H69 KILLED: a significant spread is not a Sharpe

The strategy H68 owed. Built with zero new parameter search -- same
symbols, same signal, same 90-day window, same 90th-percentile threshold,
same horizons -- and engine-graded with next-open fills and 10bp fee +
1bp half-spread + 25bp participation impact per side. All three gates
fail. Ledger 164 -> 167.

H68's S2 spread was +3.17% per 7 days with a CI excluding zero and
breadth 17/20 on 31,752 symbol-days. That is not a marginal result. The
book built from it earns Sharpe 0.89 and is beaten by equal-weight
buy-and-hold at 0.94, recomputed in the same run over the identical
window.

The reason is the transferable part. The info-grade bar has no variance
term: E[fwd|HIGH] - E[fwd|LOW] with a CI excluding zero says the
conditional means differ and says nothing about the volatility a book
must carry to collect the difference. The qualifying set averages ~1.4
names, so harvesting a panel-wide mean requires a concentrated book. The
mean survived; the ratio did not. Proposed and not adopted: report an
info SIGNAL alongside the volatility of its bucket, because two identical
spreads are not equally valuable if one is carried by 15 names and the
other by 1.4.

Gate B earned its place, and this is the part worth keeping. A1 and A2
passed -- +9.176bp/day with a CI excluding zero, CAGR +30.30% -- and on
those two bars alone this reads as a +30%/yr success. Section 2 recorded
before any code existed that the window's unconditional 7-day return is
+0.92%, roughly 60%/yr, so a long-only crypto book would clear a
profitability bar regardless of whether the signal worked, and that Gate
B existed entirely for that reason. Without it this verdict would have
recorded a beta proxy as a win.

The Section 5.2 correlation prediction was right: 0.3-0.7 predicted
before the run, +0.3929 measured. Meta-finding 5 holds and this is not an
independent edge.

Scope correction for H67's proposed tail-conditional bar: it added
nothing here (worst decile -1.575%, worst 1% -2.648% vs +0.092%
unconditional) because the linear correlation bar already failed at
+0.39. H67's blindness was specific to short-convexity payoffs driven by
squared returns, so that amendment should be scoped to asymmetric
payoffs rather than applied universally.

Recorded without being acted on: at hold=1 the book earns +39.06%/yr
while deployed only 33.0% of the time, roughly 118%/yr per unit of time
in market against buy-and-hold's 53%. The signal concentrates return but
not risk-adjusted return. 2023 (+91.48%, Sharpe 2.46) and 2024 (+94.14%,
Sharpe 2.60) are excellent and slicing them out post hoc is exactly what
pre-registration forbids; H68 section 8.6 pre-flagged 2025 as its weak
year before this study existed, and 2025 (-19.53%) and 2026 (-47.79%)
confirmed it.

An implementation deviation was found and corrected BEFORE the verdict
was written: the first run fed the engine the lake's full extent from
2017-08-17 rather than the declared 2019-01-01 start, letting the
benchmark trade the 2017 bubble (+838%/yr) and the 2018 crash while the
strategy sat in signal-less cash. Correcting it moved the benchmark
Sharpe 0.86 -> 0.94 and the strategy 0.82 -> 0.89, WIDENING the Gate B
gap rather than narrowing it. Both sets of figures are in the verdict so
the correction is auditable.

H68's SIGNAL verdict is untouched. "The effect is real" and "the effect
is tradable in this shape" are different claims and the ledger holds
both.

Engineering:
- Signal construction extracted to features/crossvenue.py so H69 consumes
  the SAME rows H68 measured rather than a second hand-rolled copy -- the
  defect the panel audit recorded when it found six copies of the
  trailing-percentile helper. H68's golden verified byte-identical after
  the refactor.
- strategies/crossvenue.py holds the ladder and the buy-and-hold
  benchmark. The ladder is stateless: exactly `hold` entry days are live
  at any time and each belongs to a distinct tranche, so replaying the
  same dates always produces the same weights.
- tests/test_crossvenue_strategy.py covers the weights contract, tranche
  roll-off, and the case that matters operationally -- a symbol that
  stops trading mid-hold frees its slot to cash instead of silently
  concentrating the book, which is XRP's Coinbase halt.

