BX shadow — daily report

Generated 2026-10-16 06:50 HKT. Read-only snapshot of Railway service bx-shadow. Source: railway status + railway logs --lines 5000.

1. Service

• status: ● Online
• region: US West
• volume: bx-xs-volume · /app/data · 0.9 GB / 48.8 GB
• deployment ID: 34ae58a9-9535-4550-bc00-f9dab1f29640
• worker version line: not in window
• log lines fetched: 5003
• heartbeat age: 3949 min (STALE — investigate)
• S3 backup: no failures in window
• worker errors: 0
• report clock: worker-heartbeat (local clock is 0.2h off — see warnings)

2. Window covered

• deploy/start: 2026-10-10 05:47 HKT (2026-10-09T21:47:55.946183+00:00)
• last heartbeat: 2026-10-13 13:00 HKT
• length: 3.30 days (79.2 h)
• hb samples: 4706

3. Events (20) — worker's durable record

Event rate: 6.06/day (backtest expectation ≈2.6/day). LO = relaxed-only (3%/2.5σ, DRY-5 only); STD = strict (4%/3.0σ). Source: the worker's own durable record (survives redeploys and log roll-off). Events are appended as detected, so one younger than the last ~10-minute bucket upload may not appear until the next run.

Detector activity (proof it is alive): largest move logged STRK +14.5% (z=3.7); max volume z=11.1; 14 of 80 hours had a coin with z>2.5; 63 move(s) ≥3% of which 54 lacked volume confirmation. The signal needs price and volume z in the same coin-hour, so both appearing separately yields no event.

4. Trades (74) — worker's durable record

mkt% = how far the coin actually moved over the hold (signed by market direction). pnl% = the same move in the trade's favour (what the book credits). w = weight. Net equity impact ≈ w × pnl% − fees.

Open positions at snapshot: 0

5. Current book state

Live equity of each book. trades this run counts only trades opened since the last worker restart — lifetime counts and all risk/return metrics (tot %, %/day, CAGR, MaxDD) are in §5b and are measured from 1.0 over era-1.

5b. Cumulative ledger — era-1 (78 trades)

Durable de-duplicated record (research/bx/daily/trades_ledger.csv) — survives redeploys and log roll-off. equity est compounds each trade's w × pnl% − fees; %/day and CAGR annualise that over the ledger's own span.

Era-1 (from 2026-10-04, worker v4, committed code). All books were reset to 1.0000 at the v4 deploy, and every trade in this ledger is era-1 (the ledger was reset with the books, so no pre-v4 trade is mixed in). Era-0 is archived under research/bx/daily/archive/era0/ — see research/bx/daily/ERA.md.

Ledger spans 2026-10-04T15:00 → 2026-10-13T01:30 UTC (8.44 days). 78 trades is far too small to trust these annualised figures — a few trades swing CAGR by orders of magnitude; %/day is the more stable read. MaxDD here is trade-level (measured at closes, so intra-trade excursions are missed) and will understate risk until many more trades accumulate.

6. STRICT vs RELAXED-only events (BX-066 pre-registration)

Cumulative era-1 ledger. DRY-4 and DRY-5 are identical except DRY-5 also admits RELAXED-only events (|r|>3% & z>2.5 that do not reach 4%/3.0), so the LO bucket is DRY-5-only and IS the incremental effect of the relaxation.

Pre-registered rule (BX-066): keep DRY-5's relaxation only if RELAXED-only trades show mean pnl% > 0 AND Σ pnl% > 0 over ≥30 such trades. Below that it is 'keep watching'. At ~1.4 relaxed-only events/day, 30 takes ≈3 weeks.

7. DRY-4 vs DRY-5 head-to-head

The Δ row is the clean causal estimate: DRY-4 and DRY-5 differ *only* in whether relaxed-only events are admitted, so (DRY-5 equity − DRY-4 equity) is the cumulative add/subtract of the relaxation. Everything else is identical by construction.

Reminder (BX-066): this is noise-dominated until ≥30 relaxed-only trades (≈3 weeks; currently 20). Do not promote or kill on a 1–2 day gap. The ledger is now pure era-1, so this Δ is a clean (if tiny) relaxation estimate — it simply lacks the sample to mean anything yet.

8. Reconciliation (accounting sanity)

Equity impact per trade ≈ w × pnl% − (w×FEE_T + w×FEE_M), summed over the trades that closed inside this run's heartbeat window, compared against the heartbeat equity move. A large gap means an accounting break. (Filtering on CLOSE time, not on trades this run, is what makes this exact — a trade that opened before the run but closed inside it still moves the run's equity.)

9. Warnings / anomalies

• ⚠️ this Mac's clock is 0.2h ahead of real time (report time taken from worker-heartbeat). launchd schedules on the local clock, so the 06:00 job fires at the wrong real time. Fix: enable 'set time automatically' (System Settings → General → Date & Time) or sudo sntp -sS time.apple.com.
• ⚠️ heartbeat 3949 min old — worker may be stalled.