novelty-check

novelty-check

A discipline layer, not a judgment layer. It does not decide whether something is novel — it refuses un-grounded novelty claims and structures the ones that pass.

What a novelty verdict must contain (the structure)

For each contribution the paper asserts:

contribution:
  id: contrib-1
  claim: "<verbatim from abstract/intro>"
  decomposition:
    method: "<what technique>"
    domain: "<what object/population>"
    finding: "<what was shown>"
    metric: "<what was measured>"
  anchors:                               # ≥ 5 required
    - canonical_id: <cid>
      closest_aspect: method|domain|finding|metric|scale
      delta: "<what's different in this paper>"
      delta_sufficient: true|false
  verdict: novel | incremental | not-novel
  confidence: 0.0-1.0
  reasoning: "<short, specific>"

A verdict without this shape is disqualified.

Agent-facing procedure

1. For each contribution the target paper claims, build the decomposition tuple.
2. Search adjacent prior work to fill anchors. Use semantic-scholar citation/reference walks + paper-search for non-indexed sources.
3. For each anchor, state the delta in one sentence. Mark delta_sufficient per delta — is the delta large enough to count as novel for this venue?
4. Commit to a verdict per contribution with a confidence number.
5. Emit the full structure as JSON and pipe to the gate:

uv run python .claude/skills/novelty-check/scripts/gate.py \
  --input /tmp/novelty-report.json \
  --target-canonical-id <cid-of-paper-under-review>

The gate exits non-zero if:

• Any contribution has fewer than 5 anchors
• Any verdict is novel but all delta_sufficient are false
• Any verdict is a hedge word ("maybe", "potentially", "could be")
• Confidence is missing

On pass, the report is written to the paper's artifact under novelty_assessment.json and a row is inserted into novelty_assessments.

Principles this enforces

From RESEARCHER.md: 6 (Name Five), 7 (Commit to a Number), 9 (Premortem — each contribution's verdict is challenged).

What this skill does NOT do

• Does not synthesize the paper — it assumes you've already read it
• Does not produce a publishability verdict (that's publishability-check)
• Does not run attack vectors (that's attack-vectors)