Separate existence from deployment: add STANDALONE-VIABLE

The standing rule "new features must beat the DEPLOYED system, not zero"
answers two questions with one answer: does this edge exist, and should it
go in the current book. It is correct for the second and was being applied
to the first, so a hypothesis that makes money standalone but less than the
incumbent adds was recorded KILLED -- the same terminal status as one with
no edge at all.

The ledger therefore cannot distinguish H04 mean-reversion ("REJECTED --
decisively", no edge) from H23, killed as redundant, whose input signal
measured a CI of [+1.25%, +5.99%]. If the deployed book is ever changed,
every edge killed for being redundant TO it is indistinguishable from the
edges that were never real. The rule protecting the book from redundancy
was also erasing the bench.

Adds a third terminal outcome. STANDALONE-VIABLE requires everything a
deployment claim requires except the comparison to the incumbent: positive
after the full cost model, CI excluding zero, DSR_global >= 0.95,
engine-grade, pre-registered. It is not softer -- it is the deployment bar
minus exactly one comparison -- and it is not a promotion path.

Deployment is unchanged: nothing enters paper or live on standalone merit.
Meta-finding 4 stands and both its examples remain correctly not-deployed.
No threshold moves, no trial count changes, nothing is re-run.

Explicitly no retroactive relabelling. Existing KILLED verdicts stand until
re-registered and re-run against the new bar; the graveyard audit's
candidates are recommendations, not reclassifications. The honest claim
today is that the partition exists, not that it is large.

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

Author: martex-dev

CLAUDE.md

@@ -17,7 +17,16 @@ killed ideas without a new pre-registered spec and a stated reason.
 - Kill test (cheap information study) BEFORE any strategy build.
   Event-driven engine is the source of truth for strategies; vectorized
   screening only pre-engine.
-- New features must beat the DEPLOYED system incrementally, not zero.
+- New features must beat the DEPLOYED system incrementally, not zero —
+  for DEPLOYMENT decisions. Amended 2026-08-27
+  (docs/research/standalone-viable-amendment.md): a hypothesis that
+  clears the full standalone bar (positive after costs, CI excluding
+  zero, DSR_global >= 0.95, engine-grade) but does NOT beat the
+  incumbent is closed STANDALONE-VIABLE, not KILLED. It is not
+  deployed; it is a live edge on the bench. Existence and deployment
+  are different questions and the incremental bar only answers the
+  second. No retroactive relabelling: old KILLED verdicts stand until
+  re-registered and re-run.
 - Paper accounts run only validated/eligible specs; one spec per record
   (spec change = archive the record, fresh $5,000 start).
 - Live/real-money actions are gated: the runbook

