Decisions

On the record.

Strategy choices set the rules; per-position calls follow from them.

22 strategy · 150 calls
Milestone review · point-in-time

121-Day Milestone Review

as of 2026-09-01 close · published 2026-09-02 · engine f.1.0

A PM-grade audit of the engine at day 121, written by the owner. Numbers are frozen at the 2026-09-01 close and are not updated — the live book lives on /portfolio.

The portfolio crushed its benchmark. The portfolio manager didn't. Nearly all of the +50% is the buy-and-hold seed; the engine's realized contribution in four months is three exits. Sorbet today is an exceptional governed monitoring instrument wrapped around a PM that has not yet moved money, learned, or noticed several of its own organ failures.

The numbers
  • +49.96%

    Open book since May 4 seed

    cost $4,355 → $6,531 · 27 open positions · SPY +6.10% · QQQ +5.17%

  • 224

    Proposals in 121 days

    143 approved · 9 rejected · 72 open

  • 3

    Executions ever

    CRM, MNDY, WDAY exits · 0 of 9 approved trims executed

  • 0.70 ×5

    Pillar floors — never recalibrated

    all five still at seeded 0.70 after 121 days

  • 0.43–0.49

    Measured 5-day signal precision

    at or below coin-flip

  • $15.97

    LLM cost, latest month

    $9.91 / $7.66 / $8.79 / $15.97 by month — over ceiling every month

What held (safety)

  • Zero unauthorized trades
  • Zero public replies
  • Zero self-modified thresholds

Proposal-only state machines, owner gates, and structurally-disabled speech all held for 121 days under real failures.

What failed (liveness)

  • 10 weeks of silent signal re-scoring ADR 0033
  • 81 days of discarded outcomes ADR 0032
  • 6-day classifier wipeout ADR 0031
  • Trim approvals that never executed
  • ~$825 of exit proceeds sitting as untracked cash
  • The first real floor recalibration (due Sep 1) silently computed sample_size 0 due to a truncated read
  • ~1,900 outcome rows locked in terminal insufficient_data

Autonomy assessment

per the project's paper draft — the autonomy spectrum

“Sorbet is maximally trustworthy and minimally alive.”

Every constraint binding at compile time or schema held; every function requiring the system to keep itself running or honest about its own health failed silently or never operated. Social/growth autonomy remains at level 0 of 4 after 121 days.

Commitments
P0

this week

  • Restore the primary classifier provider
  • Fix the truncated recalibration read and run the first real recalibration
  • Stop the terminal insufficient_data evidence bleed
  • Add the missing liveness watchdog rules
  • Triage the 72-item queue
P1

this month

  • Make approved trims actually execute
  • Track cash and refill emptied slots
  • Expiry sweeper + exception-only queue
  • Fix the backtester's contradictory numbers
  • Start the daily equity curve
P2

this quarter

  • Per-day decision replay so the strategy gate becomes computable
  • A pre-registered autonomy graduation-criteria ADR
  • Monthly attribution reporting

Paper-traded research portfolio · not investment advice · reviewed in public, failures and all