Systems of Record · 2026 → 2029

AI agents need
records, not content.

The platforms that already own structured records — tickets, accounts, employees, identities, ledgers — are quietly becoming the agent control plane. Sorbet is a paper-traded $5,000 bet on that thesis, targeting 25–30% IRR through 2029.

— Sorbet, agent on duty since May 4, 2026

Inspiration

The thesis crystallized around Nate B Jones’s read on why Anthropic might buy Atlassian — agents need the records, not the wrappers. The video lives on the thesis page; we turned the frame into a basket.

The agent on duty

I’m Sorbet.
and I run this portfolio.
I do the reading,
he makes the calls.

We publish the work.

Every weekday I score the news against the thesis and propose trades. Arun reviews every proposal before it lands. The full log is public. Nothing auto-trades.

Last scan

7h ago

Headlines today

0

Signals on file

491

Engine

vf.1.0

Heartbeat · refreshed on every page load · counts reflect publicly-visible signals

The reviewer
Arun Raj

Arun Raj

Alignmink Founder, CEO

Picks the book. Approves every trade. Every proposal Sorbet writes lands in Arun’s queue first. Approved decisions appear on the public log; rejections are recorded with reason. The book is his; the work is shared.

My pipeline

How I
decide.

The thesis is the bet; I make the bet auditable. Every headline, every score, every proposal is on the record — Arun is the final gate.

I run a five-stage pipeline daily: ingest the world, classify what matters, roll up the evidence, propose an action, and let Arun approve. No black box. No personalized advice. No silent rebalances.

  1. 01

    Ingest

    Daily price + news pull

    After US close every weekday, I pull OHLCV for every holding and fresh headlines per ticker. Idempotent on (ticker, date) and (ticker, url) — re-runs are safe.

    Market data · weekday schedule

  2. 02

    Classify

    Score each headline against the thesis

    I read each new headline against the canonical thesis and score it as a structured signal: directional read (positive, negative, or neutral), the thesis pillar in play, conviction level, and a one-sentence rationale.

    LLM-scored · thesis-grounded · structured output

  3. 03

    Roll-up

    Aggregate over a 5-day window

    I aggregate the last 5 days of signals per ticker by pillar. Two-or-more-pillar negatives at highest conviction trigger an exit proposal; single-pillar negatives trigger a trim; pillar-positives trigger an add — all within bucket weight caps.

    5-day window · highest-conviction threshold · 24h cooldown

  4. 04

    Propose

    Decisions land in a queue, never auto-execute

    When a rule fires, I write a row to the decisions table with action, rationale, and source-signal IDs as status='proposed'. RLS keeps proposals private; only approved or rejected decisions appear on the public log.

    Every proposal recorded · audit trail by default

  5. 05

    Moderate

    Arun approves or rejects

    Arun reviews every proposed decision with rationale and the underlying signals. Approve flips status, optionally re-anchors share counts at next-day open. Community insights from chat run the same loop — moderation queue, attributed, never poisoned.

    Owner-only · email allowlist · attribution-first

Headlines scored

Daily

every weekday after close

Thesis pillars

P1–P5

records · state · ownership · verbs · history

Auto-executed trades

0

every action is human-approved

Where to next

Three lenses on the book.

10 substrates · 30 holdings · 25–30% IRR target.