One human, several agents, and a review queue that grows with agent throughput instead of shrinking with it. That is the honest state of code review in this factory, and skimming diffs faster is not a strategy, it is review theatre with a shorter attention span. The question worth asking is not “how do I read faster” but “what is the actual unit of review,” and the answer this lab bet on is: not the diff.
The actual problem
A diff answers “what changed.” It does not answer “was this the right thing to make true,” and a tired human reading fifteen diffs in a row will, eventually, substitute plausibility for verification. That substitution is invisible from the outside: the PR gets approved, the code looks reasonable, and nobody can tell which approvals were real reads and which were pattern-matching on shape. Blocking on full reads defeats the point of running agents in parallel in the first place. Something has to give, and it should not be rigor.
What got scaffolded
Intent-assurance-lab took an architecture-first pass at a different unit: every change carries a falsifiable intent statement (what will be true after this change, stated so it can be checked, not just described), an evidence list that discharges it, and a review protocol that checks two things in order: is this the right goal, and does the evidence actually establish it. Diff-reading becomes a targeted follow-up where the evidence is weak, not the default unit of review.
The scaffold went further than a schema. Confidence profiles are categorical, not numeric, because a fake-precise number invites false trust. Review depth is risk-tiered T0 through T3, so a one-line config fix and a vendor-egress trust boundary do not get the same ceremony. And the target was never a standalone tool: the design points at Hekton’s core, federating evidence to the labs that already produce it rather than building a separate assurance runner that duplicates what exists.
The decision
Adopt the intent-assurance model as the target shape for review in this factory, architecture-first: design the contract before any implementation exists, so the review-unit change is deliberate rather than something the factory drifts into because reading diffs got too slow to keep up.
What next
This post is honest about where the system actually is: designed, independently reviewed, architecturally placed in the wider factory, and then parked, not proven by running code. A real Phase 0 trial did happen: a hand-written intent for a sibling lab’s still-unstarted work, using exactly the kind of vendor-egress trust boundary this piece treats as an illustrative T2 example. It proved the schema and checklist are usable by hand, on real unstarted work, and it surfaced a genuine judgment call worth recording before code exists. It did not prove the framework changes review behavior in practice, since no implementation of that work existed yet to test whether a reviewer’s attention was actually redirected by the record. The next post in this pair covers that trial in full: the candidate task it had to rule out first, and the rule that ruled it out.
What Next
Don't Write Intent for Code That Already Exists
The framework's first real trial almost ran against work that had already started. The rule that saved it: verify unstarted, don't assume it.