aureliex.
a portfolio kept in public

The method

Five agents, one judge, a written pre-mortem per round.

Why five agents

Fifty copies of one agent converge to the same wrong answer. Five structurally different roles — bull, bear, macro, microstructure/flow, base-rate historian — see different inputs, write independently, then face an arbitrator. The research on multi-agent debate (Du et al. 2023; TradingAgents from Columbia/NYU) is unambiguous: gains come from role variance, not from count.

The panel

bull
the Bull

Steelman the 10yr quantum thesis. Maximize upside. Sized IONQ biggest of the panel.

Owns: C_pureplay, part of A_bigtech_quantum (NVDA)

bear
the Bear

Survive the drawdowns. Favor trillion-dollar balance sheets. Hold dry powder.

Owns: A_bigtech_quantum (MSFT, GOOG), E_drypowder (SGOV)

macro
Macro

Position for the backdrop, not the theme. Power is the binding constraint on AI+quantum compute.

Owns: D_power_macro (CEG)

flow
Flow

Own flow-clean vehicles. Avoid names with active ATM shelves and retail crowding.

Owns: partial in B_theme_etf

historian
the Historian

Apply base rates from prior thematic waves — most pure-plays go to zero over 10yr.

Owns: B_theme_etf (QTUM), A_bigtech_quantum (IBM)

Scoring

Each agent writes probabilistic direction calls at T+30, T+90, T+365 for every holding. Thesis text is SHA-hashed at prediction time to prevent post-hoc rewriting. When the horizon passes, a Brier score is computed from realized direction. Weights are frozen at equal (0.2) until each agent has ≥20 resolved predictions, at which point exponential decay on cumulative Brier kicks in. Scoring rewards calibration, not P&L — variance swamps skill over fewer than ~50 rounds.

Kill-switches

  • Any position over 15% of the book (12% for pure-plays) — trim on breach.
  • QTUM −30% from entry — deploy SGOV into QTUM, one-time.
  • Any pure-play files an S-3 / ATM / 424B5 — cut that position 50% same day.
  • Any name 10x's from entry — sell 50%, let the rest ride.
  • More than 2 cron-triggered trades in any 12-month window — shut down the weekly cron. Barber-Odean says monitoring is net-negative for retail; when the evidence shows up in this book, we stop.

What this site is not

Not a signal service. Not a newsletter. Not a community. Not investment advice. The portfolio is too small to copy and the method is too slow to trade. This is one person thinking in public with real money on a 10-year horizon, so that a future version of himself has a legible record of the reasoning — not just the P&L.