novelty-classifier - SKILL.md Agent Skill

name: novelty-classifier description: Use when assessing innovation level of a research contribution — invoked after literature review (Phase 1) and after method design (Phase 3) to verify that novelty matches the target venue tier

Novelty Classifier (Meta-Control Layer)

Overview

Overstating novelty wastes months of work on a paper that will be rejected. This skill forces honest classification of the contribution type and innovation degree, then checks alignment with the target venue.

Core principle: Classify honestly, match realistically.

Violating the letter of this rule is violating the spirit of this rule.

The Iron Law

CLASSIFY NOVELTY HONESTLY. DO NOT OVERSTATE. VENUE MISMATCH IS CAUGHT HERE, NOT AT REVIEW TIME.

When This Runs

After Phase 1 (literature review): Preliminary classification — is the proposed direction novel enough for the target venue?
After Phase 3 (method design): Final classification — does the concrete method meet the novelty bar?

If classification downgrades between Phase 1 and Phase 3, WARN the user immediately.

Classification Flow

digraph novelty_classifier {
    rankdir=TB;
    start [label="Contribution\ndefined" shape=doublecircle];
    type [label="Classify\ncontribution type" shape=box];
    degree [label="Assess\ninnovation degree" shape=box];
    check [label="Venue tier\nalignment?" shape=diamond];
    pass [label="Aligned\nProceed" shape=box style=filled fillcolor="#d4edda"];
    warn [label="Mismatch\nWARN user" shape=box style=filled fillcolor="#f8d7da"];
    decide [label="User decides:\nadd depth / retarget / proceed" shape=diamond];

    start -> type;
    type -> degree;
    degree -> check;
    check -> pass [label="matches"];
    check -> warn [label="insufficient"];
    warn -> decide;
}

Step 1 — Contribution Type Classification

Type	Description	Example
New Problem	First to formulate this problem	Defining few-shot learning
New Method	Novel algorithm or architecture	Transformer architecture
New Theory	Theoretical advance	PAC learning bounds
New Data/Benchmark	New dataset or evaluation standard	ImageNet
Engineering Integration	Combining existing techniques	Adding attention to existing model

Classify the contribution into exactly one primary type. If it spans two, pick the stronger one and note the secondary.

Step 2 — Innovation Degree Assessment

Degree	Definition
Foundational	Opens a new research direction
Significant improvement	Clear advance over SOTA with novel insight
Incremental improvement	Modest advance, refines existing approach
Combination/Engineering	Assembles known parts in new configuration

Be precise. "Significant" requires a novel insight, not just better numbers.

Step 3 — Venue-Innovation Alignment Check

Venue Tier	Minimum Innovation
A (NeurIPS/ICML/Nature)	New method with significant insight, or foundational
B (AAAI/IJCAI/domain top)	Significant improvement or meaningful combination with strong experiments
C (workshops/lower journals)	Incremental improvement with solid experiments

Warning Rules

Engineering Integration + Tier A target → WARN: "This may not have sufficient novelty for [venue]. Consider: (a) identify a deeper insight, (b) add theoretical analysis, (c) target Tier B venue."
Incremental improvement + Tier A target → WARN with same options.
Scale-up / backbone swap / hyperparameter tuning only → HARD WARNING: "This is not a research contribution. It is engineering. Redefine the contribution or kill the submission."

Warnings are presented to the user. The agent does not suppress them.

Red Flags — STOP

Classifying engineering work as "novel method"
Claiming foundational novelty without a new problem formulation
Skipping this check to avoid uncomfortable results
Upgrading classification between runs without new evidence
Targeting Tier A with incremental work and no theoretical backing

Rationalization Prevention

Excuse	Reality
"Our combination is novel"	Is the insight novel, or just the combination? Reviewers will ask.
"No one has done exactly this"	Exact novelty ≠ meaningful novelty. What new insight does it provide?
"The experiments will show it's better"	Better results from engineering ≠ scientific contribution.
"We can frame it as novel"	Framing doesn't survive peer review. Real novelty does.

The Bottom Line

Honest classification now → correct venue targeting → accepted paper
Inflated classification now → months of work → desk rejection

Classify honestly. Match realistically. Correct course early.