You found the paper. Does it support the claim?

A writing assistant returns a valid paper link. Before keeping the sentence beside that citation, ask what a reader could verify by following it. A bibliography entry identifies a source; an evidence record explains why that source belongs beside this particular sentence.

I coauthored ScholarCopilot. This AI-assisted note offers an editorial checking workflow, with three handwritten claims. They are not model outputs, a new benchmark, or experimental results. The downloads contain records for human review; they do not automatically decide whether evidence supports a statement.

Where retrieval fits

ScholarCopilot joins scholarly writing with learned citation retrieval. Its [RET] token requests references during generation, and generation and retrieval are trained jointly. That makes it relevant to research on writing assistants and retrieval timing. The mechanism alone does not certify every sentence. See the original paper and official implementation.

A complementary evaluation perspective appears in ALCE, which studies generated text with citations and distinguishes fluency, correctness and citation quality. These works address different evaluation settings; this note does not compare their scores.

Three claims, three decisions

The examples below inspect the ScholarCopilot v2 abstract, not the whole paper. This deliberately small scope makes the evidence boundary visible. Open the source before interpreting a verdict.

The third decision is limited: “not established by the source I checked” does not prove the opposite, nor does it claim the entire paper was searched. In the second row, making the sentence sound less certain would not repair the change in metric. The measured object itself must match.

Use the worksheet on one paragraph

Download the blank record and three completed examples. Each example is explicitly marked as handwritten teaching material. No account or API key is needed to read or edit the JSON files.

1. Copy one substantive sentence exactly. Split it if it contains several independently checkable propositions.
2. Identify the original source and version. A search snippet is a lead; record where you actually inspected evidence.
3. Write the relevant passage location and a short paraphrase. Record the task, population, metric and conditions needed to interpret it.
4. Separate source identity from support. A reachable URL does not settle either question by itself.
5. Choose a revision, record unresolved questions and leave reviewer/time fields empty until a real review occurs.

For a number, check the denominator, measured object, dataset and comparison. For a causal statement, inspect the design supporting that interpretation. If evidence comes from several sources, retain separate records so each source’s contribution stays visible. These are proposed editorial steps, not a validated automatic scoring system.

A useful first exercise is to audit three adjacent sentences in your own related-work paragraph. Keep the original wording beside the revision. The result should let another reader inspect the decision without having to trust the assistant that drafted it.

ScholarCopilot research guide and citation files · All publications