618 passed, ruff clean, mypy strict clean.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

Author: martex-dev

PROJECT_MEMORY.md

@@ -3,7 +3,7 @@
 The knowledge file: ledger, results, meta-findings, lessons, open
 assumptions. PROJECT_STATE.md = what runs now; this = why and what we know.
 
-## Trial ledger: 164 registered (163 run, 1 data-blocked: H54). Every new
+## Trial ledger: 167 registered (166 run, 1 data-blocked: H54). Every new
 ## spec raises the DSR bar. Do not test without a numbered doc FIRST.
 ##
 ## SOURCE OF TRUTH for the ledger is docs/research/ledger/trials.toml, not
@@ -13,7 +13,8 @@ assumptions. PROJECT_STATE.md = what runs now; this = why and what we know.
 ## momentum family), H63 conditional carry ELIGIBLE (Sharpe 6.00, replaces
 ## H62 as the carry spec), H64 cointegration KILLED, H65 wide-universe
 ## carry STANDALONE-VIABLE, H66 cross-sectional carry STANDALONE-VIABLE,
-## H67 variance risk premium KILLED, H68 cross-venue dislocation SIGNAL.
+## H67 variance risk premium KILLED, H68 cross-venue dislocation SIGNAL,
+## H69 the strategy built on H68 KILLED.
 
 ## Hypothesis ledger (docs/hypotheses/, docs/research/)
 
@@ -181,6 +182,34 @@ assumptions. PROJECT_STATE.md = what runs now; this = why and what we know.
     a hypothesis whose returns are concentrated before 2024 should say so
     in its verdict.
 
+13. **A significant spread is not a Sharpe — the info bar has no
+    variance term (H69).** H68's S2 spread was **+3.17% per 7 days**, CI
+    excluding zero, breadth 17/20, on 31,752 symbol-days. The strategy
+    built from it with **zero new parameter search** — same symbols, same
+    threshold, same window, engine-graded with next-open fills — earns
+    **Sharpe 0.89** and is **beaten by equal-weight buy-and-hold (0.94)**.
+    **Why:** `E[fwd|HIGH] − E[fwd|LOW]` with a CI excluding zero says the
+    conditional means differ and says nothing about the volatility needed
+    to collect the difference. The qualifying set averages **~1.4 names**,
+    so a panel-wide mean has to be harvested through a concentrated book.
+    The mean survived; the ratio did not.
+    **Proposed, not adopted:** report an info SIGNAL alongside the
+    **volatility of its bucket**. Two identical spreads are not equally
+    valuable if one is carried by 15 names and the other by 1.4.
+    **The benchmark bar earned its place.** H69's A1 and A2 passed
+    (+9.176bp/day, CAGR +30.30%); on those alone it reads as a +30%/yr
+    success. Only the buy-and-hold comparison separated that from the
+    truth, and it was pre-registered specifically because the window's
+    unconditional 7d return is +0.92%. **Any long-only crypto book tested
+    on 2019-2026 needs a buy-and-hold bar or the verdict is meaningless.**
+    Correlation with rotation-stop **+0.3929**, inside the 0.3-0.7 band
+    predicted before the run — meta-finding 5 again.
+    **Scope note for finding 9's proposed tail bar:** it added nothing
+    here, because for a *directional* book the linear correlation bar
+    already caught the dependence. H67's blindness was specific to
+    **short-convexity** payoffs, so that amendment should be scoped to
+    asymmetric payoffs rather than applied universally.
+
 12. **The Coinbase premium is not the tether peg, and cross-venue
     dislocation is the first new INFO signal in a long time (H68).** The
     peg-adjusted premium between a USD venue (Coinbase) and USDT venues

PROJECT_STATE.md

@@ -506,7 +506,7 @@ presented as a discovery.
 
 ---
 
-## H68 — cross-venue dislocation SIGNALS; F2 is open and owes a strategy (2026-08-27)
+## H68 — cross-venue dislocation SIGNALS; H69 shows it is not tradable (2026-08-27)
 
 `OBSERVATION` — pre-registered (commit 3eb0fbc) before any study code
 existed; run the same day. **Ledger 152 -> 164.** Family F2's kill test.
@@ -561,17 +561,56 @@ significance, so the consistent sign is the meaningful part. But 2025 is
 a real wobble and the strategy follow-up must confront it rather than
 average it away.
 
