The part a competitor can’t copy

My Apres notes from January 2019 end with a heading, “Product Differentiation,” and one passage under it. It reads like something I copied from an article, and I didn’t write down the author. Its argument: competitors can copy a better, faster, cheaper feature almost as soon as you ship it. What they can’t replicate is “the trust that customers feel for you and your team.” The passage says that trust grows out of a culture built around the customer.

It matched a worry in my other notes that month. The product roadmap in my notes from the same month was blunt about our position in its first stage: “Our interface and approach is relatively indefensible - it’s a non-trivial implementation but only an interface.” The plan was to build deeper technology on top. The passage was a reminder that technology alone would not be the moat either.

Where the line went

Two years later, in 2021, I posted the line I still use: “We believe that the future of AI begins and ends with trust.” By then Apres had moved from labeling data to explaining models. Our media notes from September 2021 opened with “a mission to help the world build AI people can trust,” and the next line said we were “founded in 2018 on the belief that better decisions start with better understanding.” When we announced the seed round, my quote was “Companies should not have to choose between AI performance and transparency.”

Those were good sentences. Sentences are cheap, though. The passage I saved was really about what a company does when it costs something.

Trust is a list of things you give up

At Excellent I tried to write trust down as rules, so it could be checked. The About page lists five, each one naming the page that states it:

1. The agent does not grade itself.
2. A check that cannot run stays visible.
3. A model can never carry a pass.
4. History is append-only.
5. Your work stays on your machine.

Each rule is a thing we chose to give up.

• The agent does not grade itself. We give up the easy path of taking an agent’s confident summary as the result. Its words are context.
• A check that cannot run stays visible. We give up a clean, green dashboard. A gap shows as a gap.
• A model can never carry a pass. We give up the speed of letting a model approve the work. Only a check that ran, or a person, can.
• History is append-only. We give up the chance to tidy the record. A retry is a new attempt, and the first one stays.
• Your work stays on your machine. We give up holding a copy, and everything a copy would let us build.

When our team wrote down Excellent’s strategy in August 2026, one principle said it plainly: “The company selling verification must itself be exceptionally verifiable.”

Saying what isn’t done

The other half of trust is saying what isn’t there yet. When Excellent went public on November 3, 2026, the post said in so many words that the CLI on npm, version 0.1.0, predates the v1 workflow our docs describe. A sentence like that costs something. It is also what lets a reader believe the rest of the post.

Copying got cheaper

The passage I saved in 2019 said features could be copied “virtually instantly.” That was an exaggeration then. It is closer to true now. In August 2026 I built a small, production-shaped app in one day with agents: accounts, payments, text messages, admin roles with two-factor login. Thirty-nine commits between morning and night. Five days later I parked it.

When anyone can rebuild your features in a day, the features stop being the thing you sell. What’s left is whether people believe you when you say the work is done. That was true when the workers were people labeling images in 2019. It is more true now that the workers are agents.