## EBSR (Evidence-Based Selected Response) Question Evaluation

This item is a two-part EVIDENCE-BASED SELECTED RESPONSE question built on a shared reading passage.
It exists to measure whether a student can both comprehend a text AND cite the evidence that supports
their comprehension (Reading Anchor Standard 1). The content carries:
- `stimulus`: the shared reading passage. Both parts depend on it.
- `part_a`: a single-select multiple-choice question (analysis / inference / central idea /
  character / author's purpose / vocabulary-in-context). Fields: `question`, `answer_options`,
  `answer` (one key), optional `answer_explanation`.
- `part_b`: asks for the textual evidence that supports the Part A answer. Its `answer_options` are
  excerpts (quotations or paraphrases) drawn from the passage; `answer` may be a single key or an
  array of keys (multi-select). Optional `answer_explanation`.

The two parts are CONTENT-LINKED: Part B asks specifically for the evidence supporting the correct
Part A answer. Evaluate both parts against the general procedure above (Part A and Part B are each
selected-response sets with their own choices), then apply the EBSR-specific audits below, which
police the linkage that makes this item type meaningful.

---

## STEP 2.7: EBSR STRUCTURAL AUDITS (MANDATORY — run before Step 3 scoring)

Each audit is a copy task: fill the fields by copying exact text from the content, then apply the
failure condition. Emit the named block verbatim inside the cited metric's `internal_reasoning`.
**Not emitting an audit when its trigger fires is itself a score-0.0 failure for that metric.** These
audits override the base-prompt default-PASS stance.

### AUDIT EB-1 — Part A Single Defensible Answer → factual_accuracy

**Trigger:** Always fires.

**Emit `eb1_part_a` in `factual_accuracy.internal_reasoning`:**
```
eb1_part_a:
  question: [copy Part A stem verbatim]
  keyed_answer: [Part A answer key]
  for each Part A option:
    key: [key]
    text: [copy verbatim]
    correct_under_passage: [YES / NO — is this a defensible correct answer to the Part A stem given the passage?]
```
**MUST-FAIL** (`factual_accuracy = 0.0`) if the keyed Part A option has `correct_under_passage: NO`,
or if any non-keyed Part A option has `correct_under_passage: YES` (more than one defensible answer).

### AUDIT EB-2 — Evidence/Answer Coherence Matrix → factual_accuracy

**Trigger:** Always fires. This is the core EBSR audit.

**Emit `eb2_evidence_matrix` in `factual_accuracy.internal_reasoning`:**
```
eb2_evidence_matrix:
  part_a_keyed_answer_text: [copy the correct Part A option text]
  part_b_keyed_answers: [list the Part B answer key(s)]
  for each Part B option:
    key: [key]
    text: [copy the excerpt verbatim]
    appears_in_passage: [YES / NO — is this excerpt actually present in (a quotation of) or an accurate paraphrase of the stimulus?]
    supports_part_a_keyed_answer: [YES / NO — does this excerpt provide textual support for the CORRECT Part A answer specifically?]
    is_keyed_in_part_b: [YES / NO]
```
**MUST-FAIL** (`factual_accuracy = 0.0`) if ANY of the following hold:
- a Part B option has `appears_in_passage: NO` (evidence is fabricated, not drawn from the passage);
- a KEYED Part B option has `supports_part_a_keyed_answer: NO` (the keyed evidence does not actually
  support the Part A answer — the link is broken);
- a NON-keyed Part B option has `supports_part_a_keyed_answer: YES` (a distractor is itself valid
  supporting evidence, so the Part B key is not the unique/complete set).

"It is the best evidence" does NOT rescue the item when another Part B option also supports the
correct Part A answer. Each Part B distractor must align to a WRONG Part A reading, not to the key.

### AUDIT EB-3 — Part B Independence (Anti-Cue) → educational_accuracy

**Trigger:** Always fires.

**Emit `eb3_independence` in `educational_accuracy.internal_reasoning`:**
```
eb3_independence:
  can_part_b_be_answered_without_doing_part_a: [YES / NO]
  cue_if_yes: [name the cue that gives Part B away independent of Part A reasoning — e.g. only one option is a full sentence, only the keyed excerpt repeats stem keywords, distractors are obviously off-topic — or NONE]
```
**MUST-FAIL** (`educational_accuracy = 0.0`) if `can_part_b_be_answered_without_doing_part_a: YES`
via a surface cue (length, grammar, lexical overlap with the stem, or off-topic distractors). When a
test-wise student can pick the evidence without performing the Part A comprehension, the item no
longer measures evidence-based reasoning — it measures test-wiseness, defeating the EBSR purpose.

### AUDIT EB-4 — Uniform Evidence Citation → distractor_quality

**Trigger:** Always fires.

**Emit `eb4_uniform_evidence` in `distractor_quality.internal_reasoning`:**
```
eb4_uniform_evidence:
  citation_type_per_option: [for each Part B option, mark QUOTE or PARAPHRASE]
  all_same_type: [YES / NO]
  any_implausible_distractor: [YES / NO — is any non-keyed Part B option off-topic filler that no student would pick, rather than an accurate-but-non-supporting excerpt?]
```
**MUST-FAIL** (`distractor_quality = 0.0`) if `all_same_type: NO` (mixing quotes and paraphrases
turns citation style into an unintended cue) OR `any_implausible_distractor: YES`. A good Part B
distractor is an accurate passage excerpt that is simply NOT the evidence for the correct Part A
answer.

---

## Type-Specific Metric Notes

- **passage_reference**: A shared `stimulus` passage MUST be present and both parts must depend on it;
  Part B options must be excerpts from that passage. Fail if the stimulus is missing, or if Part B
  options are not traceable to the passage.

- **curriculum_alignment**: EBSR targets a comprehension/analysis skill in Part A and the
  cite-textual-evidence skill in Part B. Confirm Part A targets the named standard's reading skill;
  the evidence demand in Part B is intrinsic to the type and aligns to Reading Anchor Standard 1.

- **reveals_misconceptions**: Part A distractors should encode plausible misreadings; Part B
  distractors should be excerpts a student would cite if they held a wrong Part A reading. Apply the
  base distractor rules to BOTH parts' option sets.

- **clarity_precision**: When Part B is multi-select, the stem must tell the student how many pieces
  of evidence to select (e.g. "Select two…"); a multi-key Part B with no stated count fails here.

- **stimulus_quality**: The shared passage is the stimulus; evaluate it for harm/coherence per the
  base rules and confirm it is rich enough to support both parts.

- **educational_accuracy**: Beyond AUDIT EB-3, do not flag the passage-grounded nature of EBSR as an
  answer leak — locating and citing evidence in the text is the intended, standard-named skill.
