# Learned position-conditioned exit/hold protocol

Defined before fitting or inspecting candidate results.

## Question
Can a small fitted continuation-value head improve the existing MLP's exits, and does position context add information beyond market inputs?

## Fixed design
Use the current BTC winners, unchanged, on all five timeframes. Generate position context from the simulated candidate's own entries, never from exported TradingView position columns. The entry opportunities and trailing stop remain canonical. Fit a regularized linear value head with small nonlinear state interactions; this provides an interpretable low-variance first learned policy given the limited number of independent trades.

Predict the incremental after-fee wealth of holding for four more bars (or until the existing trailing stop/cap fires), compared with selling now. Fees follow the existing 0.5% per side convention; the entry fee is sunk at this decision. Targets can see future bars only in training and carry explicit label-end timestamps. Historical training rows include baseline in-position states plus hypothetical post-score-exit continuations up to 12 bars, stopping on a hard trailing stop. Weight each originating trade equally to avoid one long trade dominating the fit.

A score exit may be postponed by at most 12 bars; it is remembered and reevaluated every bar. Hard stops and the extension cap override all learned instructions. Primary variants: market-only veto, position-only veto, combined veto. A separately reported combined manage variant can also exit early on ordinary holding bars. Control: baseline; constant 12-bar score-exit delay under the same cap and stop logic. All heads use standardized inputs fitted on training rows only, ridge penalty 1.0 on the weighted mean squared error, and a zero incremental-value threshold. These constants are not selected using evaluation performance.

Expanding head fits before each configured historical evaluation window and before 2026-03-01. Purge any training target whose realized endpoint reaches the evaluation start. Require at least 20 distinct training trades; otherwise abstain to baseline. Start each evaluation replay flat with warm market indicators, applying the same boundary to the paired baseline. Save fold models, training sample/trade counts, cutoff proofs, predictions and actions, all paired metrics and source/input hashes.

## Interpretation and decision

Follow-up declared after the first ridge screen: add a depth-2 regression tree (at most four leaves; at least 10% of training trade weight in each leaf; deterministic seed 0) using the same data, target, purge, threshold and rollout. Compare tree market-only veto, combined veto, and combined manage. This is a limited nonlinear-capacity check, not a claim of unseen confirmation: the ridge screen's evaluation periods have already been inspected. No depth or threshold search will be performed. Initial ridge combined-veto results improved P&L in only 1/30 historical windows, motivating this check before attributing failure to the entire learned-policy hypothesis.
Historical windows are **conditional exit-head tests, not leakage-free end-to-end WFO**: the frozen upstream winners were trained across their history. The 2026 period is later than the recorded model training cutoff but was already inspected in earlier experiments and has limited trades; do not call it pristine validation.

Advance only if combined inputs beat both baseline and market-only on aggregate historical paired performance, improvements are not concentrated in one timeframe/window, drawdown does not materially worsen, and later-period trade evidence is sufficient. If a promising historical pattern lacks later evidence, recommend a strictly nested full-pipeline test or untouched forward paper sample before Pine. If learned policies fail broadly or state inputs add no incremental value, stop this branch and recommend the next hypothesis supported by event-level opportunity diagnostics. No automatic promotion, Pine export, large parameter sweep or production change.