skill-claude-generate-hypothesis

Claude Code skill for generating testable scientific hypotheses for the CVGL / NADEZHDA research project. Analyzes the vault's literature library, extracts evidence from analyzed papers, deduplicates against existing hypotheses, and produces structured If / Then / Because claims ready to drop into 2_hypotesis/.

Behaviour spec — see SKILL.md. This file is the human-facing entry point.

Installation

Drop this repo's contents into your vault's Claude Code skills directory:

git clone https://git.lissad.keenetic.name/Pikaliov/skill-claude-generate-hypothesis.git \
    .claude/skills/generate-hypothesis

Or as a submodule, if you prefer to track upstream:

git submodule add https://git.lissad.keenetic.name/Pikaliov/skill-claude-generate-hypothesis.git \
    .claude/skills/generate-hypothesis

After installation Claude Code picks up the skill on the next session. Verify with /help — /generate-hypothesis should appear in the user-invocable list.

Vault prerequisites (referenced by the skill but not bundled here):

• 1_lit_research/{2_backbone,3_fusion,4_reduction,5_multimodal_approach,6_cvgl}/ — paper analyses (P/B/F/R/M-prefixed)
• 1_lit_research/СИНТЕЗ_всех_статей_для_LUPI_CVGL.md — synthesis index
• 2_hypotesis/cvgl/HYP_*.md — existing hypotheses (used for dedup)
• 0_prompts/tag_taxonomy.md — tag whitelist for frontmatter
• CONVENTIONS.md — repo-wide RU/EN split rules

If you adapt the skill to another vault layout, update the path globs in SKILL.md §"Фаза 2: Поиск литературы".

When to use

• Designing a new experiment and need a falsifiable claim before writing code.
• Synthesizing 5–8 related papers into one actionable research direction.
• Exploring "what if we replaced X with Y" without re-reading the entire vault.
• Stress-testing an idea against existing hypotheses to find duplicates or contradictions.

Invocation

/generate-hypothesis "<research-question>" [Teacher|Student|Fusion|Loss|Edge|Dataset]

Examples

/generate-hypothesis "Поможет ли Coupled Mamba fusion для объединения drone и satellite признаков в Student?" Fusion
/generate-hypothesis "Снизит ли GGeM с FP16 кэшем latency на Jetson?" Edge
/generate-hypothesis "Заменит ли DINOv3-L текущий DINOv2-L Teacher?" Teacher

What it does (6 phases)

#	Phase	Output
1	Parse the question, route to the right NADEZHDA component	component tag, keywords
2	Targeted vault search across 1_lit_research/{2_backbone,3_fusion,4_reduction,5_multimodal_approach,6_cvgl} and the synthesis index	5–8 papers read
3	Dedup against 2_hypotesis/cvgl/HYP_*.md	extend existing or create new
4	Extract method / numbers / limitations / applicability	per-paper evidence rows
5	Synthesize the claim in If / Then / Because form (≤30 words)	hypothesis statement
6	Render via templates/hypothesis_full.md or templates/hypothesis_compact.md with frontmatter, wiki-links, math, tables	drop-in .md for 2_hypotesis/

Output contract

Every produced hypothesis must have:

1. Formal claim — If [method], then [±X% on metric Y on dataset Z], because [mechanistic rationale]
2. Confidence — High / Medium / Low + one-line justification
3. Scope — Teacher / Student / Fusion / Loss / Edge / Full system
4. Evidence — 2–4 papers, wiki-linked, with concrete numbers (R@1, params, FLOPs, latency)
5. Falsifiable test — baseline, metric, threshold, p-value
6. Experimental plan — datasets, stages, timeline
7. Assumptions — what must hold for the claim to be testable
8. Links — extends / contradicts / similar references to existing hypotheses

Hard constraints

• ❌ Никогда не выдумывать статьи, числа или wiki-links на несуществующие файлы. Если evidence отсутствует — confidence = LOW с явной отметкой.
• ✅ Текст — русский, формулы и термины — английский (см. [[../../CONVENTIONS.md]]).
• ✅ Каждое число должно иметь источник (откуда взято).
• ✅ Имя файла: HYP_<метод>_<компонент>_<суть>.md.
• ✅ Теги — только из [[0_prompts/tag_taxonomy]].
• ✅ Перед расширением похожей гипотезы — обязательная проверка 2_hypotesis/cvgl/HYP_*.md, не создавать дубликат.

NADEZHDA component cheat-sheet

Component	Current state	Where the skill looks
Teacher	DINOv3-L, ~356 M frozen, 5 modalities (sat/UAV/street-view/depth/text), Multi-FiLM-Fusion, modality dropout p=0.3	1_lit_research/2_backbone/, 5_multimodal_approach/
Student	SOFIA v7.1 / FastViT-T12 weight-shared, 2 modalities, CVD, GGeM → 512-dim	1_lit_research/2_backbone/, 3_fusion/
Fusion	Multi-FiLM-Fusion (Teacher) / late gated (Student)	1_lit_research/3_fusion/
Distillation	7 losses (InfoNCE, LUPI-MSE, feature alignment, RKD, seg distill, CVD_MI, CVD_Recon) + GradNorm + progressive staging	2_hypotesis/kd_hypotesis/, 6_cvgl/
Edge	Jetson Orin NX, < 50 ms, INT8, ≤ 5 GFLOPs (Tiny) / ≤ 500 MB (M-preset)	1_lit_research/4_reduction/

Detailed taxonomy of methods → component mapping: reference/method_taxonomy.md.

File layout

generate-hypothesis/
├── SKILL.md                    — full skill specification (frontmatter + 6-phase pipeline)
├── README.md                   — this file (human entry point)
├── templates/
│   ├── hypothesis_full.md      — full template with all 8 required sections + tables + math
│   └── hypothesis_compact.md   — short template for "draft" status hypotheses
└── reference/
    └── method_taxonomy.md      — mapping of literature methods → NADEZHDA components

Allowed tools

Read, Glob, Grep, WebSearch — read-only literature analysis. Skill does not write to the vault directly; the model returns the hypothesis as a code block, the user persists it under 2_hypotesis/cvgl/.

Related skills (sibling skills in the host vault, not bundled here)

• /analyze-paper — single-paper deep-dive that feeds evidence into this skill.
• /synthesize-review — cross-paper review that the dedup phase consults.
• /tag-vault — keeps the tag taxonomy in sync so frontmatter tags stay valid.

Worked example

Input:

/generate-hypothesis "Поможет ли Coupled Mamba fusion для объединения drone и satellite признаков в Student?" Fusion

Expected output (sketch):

• Title: HYP_coupled_mamba_Student_fusion.md
• Claim: «Если заменить bottleneck cross-attention в Student на Coupled Mamba SSM fusion, то latency на Jetson Orin NX снизится на ≥ 15 % при сохранении R@1 ± 0.5 % на University-1652, потому что SSM имеет линейную сложность O(L) против O(L²) у attention…»
• Confidence: Medium
• Evidence: [[Coupled_Mamba_NeurIPS_2024]], [[F35_Sigma_Siamese_Mamba]], [[B6_BB_2024_Efficient VMamba]]
• Test: baseline = current cross-attention fusion · metric = R@1 + Jetson latency · threshold = ΔR@1 ≥ –0.5 % AND Δlatency ≤ –15 % · 5 seeds, paired t-test p < 0.05
• Plan: 4 weeks on University-1652 + GTA-UAV, ablation per direction count

See 2_hypotesis/cvgl/ for previously generated hypotheses to study the produced format.