H70: K=2 vindicated on evidence, and the live window says the opposite

Aimed at the deployed spec rather than a new family, because the deployed
spec is still validated and still losing money live, and four consecutive
family searches had produced one standalone-viable edge and three kills.
Ledger 167 -> 170.

Hypothesis 11 line 60 reads "Long-only. K=2 FIXED". It was fixed by fiat
at the start of the rotation family and inherited untested by every
descendant, including the H42b spec running on paper today. H70 varied it
and nothing else. The K=2 recomputation reproduces the published deployed
figures to the digit -- +42.91% / 1.47 / -29.01% against +42.9% / 1.47 /
-29.0% -- so every comparison is against the real incumbent.

Gate A fails on the primary K=5 AND on every declared cell: MDD is worse
than K=2's at K=3, K=5 and K=8 alike. The verdict therefore does not
depend on which cell was primary, which makes this a vindication rather
than an inconclusive result. K=2 keeps its place on evidence for the
first time.

All three pre-registered predictions were wrong. Sharpe runs 1.47 / 1.61
/ 1.40 / 1.27 and CAGR 42.91 / 46.23 / 34.53 / 27.13 across K = 2/3/5/8 --
both peak at K=3 -- and MDD never improves on K=2. H66's carry finding
does not transfer, and the mechanism is mechanical: carry HARVESTS a
premium paid by ~20 near-independent funding streams, so averaging more
cuts variance without cutting the mean, while rotation SELECTS and the
4th-8th ranked coins are worse assets rather than more draws of the same
edge. This is the first evidence that H65's select-vs-harvest distinction
is real, arriving from the opposite direction after H66 withdrew it, and
it stays a hypothesis rather than a rule -- one measurement in each of two
families is exactly the state that produced the withdrawn refinement.

K=3 beats the incumbent on Sharpe (1.61 vs 1.47) and CAGR (+46.23% vs
+42.91%) and loses only on drawdown. It is NOT adopted: it was not the
declared primary, it fails the registered MDD bar -- the bar meta-finding
8 says decides prop-firm outcomes -- and 0.14 of Sharpe across adjacent K
on one path is not a measurement of the K surface. Its numbers go to the
owner per the charter rather than being suppressed or acted on.

The live window points the other way, and that is the most useful thing
here. Replaying every cell over 2026-07-10..2026-08-26 on
data/lake-current at L=90, which is what the paper account actually runs,
gives K=2 -6.06%, K=3 -7.54%, K=5 -3.85%, K=8 -0.85%, with MDD improving
monotonically from -15.48% to -7.09%. So concentration accounts for
roughly five of the six points rotation-stop gave up live: the drawdown
is not purely bad selection luck, it has a structural component and that
component is K. And this does not license changing K. Forty-eight days is
1.7% of the evidence behind the 2,880-day backtest, which says K=8 earns
27%/yr against K=2's 43% with a worse drawdown. Switching because the
last seven weeks favoured another setting is textbook recency-chasing.

A pre-declared inference is withdrawn in the verdict. Section 6 said a
Gate A failure would mean concentration does not explain the live
drawdown. It does not follow: the bars are computed on the frozen
backtest and cannot answer a question about the live window. The verdict
stands on the bars; the further claim does not. The sharpened open
question for the H59 divergence hunt is whether the live period is
unrepresentative or the K surface has moved, and that needs forward time
rather than another slice of the same history.

Also honoured here, and overdue: the standing commitment in
family-expansion-program.md section 5 to re-validate the deployed book as
N grows, last met at 125. rotation-stop reproduces 0.9921 at its original
104 trials against a published 0.992, then scores 0.9889 at 167.
rotation reproduces 0.9905 at 65 against 0.990, then 0.9870. Both clear
the 0.95 bar; 42 more trials cost 0.003 each. The deployed book is still
validated. What it is not is profitable live, which is the separate
question H59 opened.

data/lake-current refreshed to 2026-08-26. The refresh script touches
only the current lake, cannot bump the research epoch, and proves rather
than asserts that the frozen lake was left alone: BTCUSDT still 3,249
rows ending 2026-07-09, max |diff| 0.0 across all 3,249 shared days. H70
is registered as time-dependent because its live diagnostic reads the
moving lake; its BARS read only the frozen lake and are reproducible.

Nothing is deployed, changed, or made paper-eligible. The paper records
continue unchanged, which is the correct outcome of a vindication.

619 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: 167 registered (166 run, 1 data-blocked: H54). Every new
+## Trial ledger: 170 registered (169 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
@@ -14,7 +14,7 @@ assumptions. PROJECT_STATE.md = what runs now; this = why and what we know.
 ## 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,
-## H69 the strategy built on H68 KILLED.
+## H69 the strategy built on H68 KILLED, H70 K=2 VINDICATED.
 