-### What F2 owes next
-
-**A pre-registered strategy hypothesis on S2.** It faces the
-**incremental** bar against the deployed book, not zero. Open questions
-it must answer: Sharpe after costs; correlation with rotation-stop; and
-whether the h=1 cell — cheapest to trade at ~22bp round trip — survives
-once it is a return stream rather than a bucketed mean.
-
-**H68 deploys nothing and makes nothing paper-eligible.** Info grade
-only; no DSR deflation has been applied, and the N=164 burden lands on
-the follow-up.
+### What F2 owed, and how it turned out — H69 KILLED (2026-08-27)
+
+**The strategy hypothesis was registered (34b11e5), built and run. All
+three gates fail.** Ledger 164 -> 167.
+
+| Book | CAGR | Sharpe | MDD | DSR@167 | in market |
+|---|---|---|---|---|---|
+| hold = 1d | +39.06% | 0.90 | −73.24% | 0.0901 | 33.0% |
+| **hold = 7d (primary)** | **+30.30%** | **0.89** | −50.01% | **0.3598** | 69.3% |
+| hold = 30d | +25.23% | 0.85 | −54.76% | 0.4623 | 91.7% |
+| **equal-weight buy-and-hold** | +52.97% | **0.94** | −80.23% | 0.0005 | 100.0% |
+
+A1 (mean > 0, CI excludes zero) and A2 (CAGR ≥ 2%) **pass**. A3 (Sharpe
+≥ 1.0), A4 (DSR ≥ 0.95), B5 (beat buy-and-hold) and C6 (|corr| < 0.30,
+measured **+0.3929**) all **fail**.
+
+`INTERPRETATION` — **a significant spread is not a Sharpe.** H68's
++3.17%/7d had a CI excluding zero and 17/20 breadth on 31,752
+symbol-days. The book built from it with zero new parameter search loses
+to owning the basket. The info-grade bar has no variance term: a
+conditional-mean difference says nothing about the volatility required to
+collect it, and the qualifying set averages ~1.4 names, so a panel-wide
+mean must be harvested through a concentrated book.
+
+`OBSERVATION` — **the benchmark bar is what caught this.** On A1 and A2
+alone the book reads as a +30%/yr success. §2 of the pre-registration
+recorded, before any code existed, that the window's unconditional 7-day
+return is +0.92% and that Gate B existed entirely for that reason.
+
+`OBSERVATION` — the §5.2 correlation prediction (0.3–0.7) was **right**
+at +0.3929. Meta-finding 5 holds; this is not an independent edge and
+does not count toward the eight-edge target.
+
+`OBSERVATION` — recorded and **not acted on**: at hold = 1 the book earns
++39.06%/yr while deployed only 33.0% of the time (~118%/yr per unit of
+time in market, vs buy-and-hold's 53%). The signal concentrates return
+but not risk-adjusted return. 2023 (+91.48%, Sharpe 2.46) and 2024
+(+94.14%, Sharpe 2.60) are excellent; **slicing them out post hoc is
+exactly what pre-registration forbids**, and H68 §8.6 pre-flagged 2025 as
+its weak year before this study existed — 2025 (−19.53%) and 2026
+(−47.79%) confirmed it.
+
+**H68's SIGNAL verdict is untouched.** "The effect is real" and "the
+effect is tradable in this shape" are different claims and the ledger
+holds both. Nothing is deployed; nothing is paper-eligible.
+
+**F2 is no longer the obvious next move.** Reopening it means proposing a
+different vehicle — the premium as an overlay or filter on the deployed
+book rather than a standalone long book — which is a new
+pre-registration with a stated reason, not a rescue of H69.
 
