research-problem-scout

v3.4 portfolio discovery extension

For CVPR_BEST_PAPER_ASPIRATION or multi-motivation discovery, call research-motivation-portfolio first. Build a broad, source-grounded portfolio of different questions (default 5–8 honest motives) then explain why 2–3 have more important, defensible shortcomings, and research nearest prior work for every shortlist member. Only after this can you turn a surviving motivation into a technical Idea. See ../research-idea-discovery/references/motivation-portfolio-and-cvpr.md. A strong single-problem analysis alone is not a completed portfolio-first discovery.

Research Problem Scout — 先找真问题，再考虑 Idea

Read ../research-idea-discovery/references/motivation-first-gate.md, ../research-idea-discovery/references/motivation-and-story-protocol.md, and ../research-idea-discovery/references/search-connectors.md.

Perform bounded, neutral reading of actual benchmarks, strongest simple baselines, counterexamples, negative results and matched-condition comparisons. Produce up to three problem cards, not method proposals. Seek adverse data as actively as supportive data. Cite each Table/Figure/setting or prove evidence is unavailable. Never claim the root cause was discovered from one comparison.

Rank by problem evidence quality and scientific importance, NOT how easily three innovations can be invented. Explicitly test the null explanation: “Would fair tuning, a different metric or a matched backbone make this disappear?” Check dataset and regime scope to avoid overgeneralizing from one condition. This cheap scan may precede resource-intensive dataset feasibility; do not impose downstream G0 before deciding which problem merits further work.

For each problem call the structural scripts/motivation_gate.py when local commands are available, but also critically read its sources. The script cannot independently verify citations or scientific importance. Output PASS_FOR_IDEATION / PROBE_ONLY / REJECT_OR_REFRAME / RESEARCH_MORE; only PASS may trigger the Hypothesis Lab to generate IDEA_SEED and the costly D0/G0/N/L/M2/method design workflow. For PROBE_ONLY, offer one bounded fairness/replication check and a stop condition, with no method architecture. If all fail, report why and one targeted neutral Deep Research prompt. Label any role-playing review ROLE_SIMULATED.

v3.5 paper-anchored probe

Use research-paper-anchor-scout for quick-start research: ask counterfactual questions at each real Figure/Table, replicate failures, investigate why successful methods work, and reconcile paper disagreements fairly. All still require M0; manuscript limitations or a method's popularity alone are not established failures.