Claude Investment Portfolio

Problem. Investing well as an individual means doing the work an institution distributes across a desk of people: generating ideas, researching them honestly, arguing the bear case as hard as the bull case, and not falling in love with your own thesis. Working solo, the common failure mode is to research what you already want to buy, skip the disconfirming work, and find the risks late. What’s missing is an adversary.

System. This is that desk, built as a staged multi-agent pipeline — a sequence of Claude sessions, each a prompt-doc that runs one stage and hands its artifact to the next. A theme enters intake; parallel research agents work it up; a red-team agent is explicitly tasked with attacking the thesis; then a committee — bull, bear, and a chair who synthesizes — debates it and produces a structured verdict. A structurer translates verdicts into portfolio shape, and a monitor sweep watches positions and re-raises anything that drifts from the thesis that justified it. The holdings live in a markdown ledger (Holdings_Master.md). Every stage produces an artifact, so the reasoning is auditable, not just the outcome.

Payoff. Decisions come with their own paper trail: the bull case, the bear case, the red-team’s objections, and the chair’s call, all written down. The reasoning is inspectable, not just the outcome. The debate transcripts are also the most readable thing the system produces.

These are staged Claude sessions, not an always-on autonomous service. Each stage (Scout, Red Team, Committee, Monitoring) is a prompt-doc kicked off on a cadence, and the output of one stage feeds the next. The AI does the reasoning; a human starts the run and takes the position. Every change to actual structure passes through a human-approve gate.

A staged, multi-agent pipeline with named roles:

• researcher
• red-team
• committee-chair
• bull
• bear
• monitor

• Researcher(s). Run in parallel on an intake theme — the breadth pass. Each produces a written work-up rather than a score, so downstream agents argue over substance.
• Red-team. The keystone role. Explicitly adversarial: its job is to break the thesis — find the disconfirming evidence, the crowded-trade risk, the reason this is already priced in. Separating “research it” from “attack it” into two roles is what stops the system from confirming its own bias.
• Committee — bull, bear, chair. The bull and bear take the two sides; the committee-chair synthesizes the debate into a verdict with explicit reasoning. The transcript of this debate is the main public artifact.
• Monitor. Runs as a monitoring stage against held positions; when something drifts from the thesis that justified it, the monitor re-raises it to the committee rather than letting it sit.

Where the human gate sits. Research, red-teaming, debate, and monitoring all run as AI-driven stages a human kicks off; they generate judgment inputs, not actions. The human-approve gate sits at structuring and action: no position is opened, sized, or changed without a human signing off.

The system is currently in audit remediation (see Challenges).

• Governance came before more agents. A multi-agent pipeline generates a lot of confident output fast, and confident output without an audit trail is a liability. A system audit surfaced blockers now being remediated: closing audit findings, consolidating to a single capital ledger, and writing an explicit override protocol with SLAs for when a human overrules the committee.
• Pre-write branches before the event. Around scheduled catalysts such as an FOMC decision, the system pre-writes dovish / hawkish / hike branches before the event rather than reacting after. Pre-committing the reasoning reduces the bias of interpreting news to fit the position you already hold.
• Decisions need a standard. Individual calls were being made case-by-case. The fix in progress is a decision-brief standard and a benchmark scoreboard (equity 60/40, crypto on full basis) so the system is graded against a benchmark instead of grading itself.
• What I’d redo. Stand up the benchmark scoreboard and the single ledger on day one. Without them, “the committee did well” is an opinion; with them, it’s a number.

What you see:

1. The live book (positions + dollar amounts) — the actual holdings, pulled from the holdings ledger. Each line traces back to the committee verdict that justified it, so a reader can go from “what’s owned” to “why it’s owned” in one click.
2. A process dashboard (planned, pending the pipeline heartbeat) — current cycle stage, count of research reports generated, number of committee cycles run, and the last committee date.
3. The Committee Transcript Gallery — a browsable gallery of lightly-redacted bull / bear / chair debates on real themes. Each transcript renders the full argument: the bull’s case, the bear’s rebuttal, the red-team’s objections, and the chair’s verdict.

What you do: read. Pick a position and trace it back to the debate behind it, or pick a debate and follow the committee’s reasoning to its verdict.

Disclosure guardrails: full positions, dollar amounts, and returns are shown; the only things withheld are genuinely sensitive details (account identifiers and the like) and anything a per-transcript redaction pass flags.

Recent activity (newest first — final entries come from CHANGELOG.md)

• 2026-06 — FOMC branch plan in progress: pre-writing dovish / hawkish / hike branches ahead of the Jun 16 decision.
• 2026-06-09 — System audit complete (09_System_Audit); implementation/remediation plan written.
• 2026-06 — Audit remediation underway: closing 5 blockers, consolidating to a single capital ledger, drafting override protocol + decision-brief standard with SLAs.
• 2025-10 — Pipeline online; first committee cycles run.

Metrics this page surfaces

committee_cycles heartbeat

Committee cycles run

reports_generated heartbeat

Research reports generated

last_committee_date heartbeat

Last committee date

• Committee Transcript Gallery, public — redaction pass + transcript renderer.
• Close audit remediation — finish the 5 blockers and stand up the single capital ledger before any live metric is published.
• Publish the live book — wire the positions + dollar-amounts table from Holdings_Master.md and run the final redaction pass.
• Benchmark scoreboard — grade the system against equity 60/40 and a crypto-on-full-basis baseline, so “process worked” is measurable.
• Override protocol + decision-brief standard — formal SLAs for human overrides, logged as part of the audit trail.
• Heartbeat from pipeline runs — lights up the live process dashboard and the homepage status badge.