When an earlier investigation changes a diagnosis

Research before running a diagnostic.

1. Retrieve the prior investigation.
2. Group copies by incident and source family.
3. Read material counterexamples.
4. Keep identical symptoms with different causes separate.
5. Recheck current applicability before proposing an action.

Count independent support.

MATCH (episode:YsgLessonDemoEpisode)-[:FROM_SOURCE]->
      (:YsgLessonDemoSource {source_id: $source_id})
WHERE episode.stance = 'supports'
RETURN count(episode) AS cited_occurrences,
       count(DISTINCT episode.incident_family_id) AS independent_incidents,
       count(DISTINCT episode.source_family_id) AS independent_source_families

Two passages from one retrospective describe one incident family. They are two citations and one evidentiary source, not two independent votes.

Read the contrary evidence.

MATCH (episode:YsgLessonDemoEpisode)-[:CITES]->(evidence:YsgLessonDemoEvidence)
WHERE episode.stance IN ['counterexample', 'unresolved']
RETURN episode.episode_id AS episode, episode.stance AS stance,
       episode.incident_family_id AS incident_family,
       episode.same_symptom AS same_symptom,
       evidence.excerpt AS selected_excerpt,
       evidence.excerpt_sha256 AS excerpt_sha256
ORDER BY stance, episode

BEFORE CITING A FIELD, ESTABLISH WHAT WRITES IT.

Role: support · excerpt 7a85a1af25d8

The habit that catches all of them costs seconds: find the writer before you
quote the reader.

Role: support · excerpt ebcfc002373b

Also worth stating: all three dead were idle, but two idle runners survived, so
"idle runners die" does NOT hold — I recorded that explicitly so nobody inherits
a false correlation.

Role: counterexample · excerpt f05990d2330b

OVER-AGGREGATION IS THE CHARACTERISTIC ERROR OF THIS SEAT. I was caught three
times in one night by IDENTICAL ERROR TEXT FROM A DIFFERENT CAUSE — a test file,
a verification_reason word-for-word identical to another note's, and a refusal
code. Adjacent-but-different failure codes are the highest-risk moment. Name the
shape a failure set has, say what it implies, and let the outliers stay outliers
rather than forcing one explanation over all of them.

Role: boundary · excerpt ce68d6c3e08c

Revise the lesson and retain the boundary.

MATCH (lesson:YsgLessonDemoLesson {lesson_id: $lesson_id})
RETURN lesson.original_hypothesis AS original_hypothesis,
       lesson.hypothesis_state AS state_after_counterexample,
       lesson.revised_lesson AS revised_lesson,
       lesson.remaining_check AS remaining_read_only_check,
       lesson.review_basis AS review_basis,
       lesson.model_calls AS model_calls

Original hypothesis: An idle runner is dead.

State: contradicted

Reviewed lesson: An idle label is insufficient. Establish what writes the field, then inspect the per-entity artifact before explaining the state.

The excerpts are real. The concise hypothesis and revised lesson are an authored demonstration fixture. No model produced them.