When using AI for literature triage in physics (e.g., surfacing related models, known limits, and contradictions), which specific division of labor between AI and humans—for instance, AI clustering and tagging papers by claim type plus contradiction flags, with humans only inspecting a small curated frontier set—maximizes early detection of conflicts with prior work without flooding researchers with low-relevance citations or spurious contradictions?

Answer

Best division of labor: AI does broad, structured scanning and contradiction flagging; humans review only a small, ranked “frontier set” plus anything that blocks a key claim.

1. AI roles

• Claim extraction: Parse abstracts/intros to tag each paper with rough claim types (e.g., new model, new limit, numerical confirmation, contradiction) and key equations/parameters.
• Clustering: Group papers by (model family, regime, observable) rather than by keyword alone.
• Contradiction mining: For each draft claim (exponent, sign, scaling, regime), search for papers with explicitly incompatible statements and attach short quote+equation snippets.
• Relevance ranking: Score items by (a) direct parameter overlap, (b) formal similarity of equations, (c) citation proximity to already-trusted core papers.
• Frontier curation: Produce 3 short lists: (i) “likely direct conflicts” (5–15 items), (ii) “nearby but weaker conflicts/edge cases” (10–30), (iii) “supporting or consistent” work.
• Uncertainty & noise tags: Mark flags as low/medium/high confidence and distinguish “semantic ambiguity” vs “hard numerical/sign conflict.”

2. Human roles

• Pre-commit focus: Before seeing AI lists, humans name the few claims where conflicts would be most damaging (e.g., key exponents, new regime boundaries). AI prioritizes frontier sets for these.
• Frontier review only: For most projects, humans read in full: all “likely direct conflict” papers and only a sample of “nearby” ones, ignoring long AI citation dumps.
• Escalation rule: Any AI flag marked high-confidence hard conflict must be checked by reading the original section, not just the snippet.
• Override & feedback: Humans re-label a small fraction of AI flags each session (e.g., “not actually a conflict,” “true but irrelevant”), which AI uses to adjust clustering/ranking.

3. Safeguards against floods and spurious contradictions

• Hard caps: Default caps like ≤15 high-priority conflict candidates per focal claim, ≤40 total triage items per weekly cycle.
• Evidence-type tags (reuse): Each AI-surfaced prior result is tagged as {analytical, numerical-only, heuristic, review} to guide how seriously to treat a conflict.
• Provenance display (reuse): Every conflict flag shows raw quote, equation, and link; no free-text AI adjudication like “this disproves your claim.”
• Separate stress-tester pass (reuse): Keep contradiction mining as a discrete pass after initial modeling, so it doesn’t continuously spam suggestions while ideas are still very rough.

4. Why this division helps

• AI is used for breadth, clustering, and first-pass inconsistency spotting, where scale matters most.
• Humans are used for deep reading and judgment on a small, ranked frontier set, where expertise matters most.
• Caps, evidence tags, and provenance act as epistemic safeguards that reduce both overload and over-trust.

This setup is likely to improve early detection of real conflicts with prior work while containing noise, but it still depends on consistent human use of caps and escalation rules.