[mindsocket]

Requested by Rog · project thread

Slice 3 of reusable semantic rules: the primitives as plain rules, plus rules researched for each bundled schema, each with an eval in the #228 harness.

Rules

Schema	Rule	Primitive	Scope	Verdicts
strict_ost	outcome-product-outcome	right type	–	product-outcome · business-outcome (info) · traction-metric (info) · output (warning)
strict_ost	opportunity-refines-parent	refines parent	parent	refines · weak-link (info) · misplaced (warning)
strict_ost	opportunities-distinct	mutually exclusive	children	distinct · overlapping · duplicate (warning)
strict_ost	solution-addresses-opportunity	refines parent	parent	addresses · tangential (info) · misplaced (warning)
strict_ost	solution-serves-outcome	faithful to ancestors	ancestors	pass/fail
strict_ost	assumption-category-fits	–	–	pass/fail
strategy_general	goals-distinct	mutually exclusive	children	as opportunities-distinct
strategy_general	goals-cover-parent	collectively exhaustive	children, bodies	pass/fail (info)
strategy_general	goal-advances-parent	refines parent	parent	advances · weak-link · misplaced
strategy_general	opportunity-serves-ancestors	faithful to ancestors	ancestors, bodies	pass/fail
strategy_general	problem-statement-names-no-solution	–	own body	pass/fail
knowledge_wiki	concept-or-entity	right type	–	pass/fail
knowledge_wiki	synthesis-connects	–	own body	pass/fail
okf	okf-description-informative	–	–	pass/fail

Set-level rules sit on the parent with a when that needs two or more children of the type, so they cost one call per such parent. No verdict reports error, so none gate.

Sources. strict_ost type descriptions, rule wording and the assumption categories follow Torres: Continuous Discovery Habits and producttalk.org (opportunity mapping: "a child opportunity … is a subset of a parent opportunity", siblings are "distinct—you can address one without addressing another"; defining product outcomes: business vs product outcomes, traction metrics, outputs in outcome language). strategy_general has no single source, so its descriptions are plain definitions of vision, mission and goal; collectively exhaustive is from MECE (Minto) and complete refinement (KAOS), as in the plan doc.

Evals

14 new evals, 247 cases, 15 to 20 per rule. About a third are near-misses or mixed framings, with a comment on each: an output worded as an outcome, a solution that addresses its opportunity but works against the outcome (a daily email digest that keeps people out of the product), "the lack of a solution" worded as a problem, a named method that is still a concept. Scenarios come in contrast pairs where one detail flips the label.

Four new fixtures in fixtures.json5: ost-trees, strategy-trees, knowledge-wiki, okf. Children-scope rules get one small tree per scenario. Parent- and ancestor-scope rules share parents, since their input holds no siblings. A few labels accept two verdicts where both are defensible; --check passes on all.

Not run live yet: the container has no provider keys. To run on your MacBook:

bun run eval:semantic --provider openrouter-jev --provider openai

Other changes

• Fix: the judge now gets a type's description when its oneOf entry is a $ref. Before, strategy_general and knowledge_wiki sent no type descriptions at all.
• Base fields title, summary and tags no longer have descriptions that only restate their names ("Short summary"), so they drop out of item_types. Their values still reach the judge. parent keeps "Parent node", because otherwise the wikilink def's "An Obsidian wikilink, e.g. [[Parent Node]]" shows instead.
• A test assumed one judge input per node. That's no longer true once a node has rules with different scopes, so it now checks one judgement per rule and node, and one input per scope.
• Docs: a table of the three scope patterns in semantic-rules.md, a line in schemas.md, and a pointer in schema-authoring.md.

Out of scope: rule templates (slice 4, which needs your agreement on the dialect), and harder cases for the existing opportunity-customer-voice eval.

---

Generated by Claude Code

[claude]

Code review

No issues found. Checked for bugs and CLAUDE.md compliance.

🤖 Generated with Claude Code