feat(rules): triple-barrier exits with a vertical time stop, and the measurement (#342)

[eaitbrahim]

Research brief B-2, prerequisite-gated on #339 (closed 2026-08-17). A sixth RULE_REGISTRY kind, and the A/B it exists to make possible.

The result, in two numbers

cusum_event and triple_barrier share an entry — same CUSUM filter, same threshold, same universe, same window. Only the exit differs, so the difference is attributable.

	zero fee	1.2% taker
control (cusum_event) median PF	0.925	0.343
triple_barrier median PF	1.001	0.338
median delta	+0.033	−0.004
assets improved	17 of 24	11 of 24 (a coin flip)

The exit genuinely works. At zero fee it lifts the median profit factor across break-even and improves 17 of 24 assets — the first number in this series of documents to move the right way.

And it is worth nothing. At the rate the account pays, the median delta is −0.004. The gross gain is smaller than the friction it has to be harvested through.

Zero of 24 clear PF 1.0 at the taker rate. The only two cells above 1.0 anywhere are TON-USD at n=17 and n=16 — a sixth of the admission floor, maxima of five draws on the thinnest asset. The intersection is empty across all 192 trials.

What's new in the rule

• A vertical barrier — sell at the close after N bars, whatever price has done. No shipped rule has a holding-duration exit, and it's the leg of the source's method that transfers: executable under market fills, where a resting order at a horizontal barrier is not.
• Barriers as multiples of the product's own round trip (Friction-sized triple-barrier exits with a vertical time stop — prerequisite-gated on the Phase 10 re-measurement #342's design ask). The source's ±2.5–5% sit at or below one round trip here, so transplanted as percentages they're mechanically dead. A fixed percentage is also wrong per asset — since slippage is one global 5bp constant applied from BTC to TON #259 the backtest prices thin books to 183.8bp per leg, so the same 5% is four round trips on BTC and barely one on the tail.

Two things worth your eye

A unit trap that produces no error. median_daily_quote_volume returns a per-bar median despite its name, and slippage_for_quote_volume is anchored on $500M daily. Handing it the hourly figure unscaled reports every asset as maximally thin, clamps the universe to the 183.8bp cap, and makes every barrier four times too wide — silently. per_product_round_trip scales by bars-per-day; a test asserts the consequence rather than the call.

A bug the mutation run found in my own guard. if not stop < entry < target passes a negative stop: a large enough stop_friction_mult drives the barrier below zero, the ordering still holds, and risk = entry − stop then exceeds the entry itself — sizing a position off a loss larger than the whole holding. It's 0 < stop < entry < target now. Unreachable inside the declared param space, reachable from a stored row, which is why it isn't dead code.

The vertical axis

bars	median n	PF median	above 1.0
6	520	0.161	0
12	486	0.243	0
24	431	0.338	0
48	374	0.392	1
72	338	0.464	1

Monotone, and the source's own 24-bar barrier is mid-range rather than optimal. The direction agrees with its "wide barriers beat next-bar labeling" finding; the magnitude doesn't, because 1.2% per leg is twelve times the 0.1% that paper priced.

What this settles

Both halves of Grądzki et al. are now implemented on keel's cost structure and measured on one universe: the entry half has essentially no gross edge, and the exit half is a real improvement friction consumes entirely.

The paper is not wrong about its own venue — at 0.1% per leg a +0.033 gross improvement is worth keeping. At 2.5% it is not. That's the clearest measurement in the series of the difference between a result and a result at a price.

Verification

9 mutants, 9 killed. Full suite 5241 passed / 3 skipped; ruff and mypy clean. Ledger verifies at 92 rows.

Nothing promoted, no rule row added, config and allowlist untouched. The null grows: 0 of 90 → 0 of 114 → 0 of 138.