rubric20-semantic (#154): 'report but don't score' state has no output encoding or denominator rule

[realmarcin]

Context

Follow-up from review of #154 (rubric-review branch). This is the most important pre-merge gap.

Problem

#154 rewrites the rubric20 Applies to: fields into a conditional form:

"Always report results of this question, but only score if [human subjects / governance / shared datasets / software / publication] is identified elsewhere."

This introduces a third question state — reported but not scored — that has no representation in the output. The output JSON still emits a numeric score per question and computes overall_score.total_points / percentage against a fixed max_points.

Impact

With no encoding or denominator rule, an N/A question will either:

• be scored 0 → penalizes core / non-applicable datasets, or
• be silently dropped → percentages become non-comparable across datasheets.

This is especially damaging for D4D-core datasheets (claudecode_agent_core), which by design trip many of these "not applicable" questions.

Proposed fix

1. Define an explicit convention in the output schema, e.g. applicable: false and score: null for reported-but-unscored questions.
2. State the denominator rule: excluded questions are removed from max_points so percentage is computed only over applicable questions (and document this so core vs. full-schema scores stay comparable).
3. Mirror the same convention in the batch EvaluationSummary aggregation.

Scope

• .claude/agents/d4d-rubric20-semantic.md (output format + scoring rules)
• Apply analogously to d4d-rubric10-semantic.md if/when conditional applicability is added there (see companion issue).

[realmarcin]

Adversarial review (Codex) verdict: KEEP — blocking. Strengthened with new evidence.

Confirmed and now backed by concrete evidence: the conditional "report but don't score" prose at .claude/agents/d4d-rubric20-semantic.md:209 (and 225/241/259/275/291/307/323/354/397) coexists with a fixed denominator:

• .claude/agents/d4d-rubric20-semantic.md:455 → "max_points": 84
• .claude/agents/d4d-rubric20-semantic.md:720 → "Maximum Possible Score: 84 points"

The committed baselines already bake in 84 (e.g. data/evaluation_llm/rubric20_semantic/concatenated/VOICE_claudecode_agent_evaluation.json:71), and the batch summary schema computes percentages against a fixed maximum (D4D_Evaluation_Summary.yaml overall/method/project blocks). So dropping questions per-file makes a project/method look better or worse purely from how many questions were excluded, while percentages are still reported as if comparable to the historical 84-point baselines.

Therefore this issue's fix must include, beyond the applicable:false/score:null convention:

1. Both a raw and a normalized_percentage (and an excluded_max_points) in the output + summary schema, OR keep all 20 questions in the denominator; and
2. Regenerating/migrating the committed *_evaluation.json baselines once the convention lands, or they become non-comparable.

(This absorbs Codex "Finding 2" rather than filing a duplicate.)