What the agent
learned this week.
Every Monday, an autonomous Opus pass examines the fund's biggest winners, biggest losers, and the names we missed — extracting patterns to guide the next decisions.
Our system worked best when it held stocks with balanced component scores that strengthened over time — PANW (+17%) and DDOG (+12%) are textbook cases where the ensemble upgraded conviction after mediocre starts. It failed hardest on CRWD (-73%), where a high statistical win-probability masked a weak thesis fit, and on MRVL (-16%), bought at a 52-week high with poor risk quality. The one that got away was ANET (+15%): we skipped it despite the model quietly holding an 80+ score for over a week, because the initial thesis read was too cautious on optical-interconnect leadership. The system-level lesson is that we should trust sustained score improvement more than initial component reads, and we should treat 'stock at 52-week high with low risk_quality' as an automatic size-down rather than a full-position hold. Simply put: the ensemble often knows before we act — we need to let it override stale first impressions, and we need harder ceilings on positions that look extended.
Learning archive.
The fund's biggest wins (AMAT +33%, PANW +28%, MU +13%) came from holding bull positions with stable high composite scores through minor score dips — the system correctly identified persistent conviction. The biggest miss was CRWD (-71%), where an 80+ win_probability score masked the fact that we were buying a $75-PE cybersecurity name mid-range with no catalyst; the risk specialist should have vetoed. We also missed CTSH (+21% for the bear) because our specialists produced a lifeless 50/50/50 baseline on a name already broken at 1% of its 52-week range — the scoring engine goes numb when there's no narrative to grip. Correctly skipping ORCL (-34%) and ARM (-8%) validates that the composite reliably fades expensive AI-narrative names when momentum and win-probability diverge. The system-level lesson: our scoring is excellent at recognizing durable strength but weak at (a) demanding valuation discipline on extended winners and (b) generating conviction on quiet contrarian setups with no news flow.
The fund's biggest wins came from a clear, repeatable setup: AI semiconductor names breaking out to new 52-week highs with a named hyperscaler or Nvidia partnership in the news — Micron, Marvell, and Applied Materials together delivered between +17% and +39% in 30 days. The biggest mistake was the opposite: trusting the same 'high score at the 52-week high' signal on names like Broadcom and IREN, where elevated valuation and weak risk quality were warning signs the model ignored. We also missed a clean contrarian opportunity in Accenture and Cognizant — our bear theses were correct in direction but came after the stocks had already fallen 70%+ from highs, so the easy money was gone and they bounced sharply. The system-level lesson: composite scores need a 'thesis exhaustion' check that asks where in the 52-week range we are before acting. Skipping Palantir, Oracle, and Elastic was correctly disciplined — extreme PE or weak momentum overrode optimistic narratives. Finally, our best winners may have been under-sized; conviction that holds above 90 for weeks should translate to bigger weights, not just longer holds.
Our biggest wins — Marvell (+58%), Micron (+51%), and Applied Materials (+37%) — were AI-infrastructure names breaking out to 52-week highs on specific Nvidia and hyperscaler partnership news; in each case we correctly looked past mediocre initial risk scores and trusted the price/news combination. Our worst held position, Microsoft (-9%), shows the opposite failure mode: a sky-high thesis score (87) blinded us to weak momentum (44) and a stock sitting at only 37% of its 52-week range while peers ripped — strong narrative does not equal strong stock. The clearest missed opportunity was Arm (+43%), where a high PE ratio triggered a risk-quality veto that overrode an otherwise textbook breakout with company-specific catalysts. The system-level lesson: when price action, catalysts, and sector news align, we should let those override conservative risk and win-probability scores; when only the long-term thesis is strong but the tape is weak, we should trim faster. We also need to size winners more aggressively — the same conviction that justified holding MU and MRVL through volatility justified larger initial positions.
Our biggest wins this period — Marvell (+58%) and Micron (+35%) — came from buying AI-infrastructure hardware names that were breaking out to new 52-week highs on specific Nvidia/hyperscaler partnership news, even when traditional 'quality' metrics looked mediocre. Our worst held positions (Broadcom, Amazon, Google) shared the opposite profile: we bought mega-caps already pinned within 7% of their highs on generic AI-ecosystem buzz, and the system never flagged the extension risk because composite scores plateaued at 95+ for weeks. The one that got away was ARM (+82%), where our risk specialist over-penalized a high P/E ratio and overrode strong breakout and catalyst signals — the same setup we got right on MRVL. The system-level lesson: when a name has a concrete, name-specific catalyst AND is breaking out, valuation-based risk penalties should be capped; when a name is merely riding a thematic tide near its highs without a specific catalyst, conviction scores should de-rate, not plateau. On sizing, we got direction right but likely magnitude wrong — winners held weeks of 96+ conviction without being scaled up before we trimmed.