docs/research/standalone-viable-amendment.md

@@ -0,0 +1,129 @@
+# Methodological Amendment — separating existence from deployment
+
+Date: 2026-08-27. Status: **DECIDED — owner, 2026-08-27. In force.**
+Amends: `CLAUDE.md` standing rule *"New features must beat the DEPLOYED
+system incrementally, not zero."*
+Prompted by: `docs/research/graveyard-audit.md` §2.
+
+This is a change to what a verdict **means**, not to any threshold. No
+existing verdict is reversed by this document, and none may be relabelled
+under it without being re-checked against §3.
+
+---
+
+## 1. The defect
+
+The standing rule answers two different questions with one answer:
+
+1. **Existence** — does this edge exist at all?
+2. **Deployment** — should this go into the current book?
+
+The rule is correct for (2) and is being applied to (1). A hypothesis that
+makes money standalone, but less than the incumbent adds, is recorded as
+`KILLED` — the same terminal status as a hypothesis with no edge at all.
+
+`OBSERVATION` — the ledger therefore cannot distinguish between
+`docs/hypotheses/04-mean-reversion.md` ("REJECTED — decisively", no edge)
+and `docs/hypotheses/23-incremental-features.md` (killed as redundant,
+whose input signal H13 measured a CI of **[+1.25%, +5.99%]**).
+
+`INTERPRETATION` — this is a loss of information, and it compounds. If the
+deployed book is ever changed or retired, every edge that was killed *for
+being redundant to it* is sitting in the graveyard indistinguishable from
+the edges that were never real. The rule that protects the book from
+redundancy is also erasing the bench.
+
+---
+
+## 2. What is NOT changing
+
+Stated first, because the rule being amended has already earned its keep
+and the amendment must not be read as weakening it.
+
+- **Deployment still requires beating the incumbent.** Nothing enters a
+  paper account, the combined book, or any live path on standalone merit.
+  The incremental bar is untouched for that decision.
+- **Meta-finding 4 stands.** "Info-signal ≠ strategy improvement" was
+  learned the hard way (7d ranking real at info level, degraded the
+  walk-forward; shock signal real, fully absorbed by deployed momentum).
+  Both remain correctly not-deployed.
+- **No threshold moves.** `DSR_global ≥ 0.95` is unchanged, per
+  `mi-trial-accounting-design.md` §4.2.
+- **No trial count changes.** These trials are already counted. This
+  relabels an outcome; it re-runs nothing and spends nothing.
+
+---
+
+## 3. The amendment
+
+A third terminal outcome is added alongside `PASS` and `KILLED`:
+
+> **`STANDALONE-VIABLE`** — this hypothesis cleared a full standalone bar
+> on its own merits, and did **not** beat the deployed system. It is not
+> deployed. It is recorded as a live edge on the bench, re-examinable
+> whenever the deployed book changes.
+
+**The standalone bar is a real bar.** To be recorded `STANDALONE-VIABLE`, a
+hypothesis must meet **every** requirement a deployment claim meets, except
+the comparison to the incumbent:
+
+1. Positive expectancy **after the full cost model** — fees, half-spread,
+   participation impact. No gross-of-cost result qualifies.
+2. A 95% confidence interval **excluding zero**, by the same block-bootstrap
+   estimator its family already uses.
+3. **`DSR_global ≥ 0.95`** against the global trial count, exactly as a
+   strategy-grade claim.
+4. Engine-grade: produced by the event-driven engine, not a vectorized
+   screen.
+5. A pre-registered hypothesis document, as always.
+
+Anything failing any of these is `KILLED`, as before. There is no partial
+credit and no "point estimate was positive" route in.
+
+---
+
+## 4. The risk, and the guard
+
+`INTERPRETATION` — the honest danger is that a softer category rots into a
+dumping ground for near-misses, and the ledger quietly stops recording
+failure. That would destroy the only asset this project has.
+
+Three guards:
+
+- **The bar in §3 is not softer.** It is the deployment bar minus exactly
+  one comparison. A `STANDALONE-VIABLE` result is *more* evidenced than
+  most published retail strategies, not less.
+- **No retroactive relabelling from the armchair.** Existing `KILLED`
+  verdicts stay `KILLED` until re-registered and re-run against §3. The
+  graveyard audit's candidates (FU-B1, H02) are *recommendations to
+  re-register*, not reclassifications.
+- **`STANDALONE-VIABLE` is not a promotion path.** It cannot become
+  eligible for paper or live deployment by accumulating time or being
+  looked at again. It re-enters only through a fresh pre-registration
+  against whatever the incumbent is on that day.
+
+---
+
+## 5. What this changes in practice
+
+`OBSERVATION` — the ledger's headline currently reads as ~4 survivors from
+125 trials. Under this amendment the same history would read as three
+groups rather than two: no-edge, real-but-redundant, and deployed.
+
+`INTERPRETATION` — that is a materially different research finding, and the
+second group has never been counted. How large it is, is unknown until the
+candidates are actually re-run; the graveyard audit identified four
+suspects and proved none of them. The honest claim today is that the
+partition exists, not that it is large.
+
+---
+
+## 6. Consequence for the near-miss rule
+
+The "near-miss rule" (close a hypothesis that misses its bars, record the
+figures) is unchanged in mechanics. Its scope narrows: a hypothesis that
+misses **only** the incremental comparison, while clearing §3, is closed as
+`STANDALONE-VIABLE` rather than `KILLED`.
+
+A hypothesis that misses any §3 requirement is still closed as `KILLED`,
+including one that misses by a hair. Near-miss remains a kill.