 ---
 

docs/hypotheses/69-cross-venue-premium-strategy.md

@@ -1,7 +1,13 @@
 # Hypothesis 69 — Cross-Venue Premium, Strategy Grade (family F2)
 
-Status: **PRE-REGISTERED 2026-08-27. NO RESULT EXISTS.** Trials declared:
-**+3 → 167.** Verdict will be written into §8 and nowhere else.
+Status: **KILLED (2026-08-27).** Trials: **+3 → 167.** Verdict in §8.
+All three gates fail. H68's info signal does **not** survive contact with
+execution: the book earns +30.30%/yr at **Sharpe 0.89** and loses to
+simply owning the basket (0.94). §2's warning about a bull sample was the
+right warning, and Gate B — written for exactly this — is what caught it.
+The pre-registered correlation prediction (0.3–0.7) was **correct** at
++0.3929. H68's verdict is untouched; a real info signal has been shown
+not to be a tradable edge in this shape.
 
 The strategy hypothesis `docs/hypotheses/68-cross-venue-dislocation.md`
 §8.7 said was owed. H68 found the peg-adjusted USD-vs-USDT venue premium
@@ -249,6 +255,161 @@ the amendment has not been adopted.
 
 ---
 
-## 8. VERDICT
-
-*(Not yet run. This section is written only when the study executes.)*
+## 8. VERDICT (2026-08-27, scripts/h69_cross_venue_strategy.py, +3 → 167)
+
+**KILLED.** Gate A fails (A3, A4), Gate B fails, Gate C fails. Per §6,
+*"A fails → KILLED"* — and it would have closed on the §6 beta branch
+even had Sharpe cleared 1.0, because Gate B failed independently.
+
+Window **2019-01-01 → 2026-07-10** as §4.4 declared, 20 symbols, engine
+fills at the next bar's open with 10bp fee + 1bp half-spread + 25bp
+participation impact per side. 889 entry days carried at least one
+qualifier.
+
+### 8.1 The three declared cells, plus the Gate B benchmark
+
+All run in the same process, over the identical window.
+
+| Book | CAGR | Sharpe | MDD | mean bp/day | 95% CI (bp) | DSR@167 | in market |
+|---|---|---|---|---|---|---|---|
+| hold = 1d | +39.06% | 0.90 | −73.24% | +12.643 | [+1.627, +23.124] | 0.0901 | 33.0% |
+| **hold = 7d — PRIMARY** | **+30.30%** | **0.89** | **−50.01%** | **+9.176** | **[+1.683, +17.588]** | **0.3598** | 69.3% |
+| hold = 30d | +25.23% | 0.85 | −54.76% | +7.621 | [+0.170, +16.084] | 0.4623 | 91.7% |
+| **equal-weight buy-and-hold** | **+52.97%** | **0.94** | −80.23% | +20.712 | [+3.022, +38.236] | 0.0005 | 100.0% |
+
+### 8.2 The six bars
+
+| Gate | Bar | Measured | Result |
+|---|---|---|---|
+| A1 | mean > 0, CI excludes zero | +9.176bp, low +1.683bp | **PASS** |
+| A2 | CAGR ≥ 2%/yr | +30.30% | **PASS** |
+| A3 | Sharpe ≥ 1.0 | **0.89** | **FAIL** |
+| A4 | DSR ≥ 0.95 @167 | **0.3598** | **FAIL** |
+| B5 | Sharpe > buy-and-hold | **0.89 vs 0.94** | **FAIL** |
+| C6 | \|corr\| rotation-stop < 0.30 | **+0.3929** (n=2,652) | **FAIL** |
+
+### 8.3 The finding: a significant spread is not a Sharpe
+
+`OBSERVATION` — H68 measured the S2 spread at **+3.17% per 7 days**, CI
+[+0.51%, +6.08%], breadth 17/20, on 31,752 symbol-days. That is not a
+marginal result. The book built from it — same symbols, same threshold,
+same window, **zero new parameter search** — earns Sharpe **0.89** and is
+beaten by owning the basket equally weighted.
+
+`INTERPRETATION` — **the info-grade bar has no variance term, and that is
+the gap this hypothesis exposes.** `E[fwd | HIGH] − E[fwd | LOW]` with a
+CI excluding zero says the conditional means differ. It says nothing
+about the volatility a book must carry to collect that difference. Here
+the qualifying set averages ~1.4 names, so harvesting a panel-wide mean
+requires holding a concentrated book whose realized volatility is far
+above the panel average. The mean survived; the ratio did not.
+
+**Proposed as a general lesson, and it is cheap to act on:** an
+info-grade SIGNAL should be reported alongside the *volatility of the
+bucket*, not only the difference of its means. Two signals with identical
+spreads are not equally valuable if one is carried by 15 names and the
+other by 1.4. Stated as a proposal for the standard info template, not
+adopted by this verdict.
+
+### 8.4 It is beta, and §2 said so before the run
+
+`OBSERVATION` — §2 recorded, before any strategy code existed: *"the
+unconditional forward 7-day return over this window is +0.92%, roughly
+60%/yr… a long-only crypto book will pass a 'is it profitable' bar in
+this window almost regardless of whether the signal works. §5 Gate B
+exists entirely because of this line."*
+
+`OBSERVATION` — A1 and A2 **passed**: mean +9.176bp/day with a CI
+excluding zero, CAGR +30.30%. On those two bars alone this looks like a
+success.
+
+`INTERPRETATION` — and it is not one. Gate B is what separates the two
+readings, and it fired: the timing **subtracts** risk-adjusted return
+relative to owning everything. Without Gate B this verdict would have
