Koraki · field note 01

The first exits

For four weeks the six books were indistinguishable — same picks, same bars, every position open. That ended this week. Two names split the books apart, in opposite directions, and the experiment produced its first real observation.

Closes through 27 Aug 2026 10 picks · 21 trading days · 1 cohort complete

What happened

Two positions closed. FTAI was stopped out by all four rule books five days after entry. META lost only trailing, which quit on a pullback the other three held through — and the name then recovered.

Those are the two cases the design was built to tell apart. FTAI is a stop working: the rules cut a loser and hold kept falling. META is a stop failing: the rule exited a winner and watched it come back. Both happened in the same week, which is a useful accident — it means neither result can be read as the whole story.

The books, paired against hold

Fixed
+$452
primary · pre-registered
Ratchet
+$452
exploratory
Trailing
+$303
exploratory
Hybrid
+$452
exploratory
$0 $150 $300 $450 +$452 +$452 +$303 +$452 FIXED RATCHET TRAILING HYBRID
Dollars ahead of hold, per pick per $10k · at rule exit · n=1 cohort. Trailing is $149 behind the other three, and that gap is one position.

Read this before reading the numbers above. Every dollar of that +$452 comes from FTAI. One position, one cohort, no interval. The pre-registered read is at n=40 with clustered confidence intervals, and this is n=1 — the figure is a plumbing check confirming the paired-difference machinery works, not a result.

Case one — the stop working

FTAI, added 21 August, carried the highest implied upside in the set at +71.7% against a $360 target. It fell 7.6 points of relative strength in five trading days and hit the −7% stop. All four rule books closed at −$302. hold, which has no stop, is still in it at −$754.

0 −7.0 stop level entry · 21 Aug all four books stop out here hold · −7.8, still open DAY 0 DAY 5
FTAI · relative strength, entry to stop. Dashed because only the marked points are recorded — entry, the stop crossing, and today. The line between them is drawn, not measured.

The rules saved $452 per $10k here, which is the entire paired difference in the experiment so far. It is also the single most favourable case a stop can be handed: a fast, clean, one-directional loss with no recovery.

Worth logging while it is fresh. FTAI had the widest implied upside of any pick — +71.7%, against a set spanning +13.7% to +44% otherwise — and it was the first name stopped out, in five days.

That is the vol-scaled stop argument from the handoff, now with a case attached: a flat 7-point stop is proportionally tightest on the highest-conviction call. It belongs in LESSONS.md now rather than later — written after a seventh book exists, the same observation reads as justification rather than evidence.

Case two — the stop failing

META ran to roughly +7.0 RS. trailing's stop followed the peak up to 0.0, the name pulled back through it, and trailing closed at −$102. The other three rule books were nowhere near their −7% stop and held. META then recovered to +3.4 RS, and those three now sit at +$142.

0 +7 +3.4 −2.5 trailing stop, ratcheted to 0.0 entry · 30 Jul peak +7.0 trailing exits −2.5 +3.4 today fixed · ratchet · hybrid never came close to −7 DAY 0 DAY 21
META · the whipsaw case. Trailing surrendered 7 points from the high, which put its stop at breakeven, and ordinary volatility did the rest. Marked points only; the connecting path is drawn, not measured.

This is the failure mode written into the lineup panel before any data existed: trailing converts loses the back half of winners into never reaches the back half. It arrived on the best-performing name in the set, which is exactly where it costs the most.

Note that trailing is not behind because it is a worse idea. It is behind because the 7-point width is narrow relative to META's volatility against SPY. Same rule, wider band, different outcome — which is why the lineup note says that single parameter probably determines more of this book's result than its logic does.

Did the picks beat the index?

Separate question from whether the rules helped, and the one hold answers. Four of ten picks are ahead of SPY. The mean is −0.8 points.

−6 −4 −2 0 +2 +4 FTAI EQIX C SW CL FWONK NVDA BKR META IFF −7.6 −3.8 −2.7 −2.1 −0.6 −0.5 +0.2 +1.9 +3.4 +3.9 RELATIVE STRENGTH vs SPY, POINTS
Ten picks, entry to 27 August. Four ahead of the index, six behind, one outlier carrying most of the spread.

The number that actually matters here. Drop FTAI and the mean across the remaining nine is −0.03 points. Not roughly flat — flat to within a hundredth of a point against SPY.

So the picture after a month is: the picks as a group have done nothing, one name lost badly, and the aggregate is entirely a story about that one name. Which is what a sample of ten over four weeks should look like whether or not an edge exists.

PickAddedDaysReturnSPYRelative
FTAI08-215−7.1%+0.5%−7.6
EQIX07-3021−0.2%+3.7%−3.8
C08-0418−2.9%−0.2%−2.7
SW07-3021+1.6%+3.7%−2.1
CL08-0319+1.0%+1.5%−0.6
FWONK08-0715−1.0%−0.5%−0.5
NVDA08-1113+0.0%−0.2%+0.2
BKR08-0715+1.4%−0.5%+1.9
META07-3021+7.3%+3.7%+3.4
IFF08-0616+4.1%+0.1%+3.9

What this does and does not tell you

Established

The paired-difference machinery works end to end. A position opened, diverged across six books, closed on a rule, and produced a number measured against hold on an identical price path. That had never happened before this week, and it is the thing most likely to be quietly broken in a design like this.

Not established

Anything about whether the rules add value. One cohort, one contributing position, no confidence interval — the clustered interval needs at least two entry-date cohorts to exist at all, and the pre-registered read is at forty positions. Nothing here should move anyone's belief in either direction.

Worth watching next

The time stop. EQIX, SW, and META are at 21 days with no adds. The clock fires at 50, so around mid-October the first completions arrive from duration rather than price. Those are a different kind of observation — they measure what happens when a pick simply does nothing, which is the most common outcome in any pick set and the one a stop-based rule has the least to say about.

Whether trailing keeps lagging. One whipsaw is noise. Three or four on different names, while fixed holds through, starts to be the parameter-width story rather than a bad week.

Throughput against the kill criteria. Ten picks in four weeks is roughly on pace for the fifteen-completed-positions threshold at month twelve — but completions, not entries, are what that criterion counts, and only two have completed so far.