eprint-search

Commands

S=skills/eprint-search/scripts/eprint_search.py
python3 $S "blind rotation" "programmable bootstrapping" --months 12      # union of phrasings, last 12 months
python3 $S "oblivious key-value store" --crossref --min-year 2019          # + published versions with DOI
python3 $S "division property" --title-only                                # title search only
python3 $S "..." --json > hits.json                                        # for scripting
python3 $S --selftest

Exit status 2 means no source answered (offline or blocked) — that is "novelty check not performed", never "no hits". DBLP sometimes serves a bot challenge to scripts; the script detects it and you should add --crossref.

Novelty-check procedure ("is it already taken?")

1. Write the idea as one falsifiable sentence (what object, what improvement, under which assumptions).
2. Generate 5–10 phrasings: the technique name, the problem name, the object name, the synonym used by the neighbouring community (e.g. "digit extraction" / "lowest digit removal" / "floor function"), and the name of the closest known baseline.
3. Run them with --months 12 (always) and without a window (for the classic literature). Add --crossref for published-only venues.
4. Read the abstract of every hit whose title shares two content words with the idea; open the PDF of any hit that shares the object and the technique. Snowball: check the "related work" of the 2–3 closest hits and the papers that cite the main baseline (Crossref/Google Scholar "cited by").
5. Verdict per idea, logged in LITERATURE.md → ## Novelty checks: - YYYY-MM-DD | query: "..." | sources: eprint,crossref | window: 12m | hits reviewed: N | verdict: clear|overlap:<key>|taken:<key>
   • clear: nothing shares object + technique.
   • overlap:<key>: same object or technique, different claim → add a matrix row and a precise "our delta".
   • taken:<key>: same claim → stop, tell pi-orchestrator; salvage is a new decision, not a rewording.
6. Re-run every 2–4 weeks while the project is live and once more right before submission (ePrint moves fast; the lit-matrix linter fails if the newest check is older than 30 days).

Pitfalls

• A concurrent ePrint published during your project is normal; record it with its date — reviewers accept "concurrent and independent" only if you can show when you knew.
• Do not trust search-engine summaries of a paper's content; read the paper.
• Title-only search misses papers that use a different name for the same object — always also run full-text.