+recorded a +30%/yr strategy as a win. **The bar earned its place**, and
+that is worth more than the result it killed.
+
+### 8.5 The correlation prediction was right
+
+`OBSERVATION` — §5.2 predicted, before the run: *"Gate C will probably
+fail… predicted correlation 0.3–0.7."* Measured: **+0.3929**.
+
+`INTERPRETATION` — meta-finding 5 holds again. Every long-crypto book in
+this ledger sits 0.52–0.82 against every other; this one is 0.39 against
+rotation-stop, at the low end but well outside the 0.30 bar. **Not an
+independent edge**, and it does not count toward the eight-edge target.
+
+`OBSERVATION` — the tail-conditional statistic proposed in H67 §8.4,
+reported here only (the amendment is **not** adopted): on rotation-stop's
+worst decile H69 returns **−1.575%**, worst 5% **−1.968%**, worst 1%
+**−2.648%**, against +0.092% unconditional.
+
+`INTERPRETATION` — useful scope information for that proposal. Here the
+**linear bar already saw the dependence** (+0.39, failing), so the
+tail check adds nothing. H67's blindness was specific to a
+**short-convexity** payoff driven by squared returns. **The proposed
+amendment should therefore be scoped to asymmetric payoffs, not applied
+to every hypothesis** — for a directional book the existing bar works.
+
+### 8.6 What the signal did do, recorded without being acted on
+
+`OBSERVATION` — at hold = 1 the book earns **+39.06%/yr while deployed
+only 33.0% of the time**, against buy-and-hold's +52.97%/yr at 100%.
+Per unit of time in market that is roughly **118%/yr vs 53%/yr**.
+
+`INTERPRETATION` — the signal **does** concentrate return into the days
+it is on, by more than 2×. What it does not do is concentrate
+*risk-adjusted* return, and this project cannot lever an edge that failed
+its own validation. This is recorded because it points at a differently
+shaped question — the premium as an **overlay or filter on the deployed
+book** rather than a standalone long book — and **that would be a new
+pre-registration with a stated reason, not a rescue of this one.**
+Time in market is also not the same as leverage-adjusted exposure, so the
+118% figure is an illustration, not a Sharpe claim.
+
+### 8.7 Per-year, including the part that is tempting and must be refused
+
+| Year | n | H69 | Sharpe | buy-and-hold |
+|---|---|---|---|---|
+| 2019 | 274 | +28.38%/yr | 1.18 | −0.22%/yr |
+| 2020 | 366 | +70.43%/yr | 1.81 | +182.58%/yr |
+| 2021 | 365 | +23.64%/yr | 0.53 | +289.27%/yr |
+| 2022 | 365 | −12.93%/yr | −0.29 | −93.70%/yr |
+| **2023** | 365 | **+91.48%/yr** | **2.46** | +84.53%/yr |
+| **2024** | 366 | **+94.14%/yr** | **2.60** | +124.69%/yr |
+| **2025** | 365 | **−19.53%/yr** | −0.70 | −13.54%/yr |
+| **2026** | 191 | **−47.79%/yr** | −1.14 | −45.54%/yr |
+
+`OBSERVATION` — 2023 and 2024 are excellent on their own (Sharpe 2.46 and
+2.60, beating buy-and-hold in 2023). 2025 and 2026 are bad.
+
+`INTERPRETATION` — **this is exactly the slice that must not be acted
+on.** Selecting 2023–2024 after seeing the full result is the failure
+pre-registration exists to prevent, and the ledger has a standing rule
+that near-misses stay closed. Recorded because §7 required the per-year
+table, and because H68 §8.6 **pre-flagged 2025 as its weak year** — that
+flag was written before this study existed and it was right, which makes
+the 2025–2026 weakness a confirmed property rather than bad luck.
+
+### 8.8 Caveats that survive the kill
+
+- **§0 stands: this was never out-of-sample.** It shared H68's data,
+  parameters and window, so it could only ever have answered "is it
+  tradable". It answered no.
+- **An implementation deviation was found and corrected before this
+  verdict was written.** The first run fed the engine the lake's full
+  extent from 2017-08-17 rather than §4.4's declared 2019-01-01 start,
+  which let the benchmark trade the 2017 bubble (+838%/yr) and the 2018
+  crash while the strategy sat in signal-less cash. Corrected to the
+  registered window; the fix **helped** the benchmark (Sharpe 0.86 →
+  0.94) and helped the strategy less (0.82 → 0.89), so it widened the
+  Gate B gap rather than narrowing it. The pre-fix figures are recorded
+  here so the correction is auditable.
+- **MDD −50% on the primary and −73% at hold = 1.** Whatever else this
+  book is, it is not a low-drawdown vehicle.
+- **The 25bp participation-impact model is the project standard, not a
+  measurement**, and it is applied to concentrated positions in names as
+  small as WLD and ENA. It is more likely optimistic than pessimistic.
+- **H68's verdict is untouched.** The info signal remains a SIGNAL; what
+  is now also on the record is that it does not survive execution in a
+  long-only ladder. Those are different claims and the ledger holds both.

