Jarvis

# Weekly Flight Log — Paper Trading Allocation Bot **Period:** 13 Jul → 20 Jul 2026 (7 days) · **Runs:** 28, all decided *Paper trading only — no live money. FX (HKD→USD) is a paper approximation.* --- ## Posture & performance (per sleeve, vs its mandate) Think of the two sleeves like two crew members with different job descriptions. **Conservative sleeve** — *mandate: protect capital, only act on corroborated signals.* - Holdings: **SPY** (17 sh, ~$12,636) and **VOO** (18 sh, ~$12,297). - P&L: essentially flat — SPY **−$9.52**, VOO **−$10.08**. Combined that's about **−$20 on ~$25k**, i.e. a rounding error. Exactly what "capital protection" should look like in a choppy, broadly-down week. - Behaviour: 313 HOLDs vs 23 OPENs across the week, and **zero self-initiated trades**. This sleeve did what it's told — it never freelanced. **Aggressive sleeve** — *mandate: opportunistic, can self-initiate under a tighter position cap.* - Holdings: **QQQ** (45 sh, ~$31,290, unrealised **−$848.72**) and a tiny **XLE** position (1 sh, +$4.15). - P&L: dominated by the QQQ drawdown. Entry was ~$714, now $695 — the sleeve is sitting through a ~2.6% loss on its core position. - Behaviour: 296 HOLDs, 26 ADDs, 15 OPENs, with **6 self-initiated** actions. More active, as designed. **One honest caveat:** the data gives no realised P&L (`closed_pnl` is empty) and no benchmark. So "performance" here is *unrealised* mark-to-market only. We can judge behaviour and risk discipline confidently; we cannot yet judge whether the strategy *makes money* — nothing has been closed. --- ## What was decided, and why The market backdrop this week was described as a **broad downward trend**. In that environment both sleeves leaned heavily on **HOLD** — the correct default when there's no edge. The recurring themes across the 28 runs: - **XLE (energy)** was the star candidate. It carried a genuine breakout BUY signal *and* a **negative correlation to the book (around −0.38 to −0.46)**. Negative correlation is the valuable bit: XLE tends to zig when the equity holdings zag, so it reduces overall portfolio risk rather than piling on more of the same bet. Both sleeves acted on it — conservative with small starters, aggressive with repeated high-conviction ADDs. - **QQQ (tech)** had a Bollinger mean-reversion BUY signal. The aggressive sleeve used this to *average down* into its underwater position (buying more at lower prices on the theory it'll revert up). The conservative sleeve took only tiny QQQ starters, explicitly flagging its 24%+ volatility as too hot for the mandate. - **EFA (international)** appeared repeatedly with a full-conviction MACD BUY and low book correlation — a clean, corroborated diversifier. Both sleeves opened modest positions. - **SPY/VOO** were mostly **held, not added** — the reasoning being that they're ~1.0 correlated to each other and to the book, so adding more just concentrates the same bet. That's a genuinely sound instinct. - Everything else (GLD, EEM, TLT, IWM, GOP, NANC, DIA) was **held/passed** — either no signal, adverse momentum, or too much overlap with what's already owned. The through-line: the bot repeatedly favoured **diversification and signal-confirmation over chasing**, and repeatedly declined to add redundant equity beta. That is disciplined thinking. --- ## Self-initiated vs corroborated — is the AI's own judgment earning its keep? This is the section you're really evaluating, so let's be precise. - **Conservative: 0 self-initiated.** Perfect adherence. Every one of its 23 OPENs was backed by a strategy signal. The mandate says "corroboration required," and it delivered. - **Aggressive: 6 self-initiated** out of 41 actions. These were the AI acting on *its own reasoning* with no strategy signal behind it: - **QQQ ADD ×2** (conviction 0.70–0.75) — adding to a winner on momentum and negative correlation. - **SPY / VOO / DIA / NANC OPENs** (conviction 0.45–0.55) — all justified by *low book correlation* and mild uptrends. **My read:** the self-initiated calls share one honest logic — "this is uncorrelated to what we own, so it diversifies." That's a legitimate, conservative *rationale* even when self-initiated; the bot wasn't gambling on wild hunches. It sized them modestly (0.45–0.55 conviction) — appropriately humble for un-corroborated bets. **But** — the one self-initiated call I'd scrutinise is **adding to QQQ**. QQQ is the sleeve's biggest position *and* its biggest loser (−$848). Some of those QQQ adds were signal-backed (Bollinger), but self-init