editorial-review

A plan that is not code

Nothing in Compass mentions software. This plan is for writing a comparison piece, and it works exactly like the engineering one — which is the point: the model is domain-neutral, and this example exists so that claim is tested rather than asserted.

The situation

Three agent-memory tools claim the same ground and nobody has compared them on one workload. The piece is worth writing only if the comparison is real: run all three on one dataset, draft from it, and get an editor's sign-off before publishing.

Then one tool changes its storage engine mid-benchmark, so the dataset is no longer comparable across all three. The plan narrows to the two that held still, and says so.

What the lineage says

001  writer   Worth writing only if the comparison is real.
                 + benchmark   Run all three tools on one identical workload
                 + draft       Draft the comparison from the dataset
                 + publish     Editorial sign-off before publishing

002  writer   One tool changed engines mid-benchmark; the dataset is not
              comparable. Narrowing to the two that held still, and saying so.
                 ~ benchmark   Run the two stable tools on one identical workload
                 ~ draft       Draft the comparison, two tools, name the third's instability

What to look at in the files

• Acceptance is not about tests. benchmark accepts on a reviewed dataset; publish accepts on any(review(verdict=approved, actor=editor), waiver(actor=editor)) — an editor's approval, or an editor's explicit waiver. A human judgement is as valid a criterion as a passing test, and reads naturally.

• actor=editor is bound to the recorded author, not a typed attribute. Compass records who made an evidence claim; it does not adjudicate whether the claim is true. review(actor=editor) is satisfied by an event Compass recorded as authored by the editor — not by anyone who types editor into an attribute. Trust is "who claimed this", visible, and no stronger than that.

• The vocabulary — dataset, artifact, review, waiver, and their attributes — is this use case's own. Compass ships none of these words. They are typed constructors the author defines, so a misspelled attribute is a compile error, and Compass validates the shape of the criterion without ever knowing what "waiver" means.

Files

This plan has no name. Its identity — its PlanId — is the content hash of its origin, the first version (decision 0017), so the plan directory is cfe4f8d721d2, the same hash the 001 file carries. To a person it is its goal, "Publish a defensible comparison of agent-memory tools"; the hash is only for exactness.

• 001-cfe4f8d721d2.ts — the plan
• 002-043517240262.ts — narrowed to two tools