docs/research/family-expansion-program.md

@@ -128,8 +128,16 @@ found the peg-adjusted USD-vs-USDT premium predicts forward returns:
 momentum and concentration checks. The peg alone is noise, so the
 "Coinbase premium" is **not** tether. Dispersion (S3) and peg deviation
 (S4) are CLOSED as noise.
-**What is still owed:** a pre-registered strategy hypothesis facing the
-incremental bar against the deployed book. **What is still untested:**
+**H69 answered what was owed, and the answer was no.** The strategy built
+on S2 with zero new parameter search is **KILLED** (2026-08-27): Sharpe
+0.89 against equal-weight buy-and-hold's 0.94, DSR 0.36, correlation
++0.39 with rotation-stop. The info signal does not survive execution in a
+long-only ladder. H68's SIGNAL verdict stands — "the effect is real" and
+"the effect is tradable in this shape" are different claims.
+**F2 has now spent 15 of ~30 trials.** What remains open in it is a
+differently *shaped* question (the premium as an overlay or filter on the
+deployed book rather than a standalone long book), which needs its own
+pre-registration and a stated reason. **What is still untested:**
 the arbitrage version — the p90 premium is 11bp against a 3.17% forward
 spread, so nothing here is harvested by closing a gap, and genuine
 cross-venue arbitrage still needs order-book data and latency this
@@ -193,10 +201,12 @@ Liquidation events are mechanically forced selling with a known trigger.
    It was run ahead of F1's data-collection tail because its kill test was
    cheap (a free index, no chain needed) and it gated the largest build in
    Tier 1. That gate is now shut.
-5. **F2 cross-exchange** — **started and open (H68, 2026-08-27).** The
-   multi-venue dataset now exists in `data/venues/`. The kill test found
-   a signal; the strategy hypothesis is the next thing this family owes,
-   and it is currently the highest-value open thread in the program.
+5. **F2 cross-exchange** — **started; kill test SIGNAL (H68), strategy
+   KILLED (H69).** The multi-venue dataset exists in `data/venues/` and
+   the information is real, but it did not survive execution as a
+   standalone long book. The family is no longer the obvious next move:
+   reopening it means proposing a different vehicle, not another pass at
+   the same one.
 
 Tier 3 waits until Tier 1–2 is exhausted; it is gated on infrastructure,
 not on ideas.