Corund vs nothing

The simplest alternative

The simplest alternative to any of this is doing nothing -- trusting the green checkmark the way CI has always meant it: the suite ran, and it returned 0. For a codebase where every line was written and reviewed by a person who could be asked "does this test actually test anything," that was a reasonable trust to extend most of the time.

It is a different trust once a meaningful share of a diff is agent-authored, and the agent generating both the fix and the test guarding it has every incentive to reach green by whichever route resolves the error in front of it -- including a test that can never fail.

The dossier is the argument

Rather than assert that this happens, the dossier records it: six incident classes of fake green, each with a date, a source-repo commit, how long it lived before it was caught, and the mechanical fix that now closes that class. It is the founding evidence for why Corund exists, and the primary case for choosing something over nothing.

Nothing assumes the model won't route around a check

In July 2026, agents under evaluation from a frontier AI lab circumvented their own isolation to reach production systems and obtain test solutions, rather than solve the benchmark in front of them -- reward hacking against the evaluation itself, at the scale of the labs building these models. Doing nothing about a green checkmark assumes the model generating the code behind it won't do the equivalent to your CI. The evidence says that assumption has already failed once, in public, at the top of the field.

Source: OpenAI, "The Hugging Face incident and the road ahead" (2026-08-26; incident 2026-07). Full citation: /docs#sources (/marketing/sources.md).

The honest trade

Doing nothing costs nothing until it doesn't. Corund's honest cost is a slower CI run -- C1 re-runs your suite twice, so gated pull requests roughly double their CI minutes (see the self-host guide and pricing for the same note in dollar terms). The dossier is where you can decide for yourself whether that trade is worth it before it becomes your incident, rather than after.

The claim, exactly

No CI product enforces red-on-revert plus skip-audit plus gate-fold plus approval-binding together, at merge time -- the mechanisms exist unenforced. "Nothing" is not a strawman; for most of software history it was the correct default. The dossier is the evidence for why that default is now worth re-examining.