Content
48%Weight 40%Scale 1-5Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
The contract's sequence and checkpoints are genuinely specified — the MORPHISM pipeline, the Reaches reopening rule, and the per-completion convergence evidence give the workflow real structure — but nearly everything is written as formal Lean with redundant prose restatements, at roughly ten times the token cost the instruction content justifies. The absence of any worked example of a user-facing relay round leaves the protocol's realization dependent on the reader interpreting the formalism.
Suggestions
Replace or drastically compress the Lean block: the operative rules (collect to the limit of your own reach, cite settled facts, keep held values open for the person, show reach for what is unreached) fit in ~30 lines of prose; the current doc comments also duplicate content already in references/judgments.md.
Add one concrete worked example — a short sample relay round showing an open uncertainty with its reach, a settled fact with its citation, and a detection — so the "user-facing realization" is executable without interpreting the formal spec.
Move the formal contract to a reference file and keep a concise overview in SKILL.md with explicit load conditions, and surface the buried `references/judgments.md` §Collection pointer in a visible section instead of a mid-block Lean comment.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dominated by a ~340-line Lean 4 block whose prose doc comments restate the same contract three times over (axiom docs in SKILL.md, then again in references/judgments.md and references/round-composition.md), and the FLOW/MORPHISM/CONVERGENCE comments restate what the definitions already encode — several padded, redundant sections for a behavioral protocol whose rules fit in a fraction of the space. | 2 / 5 |
Actionability | Some guidance is concrete (the grounding table specifies what each op presents, and rules like "name a detection as one, on its own line" are specific), but there is no worked example of an actual relay round, and the user-facing realization defers to interpreting the formal spec ("Present what `surface` names and proceed"), leaving key execution details missing. | 3 / 5 |
Workflow Clarity | The sequence is explicitly written (MORPHISM: "scan → collect → read → relay → fuse"), with checkpoints present: the Reaches branch in `inquire` governs reopening, and CONVERGENCE names the evidence required at each completion ("for every uncertainty raised, how it stands — settled, with its citation… open, with its reach"). Minor gaps come from the practical steps being scattered across formal code and comments rather than stated as an executable sequence. | 4 / 5 |
Progressive Disclosure | Both references (references/judgments.md, references/round-composition.md) are real, one level deep, and conditionally loaded ("Read `references/round-composition.md` before composing when…"), but the body is a monolithic inline 340-line formal block of the size that belongs behind disclosure, and the judgments.md pointer is buried mid-block inside a Lean doc comment ("Direction: `references/judgments.md` §Collection") rather than clearly signaled. | 3 / 5 |
Total | 12 / 20 Passed |