/entry-loss triggers separately.

Retain the same full IS, subperiod, WFO, OOS, fragility and incumbent-reference gates, plus all periods' trade counts and exposure. Report missed recoveries and re-entry churn as well as clipped tail losses. Reject the experiment if it fails any gate; do not sweep 5/7.5/12.5% alternatives afterward. This two-policy experiment is selected but not executed; the requested stop after the bounded comparison is respected.

The previously archived volatility-stop experiment concerned BTC4H and changing ATR-based trail widths. It does not establish this BTC6H entry-loss hypothesis. All historical periods here have been inspected; any later success would still require a frozen qualifying artifact, matched chart/trade exports from the same TradingView state, chart validation before parity checks, and genuinely new observations without retuning. No such qualification or forward evidence is claimed now.

## Evidence and verification

Canonical evidence: `results/mlp_threshold_comparison_2026_09_21/final/`.

- `protocol.json`, `configurations.json`: declared budget and every exact configuration.
- `comparison_manifest.json`, `incumbent.json`, `all_results.json`: source/data contract, reference and every result including model hashes and period measurements.
- `summary.json`, `qualifying_configurations.json`, `gate_failure_counts.json`: per-model results, empty qualifier set and condition failures.
- `drawdown_diagnostics.json`, `*_bars.csv`, `*_monthly.csv`: loss timing, exposure and exit evidence.
- `configuration_audit.md`, `result_audit.md`, `completion_audit.json`, `verification_checks.json`: independent audits and requirement checks.

Required sync/Pine preflights passed. The 21 focused comparison/adaptive/accounting tests passed; `git diff --check` passed. Independent audit reproduced every configuration/result correspondence and zero-qualifier decision, checked common context and current input/source hashes, and verified all 924 protected files unchanged. No implementation changes required a new full-suite run.

The historical saved/replayed discrepancy remains separate and unresolved. Shared older data supports this retrospective comparison; it does not supply publication-vintage guarantees, untouched validation or live TradingView execution parity.