← Back

SecondBrain

Prototype

Signature capability

Cycle 1 ran end to end on live local data: 8 candidate findings survived an adversarial refuting verifier (which deflated inflated EV estimates to honest bands), surfacing ~135 qualified sales leads sitting cold for 20 days and driving 20 outreach emails that actually got sent.

What it is

SecondBrain is a hierarchical opportunity-discovery loop built around four levels: L1 VENTURE (durable operator context), L2 TASK (the current pass/fail standard), L3 CANDIDATES (ranked proposals awaiting approval), and L4, a 366-line read-only `l4_metrics.py` that pulls a real money scorecard from live local sources (lead pipelines, quant PnL via sqlite read-only mode, launchd job failures, API spend). The refutation gate is the core: findings are re-checked by a separate refuting verifier that deflates inflated dollar estimates to defensible bands before anything is approved to memory. It is deliberately fail-closed for safety, nothing touching outbound identity, a credential, or money is ever auto-executed; those land in MANUAL_STEPS.md for Jordan. The metrics collector is stdlib-and-curl only, fail-isolated per source so one dead input never aborts the run, idempotent, and ships with a `--selftest` that asserts JSON round-trips and type invariants.

Highlights

  • L4 scorecard is 366 lines of pure-stdlib Python: read-only, idempotent, never writes/trades/spends tokens, with each data source fail-isolated so one broken input can't abort the pass
  • Refutation gate re-checks every candidate and deflates inflated EV numbers to realistic bands; in cycle 1, 8 of the findings survived and were ranked by rechecked ROI, not first-pass optimism
  • Fail-closed safety model: anything touching outbound identity, a credential, live trading, or a GUI grant is never auto-run; it routes to MANUAL_STEPS.md for Jordan's click
  • Real output, not a demo: found 135 qualified leads (57 with emails) uncontacted for 20 days, plus a Discord webhook 404 spooling 45 unseen alerts and a launchd job silently exiting 127
  • Auto-fixed the safe waste it found (reddit_api throttle, ccusage PATH) and drafted 20 em-dash-clean outreach emails, all 20 later sent 2026-07-06
  • `--selftest` mode asserts top-level keys, list/int types, and JSON serialisability, so the scorecard's shape is verified before it's trusted

Tech · Python 3 (stdlib only: sqlite3 read-only, subprocess, glob, json), ruff, curl, launchd, markdown-as-state (L1-L4 files), git