Thesis evidence engineering

Thesis Evidence Engineering

Separate discovery from citation. A search result is not a thesis reference.

Evidence lanes

Literature

1. Search OpenAlex, Crossref, Semantic Scholar, publisher pages, DOI resolvers, or configured academic MCP tools.
2. Save results as candidate records.
3. Verify title, authors, year, venue, DOI, and relevant full-text passage.
4. Add selected records to Zotero or a controlled BibTeX file.
5. Use the real citekey in the evidence matrix.
6. Map each citation to the exact claim it supports.

Never scrape Google Scholar result pages. Never create a DOI or BibTeX entry from memory.

Experiments

1. Prefer raw instrument exports, lab records, and native plotting files.
2. Use published small papers and reviewer responses to locate facts, then reconcile them with raw records.
3. Treat presentation-derived values as provisional until traced to raw data.
4. Register conflicts instead of selecting a convenient value.

Computation

Accept a computational result only when an actual case, log, settings, and output exist. Route validation to thesis-computational-evidence.

Source priority

Read references/source-priority.md whenever sources conflict. Read references/evidence-model.md before creating or changing ledgers.

Writing boundary

• Cite literature for general knowledge, prior findings, methods, and mechanism hypotheses.
• Cite experiment IDs or figure/table records for the author's results.
• Label interpretations as interpretations when direct evidence is unavailable.
• Do not cite a review as proof of a specific primary experiment unless the review itself is the claim source.
• Do not cite a source that mentions the topic but does not support the sentence.

Deterministic tools

Audit Markdown citations against a verified bibliography:

python scripts/audit_citations.py chapter.md references.bib --strict \
  --report reports/citations.json

Audit the claim ledger:

python scripts/audit_claim_ledger.py reports/claim_ledger.csv \
  --project-root . --strict --report reports/claims.json

Build an evidence-matrix shell from search candidates and BibTeX:

python scripts/build_evidence_matrix.py candidates.csv references.bib \
  --output literature/evidence_matrix.csv

Read references/citation-workflow.md for Zotero and CSL responsibilities.

Stop conditions

Block final prose when:

• the citekey is missing;
• metadata conflicts with DOI/publisher records;
• the supporting passage cannot be located;
• a numerical claim lacks an experiment or model source;
• a mechanism claim exceeds the analytical method;
• the same sample label maps to conflicting data and no authoritative correction exists.