Evidence over vibes

Agent-authored code failed a quiet test long before it failed CI: we started trusting narration. A model explains a patch eloquently, opens a pull request, and the review becomes a reading comprehension exercise. The missing artifact is not another summary. It is a receipt.

What a receipt is

A useful receipt answers questions that prose ducks:

• Did the project still build?
• Did types, lints, and formatters agree with the story?
• Did public APIs drift?
• Did smoke tests run, and what did they observe?
• Which dependencies changed, and was that change justified by the diff?

Aftermath exists because those questions should not depend on the agent’s self-esteem. Verification has to be independent, re-runnable, and durable enough to outlive the chat that produced the change.

Why independence matters

If the same system that wrote the code also grades the code, you have theater. The interesting architecture is a second plane: something that examines artifacts after the fact and is allowed to disagree. That disagreement is the product.

A bias worth keeping

Ship tools that make “show me” cheaper than “trust me.” Fluency is abundant. Evidence is not.