# Sunday, August 2, 2026

*Fired on time at 23:05 ET — second on-schedule nightly in a row, so the sleep-burst did not recur over the weekend.*

## TODAY — Top 3

1. **Call the revbreak_long demotion — 6 days stale, and Friday added more evidence.** Setup League already had it confidently negative live (n=259, 18% win, -0.46 avgR, 90% lower bound -0.54). Friday 7/31 added 6 more signals for -0.7R, and the Week Ahead page built this morning still lists it as a live play for Monday. Every day it stays in the validated set it fires again and costs money. This is the only open item that is directly P&L.
2. **Pick one of the 3 SIP scoring options so blue-line can unblock.** Catalyst scores 21.6–36.0 pts vs mover 1.5 / rrg 2.2 / liquidity ≤10, so PVP / PBS / GAPMA / LYNCH technical hits mathematically cannot reach the top-10. This morning's Week Ahead is the proof: the 20 prepped names are ordered catalyst-first, so the technical scanners contributed almost nothing to Monday's study list.
3. **Decide BOAT v4.6 — lock the cap=1 selectivity build, or shelve it.** Today's sweep gives a clean answer to test rather than more sweeping. The signal beats random control everywhere, so it is real, but it does not clear a 0.10% round trip: **+0.006 at zero cost, -0.042 at 0.10%**. The cap=1 build holds up out of sample; the uncapped one decays. Commit and attack the cost side, or shelve it before it eats another week of bench time.

## What shipped

- **[LYNCH Master Combinatorial Backtest](/system/LYNCH_MASTER.html)** (22:00) — 3,618 Grade × Sizing × Exit × Regime-gate combinations, 5 years, real dollars, whole shares, $50k position cap, SMB commissions. **998 survived the both-halves honest test.** Winner on *both* return/DD and raw dollars is the same recipe: A+B+C · UVXY-mom × Reactive · Swing pure (hold 5d) · no gate → **+$71,454 with $14,992 max drawdown** (477% return/DD, 45.1% win, Sharpe 0.74). The top-50 concordance is the more useful read: grade appears in 100% of the top 50 and the sizing rule in 84%, but the best exit tops out at 26% and the gate splits 36/36/28 — **grade and sizing are the edge; exit and gate are close to noise.**
- **BOAT v4.6 filter + cost sweep** benched on a2 (`~/j3p-v269/`) — 139 dates, top 8 gappers/day, sweeping the efficiency gate, opening-range %, per-symbol-day cap and a timezone fix. It beats its random control decisively at every setting (-0.042 vs -0.191 capped), so the signal is real — but the whole verdict is the cost band: +0.006 / -0.018 / -0.042 / -0.066 at 0.00 / 0.05 / 0.10 / 0.15%. Selectivity separates the two builds: **cap=1 gives 1.0 trades/day and gets *better* out of sample** (2nd half -0.032 vs 1st -0.052), while the uncapped in-play-only gate gives 7.6/day and **decays** (-0.094 vs -0.041). `kind=BOAT` loses to `kind=V` in every cut.

## What moved forward

- **J3P Week Ahead auto-rebuilt** at Sunday 10:00 with 20 names prepped for Monday 8/03, leaned long/short off the 5-DMA, each with the scanner that found it. It also carries a blunt fact for decision #1: **every one of Friday's five plays lost** — vwap_reject_short 13 signals / -8.3R, vwap_pullin 5 / -5.0R, revbreak_long 6 / -0.7R, earn_gap_reclaim -1.0R, trend_ride -1.0R.
- **EOD catch-up ran clean at 16:45** — 8/02 scorecard (empty, no weekend trading), `you-vs-you`, manifest, and the 7/29–7/31 six-trader backfill again at ok=18 fail=0. You-vs-you is unchanged from Saturday: -$887 net over 20 days, 43.9% win, 3,663 trades, and the sub-2-minute leak still the whole story (1,192 trades for -$2,091; remove them and the period flips to +$1,204).
- **multi-tf daemon ran its Sunday 18:00–20:00 window** — 118 full + 276 tick refreshes, 8/02 page, index, alerts and preview-verdict all written. First two rebuilds failed (rc=1) then it self-corrected on #3.

## What's blocked

- **Both P0 decisions are now six days old** — revbreak_long demotion and the SIP scoring rebalance, untouched since 7/28. Nothing else routes around them.
- **New chronic bug: the multi-tf "verdict preview timed out (>60s)" has fired every single day since at least 7/20** and has never been flagged. Per-day counts 7/20→8/02: 11, 11, 22, 10, 10, 2, 6, 6, 8, 6, 11, 3. The preview does land at daemon exit so the page is not stale, but a step that blows its budget several times a day, every day, for two weeks is not a blip.
- **The nightly scan is still unfixed, and tonight found a second failure mode.** WebDAV Call A finally returned after ~50 minutes and put a hard number on the `os.walk` bug: **549 rows for 4 unique files — exactly 137 duplicates of each.** The visited-set is keyed on the raw path instead of realpath, so the same tree gets re-walked 137 times. (Its file list matched the ssh scan exactly, so nothing in tonight's report was missed — the WebDAV path is correct, just ~750× slower than it needs to be.) But the ssh replacement is not a drop-in either: a recursive `find` over `~/public_html` on a2 returned **zero rows and exit 0** at maxdepth 1, 2 and 3, while the identical predicate inside `~/public_html/scorecards` correctly returned the three files written at 16:45. Same-shape finds over `~/setup_daily_reporting/reports` and `~/gt_dev` worked fine. **Root cause found on the last scan of the night:** a wide 11-root find ran past 3 minutes and died with `Connection reset by 185.181.254.167 port 22`, exit 255 — the *host* is killing the session, almost certainly a CloudLinux CPU limit. The earlier runs only looked like clean empty results because the reset landed on the far side of a `| sort`, which then exited 0 on empty input. So: don't pipe the remote find through anything, keep each ssh call to one or two roots, and treat an empty result as a failure — never as a quiet day.
- Minor: `accountability-v2/api/error_log` and `gt_dev/php_error_analysis.log` each grow every 60 seconds with *successful* DB-connection messages that print the database name and user into a file inside the web tree. Buries real errors; logged as a P3.

## Not a regression

No MGP and no `system/briefings` today. Neither ran on 7/25 or 7/26 either — **weekends are by design.** The 8/01 Saturday MGP was the burst-fire anomaly, not the rule.

## Files touched

- `public_html/system/LYNCH_MASTER.html` — combinatorial backtest rebuilt (22:00)
- `public_html/system/J3P_WEEK_AHEAD.html` — Sunday 10:00 auto-rebuild, 20 names for Mon 8/03
- `public_html/system/multi-tf-stack/2026-08-02.html`, `index.html`, `alerts.html`, `preview-verdict.html`, `data/`, `history/`
- `public_html/scorecards/trade_scorecard_2026-08-02.html`, `manifest.json`, `you-vs-you.html` — 16:45 catch-up
- `a2:~/j3p-v269/` — `boat_bench.py`, `boat_bench2.py`, `bench_frozen.py`, `bench_tz.py`, `target_sweep.py`; `runs/filters.log`, `runs/liquid.log`, `runs/tzfix.log`, `runs/boat_v46.json`
- `a2:~/gt_dev/lynch/data/universe.json`, `gt_dev/breadth/extended_live.json` — 20:00 data refresh feeding the LYNCH run
- `setup_daily_reporting/reports/` — `eod_catchup.log`, `multi_tf_daemon.log`, liquidity/companies caches

---
*Generated by nightly-project-manager — 2 shipped, 3 moved forward, 4 blocked*
