Jarvis

# Weekly Flight Log — Paper Trading Allocation Bot **Period:** 20 Jul – 27 Jul 2026 (7 days) · 28 decision runs, all completed > Quick orientation, Captain: think of this bot as having two crew members. The **conservative sleeve** protects capital and only acts when the evidence lines up. The **aggressive sleeve** is allowed to chase opportunities harder, with a tighter risk cap to keep it honest. Every claim below is grounded in this week's data. A note on the money: the account was funded with **1,000,000 HKD** and all P&L is shown in USD as a *paper approximation* — treat the exact FX as rough. --- ## Posture & performance (per sleeve, vs its mandate) **Conservative sleeve** — currently holding four positions: - **XLE (energy)**: +$189.55 unrealised — the standout winner, riding a genuine breakout. - **DIA (Dow)**: +$53.76 unrealised — quiet, low-volatility, doing its job. - **SPY**: –$38.90 · **VOO**: –$40.15 — both slightly underwater, both broad US equity (essentially the same bet twice; more on that below). - **Closed trades this week**: 14 trades, **3 winners**, net **–$651.68**. The realised loss is honest and worth teaching on: the sleeve was *right to be cautious* — most of its closed losses came from broad-equity churning while the market trended down. Its mandate is capital protection, and its live book is net positive on unrealised P&L, so the ship is trimmed correctly even if some past trades cost a little. **Aggressive sleeve** — currently holding one position: - **QQQ (tech)**: –$150.29 unrealised — bought a dip that hasn't yet reversed. - **Closed trades this week**: 28 trades, **10 winners**, net **–$532.23**. The aggressive sleeve traded roughly twice as much and lost slightly less in dollars — but with a similar win rate (~36% conservative closed wins vs ~36% aggressive). Neither sleeve had a *good* realised week. Both are down. That's the plain truth. --- ## What was decided, and why The dominant action across both sleeves was **HOLD** — conservative held 293 times, aggressive 246. That is the system saying *"no clear edge, so don't touch."* For a captain, that's the equivalent of not making a course change without instruments agreeing. Good discipline. The recurring themes in what it *did* act on: - **XLE was the conviction trade everywhere.** Both sleeves repeatedly cited a breakout + momentum BUY signal, a +8–10% 20-day return, and — importantly — a *negative correlation to the rest of the book* (around –0.35 to –0.43). Negative correlation means XLE tends to zig when the rest zags, so it doubles as a hedge. This is the one idea the bot has genuinely high confidence in. - **GLD (gold)** was added by both sleeves as a defensive diversifier on an SMA-crossover BUY signal. - **EFA (international equity)** and **SPY** were opened at modest size as diversifiers. - The aggressive sleeve **closed DIA** (no signal, drifting down, no diversification value) and **trimmed QQQ, GLD and GOP** to raise cash when it ran short — a sensible cash-management reflex. What it pointedly **did NOT do**: it repeatedly refused **QQQ, EEM, IWM** and others despite tempting recent numbers, because there was no strategy signal or the volatility was too high for the mandate. The clearest example: EEM had a –10.2% 20-day drop and ~33% volatility — the bot called it a "falling knife" and stayed out of it even in the aggressive sleeve. That is exactly the judgment we want. --- ## Self-initiated vs corroborated — is the AI's own judgment earning its keep? **This is the number you're evaluating, and it's unambiguous this week:** - Conservative self-initiated: **0** - Aggressive self-initiated: **0** **Every single action in both sleeves was backed by a strategy signal.** The AI did not once act on its own hunch with no signal behind it. So the honest answer to "is the AI's own judgment earning its keep?" is: **it wasn't tested this week** — it never went off-script. What the AI *did* contribute was **sizing and vetoing judgment**: e.g. seeing a valid QQQ BUY signal but deliberately sizing it small (0.45–0.5 conviction) because volatility was high, or holding XLE rather than pyramiding into it to avoid over-concentration. That's judgment layered *on top of* signals, not judgment replacing them. Bottom line: no evidence this week of the AI freelancing recklessly. But also no evidence yet of it generating independent alpha, because it took no independent shots. --- ## What I'm watching / uncertain about 1. **SPY and VOO are the same trade held twice.** The bot itself flagged their ~0.92 correlation repeatedly and correctly declined to add more — but the conservative sleeve still *holds both*, both underwater. That's mild redundant concentration we should keep an eye on. 2. **Missing correlation data** appears in several rationales (XLE and EFA at times). The bot correctly reduced conviction when data