Shipped four approval-console tickets to the always-on box in one day
Every ticket survived three adversarial critic rounds with real blocks before it went live.
8 entries
Every ticket survived three adversarial critic rounds with real blocks before it went live.
Two mockups got rejected. The third was built desktop-first from panels and computed styles measured off a live reference app, read-only.
Ten tickets, every one run past a critic before it could ship, and roughly ninety root-cause fixes landed before night one.
From a competitor teardown to a fully autonomous daily pipeline in two days: news sweep, de-slop-gated script, cards, cloned voice and music, render, avatar bubble, upload and schedule with read-back.
A credential broker that types secrets from 1Password without the value ever entering the container, a closed nine-operation browser vocabulary behind an origin allowlist, and a readiness ladder proven green twice.
A two-arm A/B, identical agents on identical work with only one arm's guardrails removed, against a sealed failure criterion and seven sealed predictions. The criterion fired and all seven predictions were wrong.
The agent server now serves its own decision console over Tailscale with TLS. Findings, drafts and approvals flow through one recorded writer, and the first real approval was executed end to end from the console.
A reimplementation of the Dreams API curation pattern: extract, gate, single-model judge, then a deterministic script that re-verifies every evidence quote verbatim and enforces a two-session promotion gate.