docs: batch-three findings (T11 TTL optimal, T12 no exit rule, T14 imbalance band, T15 churn) + backtest paper-parity + Profitable bench

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

FINDINGS.md

@@ -704,6 +704,50 @@ can't flatter the verdict. Either outcome is a finding: the premium
 median kills the mirage in three days, or the paper ledger prices the
 first edge that survives its own execution.
 
+**Batch three (2026-07-23 night, while the harness deployed):** four
+more questions answered on already-flowing data, plus the backtest
+itself brought under the discipline.
+
+- **T11, the patience curve:** #20's 60-second maker TTL is already
+  optimal. Touches arrive in 1–2s (p90 ≤7s), fill rate flatlines by 5
+  minutes, and the marginal late fills LOWER EV/signal (+$16.62 @60s vs
+  +$15.36 @5m+) — at 5m+ every eventual loser reaches the bid and the
+  3% that never fill are winners running away. Patience past a minute
+  buys nothing but adverse selection.
+- **T14, depth structure:** one real cross-harness regularity in 1,923
+  chain-graded fills — book imbalance at fill time is a BAND, not a
+  dial. Extremes are adverse both ways (empty-bid books = falling
+  knives, bid-heavy books = the winner's curse); the balanced 0.25–0.47
+  middle is least-bad in both instruments. A portable fill-quality
+  guard candidate for #23's thin siblings. The slope feature's
+  "consistent" gate was a quantile-tie artifact (books cluster at 4-5¢
+  across five levels; the quartile boundary landed inside tied values)
+  — discarded, lesson logged: check feature variance before trusting a
+  quantile gate.
+- **T12, unwinds:** maker-sharp inventory take-backs are NOT exit
+  signals. STAY on the lean after the sharp unwinds half: +$5.95/unwind
+  (n=265, 2.6× the pre-declared bar); FADE it: −$5.22. #21's finding
+  one species over — the lean's information outlives the sharp's own
+  profit-taking. Study C keeps no exit rule. (Concentration caveat in
+  the file header: top-5 assets carry the total, the tail bleeds —
+  decisive for the exit-rule question, not a new edge.)
+- **T15, sharp half-life (first half):** the published sharp screen
+  churns HARD — 49% of newly detected wallets are gone within a day,
+  34% survive three days, ~10% a month. Whether that is edge death or
+  threshold jitter awaits the EV-by-age grade (the chain-ensure over
+  these wallets' pre-pipeline history is the long pole).
+- **The backtest now replays the paper book's execution** (portfolio.py
+  reads the follow flags — maker feeless entries as the DECLARED
+  optimistic bound, hold-through exits; flip the config back and the
+  backtest follows). Its first maker/hold run surfaced #21's cost side
+  honestly: holding to resolution locks capital — the $1k replay places
+  ~4× fewer bets (86 resolved vs 390 under mirror-exits) while missed
+  bets swell. The /trading "Profitable bench" section now scores the
+  15-wallet bench on realized + would-be-missed P&L: 12 of 15 carry
+  their weight, several only BECAUSE of their misses (EdwardIN −$81
+  realized, +$791 would-be) — and JuiceFarm is trimmed (+$33 realized,
+  −$72 would-be: its misses were good misses).
+
 ## Repo layout
 
 - `insider.py` — the detector: z-score/p-value, timing/freshness/sizing signals,

HANDOFF.md

@@ -138,7 +138,25 @@ push through.
   T6 lean-follow → #22 (fade arm FAILED its concentration gate,
   report-only); T9 lead-lag → #23; T1 maker-quote Stage-1 KILL; T7
   settlement-discount industrialized, parked; T8 crater-rejects were
-  good misses; T10 age gradient needs sample.
+  good misses; T10 age gradient needs sample. Batch three (same night)
+  — T11: #20's 60s TTL already optimal, patience past 60s = adverse
+  selection; T12: maker-sharp unwinds NOT exit signals (STAY
+  +$5.95/unwind n=265; Study C keeps no exit rule; #22 comment); T14:
+  balanced-book imbalance band (0.25-0.47) = portable fill-guard
+  candidate for #23; T15: sharp-screen churn 49%/1d, ~10%/28d
+  (EV-by-age grade pending its chain-ensure).
+- **Backtest = paper-execution parity (2026-07-23, 004911b)**:
+  live/portfolio.py replays the PAPER book's follow flags (maker
+  feeless entries — declared optimistic bound; hold-through exits;
+  sold-unresolved carry as open at the exit print). The backtest is NO
+  LONGER a taker control — it tracks whatever copybot.paper.json runs.
+  First maker/hold run: equity +36%, 0 fees, 86 resolved (vs 390 under
+  mirror) — hold-mode LOCKS capital, misses swell to 1,588; that is
+  #21's cost side, visible. /trading gained a "Profitable bench"
+  section (realized + would-be-missed > 0): 12/15 clear it; JuiceFarm
+  trimmed (+$33 realized, −$72 would-be). JuiceFarm's conviction floor
+  pins at the next 08:00 sync_floors run (whole-book copies until
+  then — flagged 2026-07-23 evening).
 - Dashboards: jaxperro.com/{trading,live,test,value} — /test = all four
   studies on one page (old /surge + /oracle URLs redirect) · daily pipeline
   on the Mac at 08:00 (launchd, lockfile) — floors, bench forward table,

research/README.md

@@ -84,6 +84,26 @@ Layout:
                         m=4c/R=1s under queue-front optimism; every
                         R=4s cell −$7-9/fill — staleness IS the
                         adverse selection; ~31-50k graded fills/cell)
+  copy_maker_patience.py  T11 — #20's 60s maker TTL is ALREADY OPTIMAL:
+                        touches land in 1-2s, fill rate flat after 5m,
+                        and the marginal late fills lower EV (loser
+                        fill-rate 100% at 5m+ — patience past a minute
+                        only harvests adverse selection)
+  maker_unwind.py       T12 — maker-sharp unwinds are NOT exit signals:
+                        STAY +$5.95/unwind (n=265, 2.6x the bar) vs
+                        FADE -$5.22 — #21 one species over; Study C
+                        keeps NO exit rule (concentration caveat in
+                        the header; --grade-only re-grades the dump)
+  depth_imbalance.py    T14 — book imbalance at fill is a BAND: extremes
+                        adverse both ways, balanced 0.25-0.47 least-bad
+                        in BOTH harnesses (portable #23 guard
+                        candidate); slope "gate" = quantile-tie
+                        artifact — check feature variance first
+  sharp_halflife.py     T15 — published-set churn is violent: 49% of
+                        detected sharps gone in 1 day, ~10% at 28d
+                        (real as-of git history of watch_sharps.json);
+                        EV-by-age = churn's meaning — pending its
+                        81k-cond chain-ensure
   requote.py       crater→requote timing (feeds the bots' per-niche retry)
   forward.py       scores frozen studies on new tape days → forward_ledger
                    (payouts_for chain overlay mandatory; controls + sub5c