 ## Hypothesis ledger (docs/hypotheses/, docs/research/)
 
@@ -182,6 +182,36 @@ 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.
 
+14. **Breadth feeds edges that HARVEST and starves edges that SELECT —
+    now with evidence from both directions (H70).** H65 proposed this,
+    H66 withdrew it, and H70 supplies the missing half. Varying the
+    deployed rotation's slot count — the one number hypothesis 11 fixed
+    by fiat ("Long-only. **K=2 FIXED**") and every descendant inherited
+    untested — gives Sharpe **1.47 / 1.61 / 1.40 / 1.27** at K = 2/3/5/8
+    and MDD **worse than K=2 at every higher K**. All three
+    pre-registered predictions (Sharpe rising, CAGR falling, MDD
+    improving, all monotone) were **wrong**.
+    **The mechanism is mechanical:** carry harvests a premium paid by ~20
+    near-independent funding streams, so averaging more cuts variance
+    without cutting the mean. Rotation *selects*, and the 4th-8th ranked
+    coins are worse assets rather than additional independent draws of
+    the same edge. Diluting a selection edge lowers the mean faster than
+    the variance. **Still a hypothesis, not a rule** — one measurement in
+    each of two families is precisely the evidential state that produced
+    the withdrawn refinement last time.
+    **K=2 is vindicated on evidence for the first time.** K=3 beats it on
+    return (Sharpe 1.61, CAGR +46.23%) and loses on drawdown; it was not
+    the declared primary, it fails the registered MDD bar, and it is
+    **not adopted** — acting on it needs a fresh registration.
+    **And the live window says the opposite:** over 2026-07-10..08-26 the
+    same cells give K=2 −6.06%, K=5 −3.85%, K=8 **−0.85%**, with MDD
+    improving monotonically. So concentration is a **real contributor to
+    the live drawdown** — about five of the six points — and 48 days is
+    1.7% of the evidence behind the 2,880-day backtest. Both facts are
+    true; neither licenses changing K. **The sharpened open question:
+    is the live period unrepresentative, or has the K surface moved?
+    That needs forward time, not another slice of the same history.**
+
 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
@@ -243,6 +273,21 @@ assumptions. PROJECT_STATE.md = what runs now; this = why and what we know.
     in the units the position actually pays in, not the units the
     phenomenon is quoted in.**
 
+## DSR re-check at 170 trials (2026-08-28) — the standing commitment, honoured
+
+`family-expansion-program.md` §5 requires re-validating the deployed book
+as N grows. Last honoured at 125; run again at 167/170.
+
+| Book | Reproduced at its original N | DSR @167 | Bar 0.95 |
+|---|---|---|---|
+| rotation-stop (deployed) | 0.9921 vs published 0.992 | **0.9889** | CLEARS |
+| rotation | 0.9905 vs published 0.990 | **0.9870** | CLEARS |
+
+Forty-two more trials cost **0.003** each. Third confirmation that the
+DSR bar is far less sensitive to ledger growth than was once feared. The
+deployed book remains validated; what it is not is profitable live, which
+is a different question and the one H59 opened.
+
 ## DSR re-check at 125 trials (2026-08-11, scripts/dsr_recheck.py)
 
 Correction candidate 7 CLOSED. Reproduce-first guard passed on both books

PROJECT_STATE.md

@@ -436,7 +436,7 @@ that moves underneath them cannot serve as their witness.
 | Path | Contents | Role |
 |---|---|---|
 | `data/lake` | through **2026-07-09** | the FROZEN research lake. The input set every published figure was computed on. Immutable until a deliberate, recorded epoch bump. |
-| `data/lake-current` | through **2026-08-10** | the CURRENT lake. New research, and the divergence hunt. |
+| `data/lake-current` | through **2026-08-26** (refreshed 2026-08-28) | the CURRENT lake. New research, and the divergence hunt. |
 
 The frozen one keeps the plain name because every committed script points at
 `data/lake`; renaming would touch the whole corpus and would itself invalidate
@@ -506,6 +506,104 @@ presented as a discovery.
 