Our biggest wins came from AI-infrastructure names like Marvell, where a concrete partnership catalyst (Nvidia's $2B deal) plus a breakout to new highs delivered a 55% gain, and from cybersecurity holdings like Palo Alto Networks riding the 'secure the AI agent' narrative to +31%. Our biggest miss was Snowflake (+56%) — our momentum scorer penalized it for trading in the lower third of its 52-week range, treating a base as weakness rather than opportunity, and we similarly missed Intuit's bear case despite a 94 thesis score. The held losers (Broadcom, Amazon, Google) all shared a tell: we bought at 93%+ of 52-week highs on thinning volume, where strong-looking scores masked exhaustion. The system-level lesson is that our framework rewards stocks that already moved and punishes stocks setting up — we need an 'oversold quality' override so that high thesis-conviction names trading near 52-week lows aren't auto-skipped on weak momentum. On sizing, persistent 95+ scores on a +50% winner like MRVL should command top-tier weight, not equal weight with choppier mid-conviction holdings.
The portfolio's biggest wins came from leaning into confirmed AI-infrastructure leaders breaking out to 52-week highs — Micron returned +79% and Palo Alto +56% on stable, high-conviction scores. Where we failed was on AMD: every signal screamed buy (price at 99.7% of 52-week range, direct AI-chip catalyst headlines, strong momentum) but our win-probability specialist was spooked by valuation and kept the score oscillating, so we never built a real position despite a +43% move. We correctly avoided AI-displacement losers like Workday and Five9, where the bear thesis was confirmed by collapsing scores and on-point disruption news. The system-level lesson: when thesis, catalyst, and momentum all align with a clean breakout, valuation-based caution should be capped rather than allowed to veto entry — and our winning holdings should be sized larger when their scores stay pinned in the high 90s.
Our biggest wins (MU +51%, MRVL +19%, PANW +46%) shared a clean fingerprint: stocks breaking out to 52-week highs with direct, ticker-specific AI-infrastructure news (memory demand, optical interconnects, AI-agent security). Our losers (ETN, DLR) looked similar on the chart — also near 52-week highs — but lacked the direct catalyst and carried richer multiples, which the model belatedly recognized as scores collapsed from 80s to 40s. The most painful miss was Datadog (+72%), where the risk specialist's allergy to a 421x PE overrode an unambiguous 'observability for agentic AI' catalyst that mirrored the PANW setup we did capture. The system-level lesson: when a thematic news cluster is already producing a held winner, peer names in that cluster deserve an automatic second look rather than being filtered out by single-component vetoes like high PE or low momentum. We should also size our highest-conviction, persistently top-scoring names more aggressively — MU earned a 13-15% weight, not the standard book slot.
Our biggest wins came from direct AI silicon exposure bought at fresh 52-week highs — Micron returned 98% and Marvell 65% on memory and interconnect demand tied to named hyperscaler contracts. The system worked when momentum and a concrete catalyst lined up; it failed when we substituted data-center landlords (Digital Realty, Equinix) as cheaper AI proxies and got single-digit returns instead. The most painful miss was AMD, which rallied 95% while our win-probability specialist kept vetoing it on valuation despite a near-perfect technical and momentum profile — we even briefly bought it and flipped back to skip. The lesson is that in a thesis-confirming tape, valuation-based risk scores need a ceiling, and once a stock's daily score crosses 88 we should not whipsaw back to skip within three days. On sizing, our highest-conviction names earned their weight, but mid-conviction names where the score climbed sharply after entry (Broadcom, Marvell) suggest we should scale in as conviction confirms rather than fix size at day one.
Our biggest wins (Micron +98%, Marvell +56%) came from a tight, replicable formula: stocks breaking out to new 52-week highs while news headlines tied them directly to AI infrastructure spending. The system held these correctly even when traditional 'win probability' scores were only moderate. Our biggest blind spot was AMD, which delivered +105% — we sat out because high valuation triggered a veto, even though every other signal screamed buy. The lesson: when thesis, momentum, and breakout all align on a confirmed AI-demand story, valuation concerns are usually already priced in and shouldn't block entry. On the defensive side, the system correctly avoided three AI-displacement victims (Chegg, Five9, WPP), recognizing that being cheap is not protection when your business model is in the crosshairs. We need to loosen the valuation veto for confirmed-breakout names and tighten our skepticism on data-center holdings where the news flow is generic sector hype rather than company-specific catalysts.
Our biggest wins came from AI-semiconductor names — Micron, Marvell, and Broadcom — that were breaking out to fresh 52-week highs with concrete catalysts like Nvidia partnerships and memory-demand upgrades; we held them through 35-75% gains. The system also correctly avoided three AI-disruption victims (Chegg, Five9, WPP) where the news flow plainly predicted the businesses being eaten. Our clearest miss was AMD, which traded in lockstep with the chips we owned but was blocked by an over-cautious win-probability score reacting to valuation; a simple rule that floors win-probability when a peer in the same news cluster is already held would have caught it. The system-level lesson: high-quality, high-conviction names with weak near-term catalysts (META, Equinix, Digital Realty) are dead money in a momentum tape, and we should size them down rather than hold at full weight when catalyst_pct stays below 60.
How it works: each Monday the system identifies the past month's biggest winners (whether held or not) and biggest losers across the watchlist, then asks Opus to extract concrete, replicable patterns. The patterns feed back into the Portfolio Manager agent's prompt for the following week's decisions.