ai-research-skills/docs/example-topic-dossier.gaps.yml at main · WenyuChiou/ai-research-skills · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
dossier: LLM applications in water resources
generated: "2026-05-22"
run_type: gap-to-topic re-test (research-hub plugin v0.3.6 — §1 applies search --screen)
downstream_consumer: research-design-helper   # v0.3.11+ contract reader (Stage 3a)
recall:
  tool: research-hub search --adversarial --screen --json
  query_phrasings: 6
  unique_papers: 435
  tool_confidence: medium
  dossier_confidence: medium
  screen:
    gate: fit-check BM25 relevance gate (search --screen)
    retrieved: 25
    kept: 25
    screened_out: 0
    tier: cold-start
    note: >-
      No blatant >=5x bimodal split in the returned batch, so the gate
      (conservative by design) kept all 25. The agent still applied
      relevance judgement when selecting the prior-art corpus.
  note: >-
    semantic-scholar backend rate-limited (HTTP 429) part of the run;
    recall rests on crossref + openalex + arxiv.
pipeline:
  - research-hub search --adversarial --screen --json   # §1 step 1
  - literature-triage-matrix -> literature_matrix.md     # §1 step 2 (on-topic kept:true)
  - .bib from on-topic search --json results             # §1 step 3
  - gap-to-topic gates §1-§4
gaps:
  - id: G1
    name: "LLM-driven water resources management"
    statement: >-
      Use LLMs / LLM agents for water resources management and decision
      support - hydrological forecasting, reservoir/river operations,
      monitoring, planning.
    type: B
    open: occupied
    dead_end_status: not-assessed   # Gate (1) no-go
    contribution_type: not-assessed # Gate (1) no-go
    feasibility: not-assessed       # Gate (1) no-go
    verdict: no-go
    verdict_reason: >-
      Fails Gate (1) - occupied, firmly: primary studies plus a hydrology
      LLM benchmark (HydroLLM), a systematic review (AI in water
      regulation) and a bibliometric analysis of LLMs in hydrology.
    linked_claim: null
  - id: G2
    name: "LLM agents for human behaviour in socio-hydrology"
    statement: >-
      Use LLM agents to represent heterogeneous stakeholder/household
      decision-making in coupled human-water (socio-hydrology) modeling,
      calibrated and validated against observed water-use/adaptation
      behavior.
    type: B
    open: open
    open_confidence: medium
    dead_end_status: partially-attempted
    dead_end_evidence: >-
      Schuck 2026 (10.3389/frwa.2026.1749745) - a peer-reviewed perspective
      cautioning on LLMs for human data in human-water research - engages
      the exact G2 premise as a standing caution.
    contribution_type: borderline
    borderline_reason: B1   # partial novelty - Batista 2025 (LLM sentiment
                            # for water governance), Braga 2025, Khaki 2025
                            # are weaker realised forms of the capability
    feasibility: feasible-with-effort
    verdict: conditional-go
    verdict_reason: >-
      Open (medium confidence), partially-attempted on dead-end history,
      borderline/B1 on contribution, feasible-with-effort. Worth-it call
      handed to researcher + advisor (dossier section 4).
    linked_claim: null
open_questions:
  - id: Q1
    text: >-
      Is G2 still open after a full-recall re-run with semantic-scholar
      enabled? The openness verdict is medium-confidence pending that.
  - id: Q2
    text: >-
      Can G2's validation design answer Schuck 2026's caution AND
      out-distance Batista 2025's LLM-sentiment approach to water
      governance? Batista 2025 is now a close in-domain analogue.
  - id: Q3
    text: >-
      Does the empirical-calibration-against-observed-water-behaviour
      requirement make G2 problem-solving or incremental? Turns on framing
      (B1).
  - id: Q4
    text: >-
      Can an existing socio-hydrology dataset/model (e.g. Jiang et al. 2026,
      North China Plain) be reused as the behavioural-validation base, or
      does G2 need primary data collection? Binding constraint on timeline.