Research brief B-2 from the 2026-08-17 scout, filed as implementation tracking. Prerequisite-gated on the Phase 10 re-measurement (#339), same rationale and same honest prior (the null) as #341.
Triple-barrier exit with a vertical time stop, barriers sized to friction: keel's per-kind exits are already stop/target/signal; what is genuinely new is (i) a holding-duration exit no kind has, (ii) barrier distances as explicit multiples of round-trip friction rather than ATR/structure multiples, and (iii) the label-balance grid method for choosing the vertical barrier. Evidence: Grądzki et al. 2025 (24-period vertical barrier + ±2.5–5% barriers beat next-bar labeling; vol-adjusted barriers did NOT help); arXiv 2504.02249 (Korean stocks 2006–2024, vertical barrier grid-searched 5–29 days for label balance — no costs modeled; label balance is not P&L); Alvarez 2019 (keel's shape — indicator OR N-day time stop at next open — is the industry baseline; the limit-order half that helped there is unavailable to keel's market fills).
keel-specific weaknesses to design against: the source's 2.5–5% barriers sit at or below keel's 2.5% round trip — mechanically dead until scaled several-fold; the time-stop leg IS executable under market fills ("sell at close after N bars"); per-kind stops already swept flat; #259's per-product slippage now prices thin books up to 50bp, which the barrier sizing must incorporate.
Research brief B-2 from the 2026-08-17 scout, filed as implementation tracking. Prerequisite-gated on the Phase 10 re-measurement (#339), same rationale and same honest prior (the null) as #341.
Triple-barrier exit with a vertical time stop, barriers sized to friction: keel's per-kind exits are already stop/target/signal; what is genuinely new is (i) a holding-duration exit no kind has, (ii) barrier distances as explicit multiples of round-trip friction rather than ATR/structure multiples, and (iii) the label-balance grid method for choosing the vertical barrier. Evidence: Grądzki et al. 2025 (24-period vertical barrier + ±2.5–5% barriers beat next-bar labeling; vol-adjusted barriers did NOT help); arXiv 2504.02249 (Korean stocks 2006–2024, vertical barrier grid-searched 5–29 days for label balance — no costs modeled; label balance is not P&L); Alvarez 2019 (keel's shape — indicator OR N-day time stop at next open — is the industry baseline; the limit-order half that helped there is unavailable to keel's market fills).
keel-specific weaknesses to design against: the source's 2.5–5% barriers sit at or below keel's 2.5% round trip — mechanically dead until scaled several-fold; the time-stop leg IS executable under market fills ("sell at close after N bars"); per-kind stops already swept flat; #259's per-product slippage now prices thin books up to 50bp, which the barrier sizing must incorporate.
Rulesubclass + registry entry (human-reviewed); measured on An hourly-cadence paper profile — evidence at a collectable rate #337's hourly profile with Pool min_trades across same-parameter rules, with a cross-sectional diversity floor #338's pooled counting; the DCA ablation's fee-drag result (in Re-measure under the honest cost model: simulate re-run, DCA dip ablation, paper DCA re-enabled #339) disciplines the friction-sizing thesis before anyone builds.