Require DeepSeek for literature screening

SKILL.md

@@ -14,9 +14,10 @@ structure, properties, mechanism, device/application performance, and limits.
 
 It can also route through the user's Zotero/Obsidian literature-note workflow:
 extract keywords from the manuscript topic or draft, use QQnote-generated
-Zotero child notes or paper tables as literature inputs, separate strongly
-related and weakly related papers inside QQsci, then use those papers to improve
-positioning, introduction logic, comparison, novelty framing, and citations.
+Zotero child notes or paper tables as literature inputs, ask DeepSeek to screen
+those candidate notes into strongly related and weakly related papers under
+QQsci's rules, then use those papers to improve positioning, introduction
+logic, comparison, novelty framing, and citations.
 
 ## Core stance
 
@@ -123,8 +124,10 @@ DeepSeek/AwesomeGPT Zotero notes generated by QQnote or equivalent sources.
 2. If the target journal is absent, ask the author for it. If they want advice,
    open `references/journal-positioning.md` and recommend several tiers.
 3. Open `references/literature-routing.md` and perform QQsci's own relevance
-   screening against the existing notes or QQnote-generated comparison tables.
-4. Screen notes into:
+   routing against the existing notes or QQnote-generated comparison tables.
+   The large-scale strong/weak paper screening pass must be run by DeepSeek when
+   DeepSeek is configured.
+4. Ask DeepSeek to screen notes into:
    - `strongly related`: same material, same material family plus same device,
      same mechanism, same target journal/neighbor journal, or direct benchmark.
    - `weakly related`: same direction, adjacent material family, same

references/configuration.md

@@ -80,7 +80,8 @@ for QQsci's manuscript reasoning.
 QQsci owns:
 
 - manuscript keyword extraction
-- strong/weak related-paper screening
+- DeepSeek screening prompts, rules, and review of strong/weak related-paper
+  results
 - target-journal positioning
 - comparative review against strong papers
 - submission-package audit
@@ -93,4 +94,7 @@ QQnote owns:
 - generating paper comparison tables from Zotero items
 - optional QQnote note-state maintenance when explicitly requested
 
-QQnote output is an input to QQsci, not the decision engine.
+DeepSeek owns the large-scale screening pass over QQnote/Zotero/Obsidian note
+candidates when QQsci needs strong/weak related papers. QQnote output is the
+candidate corpus; QQsci owns the criteria, prompt, sanity check, and downstream
+writing brief.

references/literature-routing.md

@@ -44,6 +44,15 @@ Return a query pack:
 
 ## DeepSeek screening brief
 
+Strong/weak related-paper screening must be executed with DeepSeek when a
+usable DeepSeek configuration is available. QQsci prepares the rules and inputs;
+DeepSeek reads the QQnote/Zotero/Obsidian note candidates and returns the
+screened set.
+
+Do not ask Codex to silently classify a large note library from memory. Codex may
+do a small sanity check after DeepSeek returns results, but DeepSeek is the
+screening engine for the note set.
+
 Give DeepSeek a compact, structured task:
 
 ```text
@@ -68,6 +77,59 @@ Screen the notes and return:
 4. Do not include papers only because of a broad keyword match.
 ```
 
+## DeepSeek output contract
+
+Require DeepSeek to return structured Markdown or JSON that can be handed to
+QQsci and then QQwrite:
+
+```json
+{
+  "keyword_pack": {
+    "must_match": [],
+    "should_match": [],
+    "context": [],
+    "exclude": []
+  },
+  "strongly_related_papers": [
+    {
+      "title": "",
+      "year": "",
+      "journal": "",
+      "doi": "",
+      "zotero_key": "",
+      "relation_type": "",
+      "usable_for": [],
+      "why_strong": "",
+      "notes_evidence": ""
+    }
+  ],
+  "weakly_related_papers": [
+    {
+      "title": "",
+      "year": "",
+      "journal": "",
+      "doi": "",
+      "zotero_key": "",
+      "relation_type": "",
+      "usable_for": [],
+      "why_weak": "",
+      "notes_evidence": ""
+    }
+  ],
+  "rejected_near_matches": [
+    {
+      "title": "",
+      "doi": "",
+      "reason": ""
+    }
+  ],
+  "missing_or_uncertain_metadata": []
+}
+```
+
+If the output will feed QQwrite, QQsci should convert the screened papers into a
+`qqwrite_brief`-style section plan rather than passing raw notes directly.
+
 ## Strong vs weak relation rules
 
 Strongly related:

scripts/material_literature_router.py

@@ -1,9 +1,9 @@
 #!/usr/bin/env python3
 """Create a manuscript keyword pack and DeepSeek screening prompt.
 
-This helper is intentionally local-config only. It does not call DeepSeek or
-Zotero by itself; it prepares a reproducible prompt that the skill can use with
-the user's configured note-search route.
+This helper is intentionally local-config only. It prepares a reproducible
+DeepSeek prompt for the required screening pass over QQnote/Zotero/Obsidian
+literature notes. It does not call DeepSeek or Zotero by itself.
 """
 
 from __future__ import annotations
@@ -81,8 +81,16 @@ Manuscript profile:
 Return:
 1. Strongly related papers: same material, same material family plus same device, same mechanism, direct benchmark, or same target-journal positioning.
 2. Weakly related papers: same direction, adjacent material family, same characterization logic, similar device architecture, or useful writing pattern.
-3. For each paper: title, year, journal, Zotero key if present, relation type, useful claim/evidence, and why it matters for the manuscript.
+3. For each paper: title, year, journal, DOI, Zotero key if present, relation type, usable_for, useful claim/evidence, and why it matters for the manuscript.
 4. Reject title-only broad keyword matches.
+5. Return rejected near matches and missing/uncertain metadata separately.
+
+Output JSON keys:
+- keyword_pack
+- strongly_related_papers
+- weakly_related_papers
+- rejected_near_matches
+- missing_or_uncertain_metadata
 """