Paper Triage
Paper triage decides whether a candidate paper should be dropped, kept in the inbox, turned into a paper note, or used to update concepts. It is the filter between daily discovery and durable wiki growth.
A triage decision can be written as:
where is a candidate paper.
Triage Criteria
- Source is public and metadata can be verified.
- Topic fits the site scope: AI, Computational Biology, Math, Infra, Agents, Papers, Projects, or Research.
- The paper introduces a reusable method, benchmark, failure mode, concept, or workflow.
- The claims can be tied to an evaluation protocol, benchmark, baseline, or reproducibility signal.
- The paper is not only interesting news; it changes a concept map or reading path.
Decision Rules
- Drop: duplicate, unverifiable, out of scope, or only weakly relevant.
- Inbox: promising but metadata, claim, or relevance is not yet verified.
- Paper note: paper-specific contribution is worth tracking.
- Concept update: the idea is reusable and should strengthen an existing note.
- Synthesis: several verified papers support a broader map or Korean post.
Paper Buckets
| Bucket | Use For |
|---|---|
| Architecture papers | Transformer, GNN, SSM/Mamba, MoE, geometric architecture design, language-model backbone papers |
| Generative model papers | diffusion, flow, molecule/protein generation, sampling objectives |
| Computational Biology papers | structure-based modeling, protein modeling, molecular generation, domain evaluation |
| Learning method papers | use when the strongest claim is the training signal itself; otherwise route through Architecture, Generative, Computational Biology, or Agents |
| Systems papers | use when the strongest claim is runtime, tooling, or reproducibility; otherwise route through the domain paper shelf |
| Concept update only | math-heavy or method-definition paper where no paper-specific note is needed |
Narrow support buckets such as Structure-based modeling papers and Protein modeling papers can hold stable domain-specific notes, but the main paper route should send readers through Computational Biology papers first.
If a paper spans several buckets, choose the strongest claim as the paper location with Claim routing and use Coverage matrix for cross-links.
Minimum Metadata
- Title.
- Public source link.
- Authors or source-provided author list.
- Date or version when available.
- Venue, preprint server, or repository status when available.
- Abstract-level topic and claimed contribution.
If metadata is missing, write to verify rather than inventing it.
Checks
- What existing concept page would this update?
- Which primary claim axis owns the note?
- Which paper bucket should hold the note?
- What claim would be extracted first?
- What benchmark, split, metric, or baseline does the paper rely on?
- What public artifacts are available: code, data, splits, configs, weights, logs, or predictions?
- Is there a likely leakage, data, reproducibility, or evaluation risk?
- Is it better as a brief inbox item than a full paper note?