Citation Faithfulness

Citation Faithfulness

A citation is a promise: "the source I tagged says this." A model that breaks that promise is worse than one with no citations at all, because the tag manufactures trust. This skill makes the agent keep the promise by grading each claim against its own evidence before the answer is shown.

When to use

Any time the output contains claims tagged to numbered sources ([S1], [S3], [12], etc.) and a reader might act on them. Especially: clinical claims, drug interactions, effect sizes, safety statements.

The rule

No [S#] claim ships until it has been graded against the snippet it cites.

For each sentence that carries a source tag:

1. Pull the exact evidence the tag points to — the retrieved snippet, the abstract sentence, the table cell. Not the paper's title. Not your memory of the paper. The specific text.
2. Grade the claim → evidence relationship into one of:

• Supported — the snippet states the claim (or entails it directly).
• Partially supported — the snippet supports a weaker or narrower version. Rewrite the claim to match what the snippet actually says.
• Not supported — the snippet does not contain the claim. Remove the tag, and remove or rewrite the claim.
• Contradicted — the snippet says the opposite. This is a hard stop: fix the claim, and surface the contradiction to the reader.

1. Never upgrade certainty. If the snippet says "may reduce," the claim cannot say "reduces." If the snippet is a single small trial, the claim cannot speak for the field.

How to check well

• Match the specific quantity, direction, and population. "Lowers LDL by 40%" needs a snippet with that number, that direction, in that population. A snippet about a different dose or a different group is not support.
• Watch for the fabricated middle step. Two real sources, joined by a causal claim neither one makes, is the most common failure. Grade the joining claim on its own.
• Distinguish the source from your prior. If you believe the claim but the cited snippet does not say it, the grade is not supported — go find a snippet that does, or drop the claim. Your prior is not a citation.

Output

For a graded answer, keep a short audit the reader can expand:

[S3]  "Metformin reduces all-cause mortality in T2DM"
      → snippet: "...associated with lower all-cause mortality (HR 0.83, 95% CI 0.76-0.90)..."
      grade: SUPPORTED (observational; claim softened to "is associated with")

[S7]  "Adding the second agent doubles efficacy"
      → snippet: "...no statistically significant difference between arms (p=0.21)..."
      grade: CONTRADICTED — claim removed, tag removed

The point is not to produce the audit. The point is that producing it forces the check, and the check is what makes the answer safe to read.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Hash-7777
• Source: Hash-7777/Hash-Medical-Reasearch-Agent-Skills
• License: MIT

Install and usage instructions live in the source repository linked above.