Extracting Literature Requirements

Given one PDF and a research question, produce exactly three outputs:

1. Study profile (keyword-level)
2. Ranked requirement-candidate table (≤15 rows)
3. Lists: contradictions/limitations, unusable parts, uncertainties

No design principles. No evaluation beyond the text. No knowledge outside the PDF. If the PDF is unreadable or empty, state this immediately and stop.

Progress:
- [ ] Step 1: Confirm PDF is readable; if not, report and stop
- [ ] Step 2: Read full text; build study profile (type, sample, context, main finding)
- [ ] Step 3: Extract candidate statements tied to the research question's layers
- [ ] Step 4: For each candidate, locate verbatim sentence + section/page
- [ ] Step 5: Classify type (finding / recommendation / implemented feature / conjecture)
- [ ] Step 6: Assign layer (1-8), write mechanism/rationale from text only
- [ ] Step 7: Assess transferability (direct vs indirect) with brief justification
- [ ] Step 8: Rank candidates by relevance to RQ; cap at 15, note remainder by layer
- [ ] Step 9: Collect contradictions/limitations (verbatim) stated by authors
- [ ] Step 10: List unusable parts (half-sentence each) and flag own uncertainties

Step 2 detail: Study Profile

Extract only what's stated, in keyword form:

• Type: empirical / conceptual / review / design study / experiment — "not in text" if absent
• Sample: N, population, context of study — "not in text" if absent
• Context: setting (e.g., university course, lab study) — "not in text" if absent
• Main finding: two sentences, taken from or tightly paraphrased from the text, citing location

Step 3-6 detail: Candidate Extraction

For each candidate statement:

• Phrase as "The system should ..., so that ..." — but only if this follows logically from the source sentence. Do not invent justifications not supported by text.
• Pull the exact verbatim sentence (max one sentence) that supports it. If the supporting evidence is only in the abstract, mark "abstract only".
• Classify Type:
  • finding: empirically observed result
  • authors' recommendation: explicit prescriptive statement by authors
  • implemented feature: a system/tool feature described as built and used
  • conjecture: author speculation, future work, hypothesis not tested
• Assign exactly one Layer (1-8 per the scheme below). If ambiguous, pick primary layer and flag in uncertainty list.
• Mechanism/rationale: state only the causal or explanatory logic the paper itself gives (e.g., "increases awareness because X"). If paper gives no mechanism, write "not in text".
• Transferability:
  • direct: applies straightforwardly to AI-chat self-regulation dashboards
  • indirect: applies to a different context (e.g., general SRL, LMS dashboards); briefly state the gap

Layer reference:

Step 8: Ranking & Overflow

Rank by direct relevance to the RQ (self-regulated AI use, metacognitive processes, dashboard design). If more than 15 strong candidates exist, output the top 15 and append: "N additional candidates exist, primarily in layers [X, Y, Z]."

Step 9-10: Contradictions, Limitations, Unusable Parts, Uncertainties

• Contradictions/limitations: verbatim sentence + brief note on what it contradicts or limits (e.g., re: dashboards generally, re: self-regulation claims, re: AI-chat context specifically)
• Unusable parts: half-sentence summary of sections not relevant to RQ (e.g., "Section 4.2: unrelated statistical validation of survey instrument")
• Uncertainties: flag your own classification judgment calls (layer assignment ambiguity, borderline type classification, paraphrase vs. verbatim boundary cases)

Example 1: Input: PDF states: "Students who received weekly summaries of their query patterns reported increased awareness of over-reliance on AI for factual recall (p<.05)." Output:

Example 2: Input: PDF only discusses general SRL theory in education (not AI-specific), stating: "Monitoring is a prerequisite for effective control in self-regulated learning cycles." Output: