Three defect families account for 121 of the 132 blockers in a 170-goal planning sweep

The gate layer runs before the critic. A plan that breaks a hard rule never reaches the reviewing model, so the critic only spends tokens on plans that are already valid against the rule set ^[1]. The gates make no LLM calls at all, and the post clocks them at about 4.7 seconds for 1,295 tests ^[8], roughly 3.6 milliseconds each ^[5].

Two of the three families are decidable in code. The requirement that some earlier task establish every precondition is a reachability question over the task graph. The requirement that no task run before its prerequisites is a topological check. Those two families hold 103 of the blockers between them ^[1]. In the model-serving example the post prints, the plan cut traffic over to 100 percent with nothing verifying that the 10 percent and 50 percent stages were healthy ^[16].

The third family does not reduce that neatly. The weak_rollback blocker the post quotes says "Rollback only switches to single-write mode without addressing the potential inconsistencies dual-write may have introduced" ^[15]. Deciding whether a rollback actually undoes a dual-write is a judgement about data state, not a graph property, and that family drew the fewest blockers of the three ^[6].

The model comparison is the softest part of the account. The author ran GPT-4o as planner with the small model as critic, then GPT-4o in both roles, and reports the same defect pattern in better prose ^[9]. The post does not include blocker counts for those two configurations, so the conclusion rests on his characterisation of the pattern. "You cannot prompt your way out of a structural problem," he wrote ^[13].

The revision loop has a defined stopping condition. The planner fixes one blocker and introduces another, reshuffles task order without closing the dependency gap, or attaches rollback to the wrong task ^[11]. He is careful about where the fault sits: across revisions the critic found the same blockers reliably, so what failed was the planner's ability to repair structure ^[12]. By the hundredth goal, he wrote, "I had stopped being surprised and started being annoyed, because the plan was plausible and still wrong" ^[14].

For those three names to be your top three, your gates have to check those three properties, because a blocker count is a count of rule hits. The 121 figure is 92 percent of the 132 concrete blockers ^[3], and eleven fell outside the three families ^[2]. The goals were drawn from 40 domains including identity management, multi-agent ops, SRE, supply chain policy and FinOps ^[2]. The distribution transfers if your planner is drafting multi-step change plans over comparable domains and your rule set encodes the same three checks.

The sweep cost $0.49 for 170 goals ^[3], about a third of a cent each ^[4], with a small model in the critic seat by default ^[9]. The $0.49 covers the models he called. The next build he names is a precondition closer, a deterministic linter that runs after the planner drafts and verifies every precondition ^[17].