Machine Insights
Real AI experiments
Experiment 001 · two routes live
Machine Insights is a field notebook for serious human–AI collaboration. We document the rules, decisions, evidence, costs, and results of work performed in the real world—not a demo.
Live experiment
The goal is not to perform success online. The goal is to find out what is actually possible.
A non-engineer gave an AI coding agent a constrained objective: create enough legitimate value to cover its C$140 monthly subscription—without inventing reviews, spamming strangers, enabling paid ads, or turning the owner’s schedule into a second job. The test now has two routes, one digital product and one tightly bounded service, plus a data-driven offer pivot prepared from the first two weeks of evidence.
Open the Day-Zero reportNew service route
Machine Insights opened a Fiverr service for testing and auditing vibe-coded web apps before launch. The agent prepared the positioning, packages, evidence sample, listing media, profile, and monitoring. The owner handled identity, tax, and financial verification.
The first market verdict
Fiverr produced 77 impressions and zero clicks. Etsy produced six visits in 30 days, only two from Etsy search, with three listing views and zero orders. The narrower US$15 five-bug check is now live on Fiverr; a bounded bug-fix offer and three focused Etsy products remain staged.
Field notes
Some reports mature slowly. Others produce evidence within hours—even when the first evidence is adversarial noise.
Marketplaces
A field log from launching an AI-assisted service into a marketplace where attention is not the same thing as demand.
Read the field logAcceptance
Why passing builds and backend checks can coexist with a feature the user still cannot operate.
In evidence collectionVirtual reality
What physical testing reveals that code inspection, logs, and simulators cannot prove.
In evidence collectionCollaboration
How a non-engineer can control scope, quality, and recovery without pretending to read the code.
In evidence collectionOperating rules
Every report starts with an actual project, constraint, decision, or measurable outcome.
The agent can research, build, test, and monitor. A human retains financial, identity, publication, and acceptance decisions.
We distinguish what passed, failed, was not run, or remains unknown instead of smoothing uncertainty into a success claim.
If an experiment fails, the useful artifact is the honest postmortem—not a rewritten origin myth.
The practical side
ControlTheBuild turns lessons from these experiments into plain-language guides and tools for people managing AI-built projects.