 ---
 
+## H70 — K=2 VINDICATED, and the deployed book re-validated (2026-08-28)
+
+`OBSERVATION` — two things owed to the **deployed spec** rather than to a
+new family. Ledger 167 -> 170.
+
+### 1. The standing re-validation, honoured
+
+`family-expansion-program.md` §5 requires re-validating the deployed book
+as N grows. Last honoured at 125.
+
+| Book | Reproduced at its original N | DSR @167 | Bar |
+|---|---|---|---|
+| **rotation-stop (deployed)** | 0.9921 vs published 0.992 | **0.9889** | CLEARS |
+| rotation | 0.9905 vs published 0.990 | **0.9870** | CLEARS |
+
+Forty-two more trials cost 0.003 each. **The deployed book is still
+validated.** What it is not is profitable live — a different question.
+
+### 2. H70: was K=2 ever the right number?
+
+`OBSERVATION` — hypothesis 11 line 60 reads *"Long-only. **K=2 FIXED**"*.
+It was fixed by fiat and inherited untested by every descendant, including
+the H42b spec on paper today. H70 varied it and nothing else. The K=2
+recomputation reproduces the published deployed figures exactly
+(+42.91% / 1.47 / −29.01% vs +42.9% / 1.47 / −29.0%).
+
+| K | CAGR | Sharpe | MDD | DSR@170 |
+|---|---|---|---|---|
+| **2 (incumbent)** | +42.91% | 1.47 | **−29.01%** | 0.9994 |
+| **3** | **+46.23%** | **1.61** | −32.40% | 0.9998 |
+| 5 (primary) | +34.53% | 1.40 | −31.46% | 0.9979 |
+| 8 | +27.13% | 1.27 | −31.82% | 0.9931 |
+
+**Gate A fails on the primary and on every cell** — MDD is worse than
+K=2's at every higher K — so the verdict does not depend on which cell was
+primary. **K=2 is vindicated, on evidence for the first time.**
+
+`OBSERVATION` — all three pre-registered predictions were **wrong**.
+Sharpe and CAGR peak at K=3 and fall; MDD never improves.
+
+`INTERPRETATION` — **H66's carry finding does not transfer, and the
+mechanism is mechanical.** Carry *harvests* a premium paid by ~20
+near-independent funding streams, so averaging more cuts variance without
+cutting the mean. Rotation *selects*, and the 4th–8th ranked coins are
+worse assets, not more draws of the same edge. This is the first evidence
+that H65's select-vs-harvest distinction is real — arriving from the
+opposite direction after H66 withdrew it — and it stays a **hypothesis,
+not a rule**.
+
+### 3. K=3, presented and not acted on
+
+K=3 beats the incumbent on **Sharpe (1.61 vs 1.47) and CAGR (+46.23% vs
++42.91%)**, losing only on drawdown (−32.40% vs −29.01%). It is **not
+adopted**: it was not the declared primary, it fails the registered MDD
+bar — the bar meta-finding 8 says decides prop-firm outcomes — and 0.14
+of Sharpe across adjacent K on one path is not a measurement of the K
+surface. **The trade is +3.3pp CAGR and +0.14 Sharpe for 3.4pp more
+drawdown.** If wanted, it needs its own registration and a prop-sim
+pass-rate comparison.
+
+### 4. The live window says the opposite — the H59 divergence, partly answered
+
+Every cell replayed over the live paper window (2026-07-10 → 2026-08-26,
+`data/lake-current`, L=90 as the account actually runs):
+
+| K | live return | MDD |
+|---|---|---|
+| 2 (incumbent) | **−6.06%** | −15.48% |
+| 3 | −7.54% | −15.50% |
+| 5 | −3.85% | −11.16% |
+| **8** | **−0.85%** | **−7.09%** |
+
+`INTERPRETATION` — **concentration accounts for roughly five of the six
+points rotation-stop gave up live.** The live drawdown is not purely bad
+selection luck; it has a structural component and that component is K.
+
+**And this does not license changing K.** 48 days is 1.7% of the evidence
+behind the 2,880-day backtest, which says K=8 earns 27%/yr against K=2's
+43% with a worse drawdown. Switching because the last seven weeks favoured
+another setting is textbook recency-chasing.
+
+`OBSERVATION` — **a pre-declared inference was withdrawn.** §6 said a Gate
+A failure would mean concentration does not explain the live drawdown.
+That does not follow: the bars are computed on the frozen backtest and
+cannot answer a question about the live window. Drafting error, corrected
+in the verdict.
+
+**The sharpened open question for the divergence hunt: is the live period
+unrepresentative, or has the K surface moved? That needs forward time,
+not another slice of the same history.**
+
+`OBSERVATION` — `data/lake-current` was refreshed to 2026-08-26
+(`scripts/refresh_current_lake.py`). It touches only the current lake and
+proved rather than asserted that the frozen lake was untouched: BTCUSDT
+still 3,249 rows ending 2026-07-09, max |diff| 0.0 on all shared days.
+
+---
+
 ## H68 — cross-venue dislocation SIGNALS; H69 shows it is not tradable (2026-08-27)
 
 `OBSERVATION` — pre-registered (commit 3eb0fbc) before any study code