Explain

W05 · 5 h

Generation, citation, and absorption

Essential question

Was a source merely displayed, or did it shape the answer?

What you should be able to do

  1. Score citation correctness and completeness.
  2. Distinguish mention, citation, entailment, and absorption.
  3. Diagnose a claim-level attribution failure.
PrerequisitesW02 claim ledger.W03–W04 pipeline vocabulary.
Builds on

W02 claim ledger. · W03–W04 pipeline vocabulary.

Investigates

Was a source merely displayed, or did it shape the answer?

Feeds

L04 CES graph review.

Arrive with a prepared artifact

Reading route
Core PAPER-24/PAPER-33; Extend PAPER-13; foundation candidates A05/A07/A08.
Viewing route
Open the complete text-first lecture packageFive-minute annotation demo with a compound claim and misplaced citation. The current equivalent is notes, slide script, worked case, and no-video transcript; no recording is claimed.
Readiness check
Segment one supplied paragraph into atomic claims.
Bring
Five annotated claims with exact source passages.

Concepts, assumptions, and boundary

Citation has several failure modes

A response can contain fluent claims and visible references while remaining weakly supported. Citation correctness asks whether the cited passage entails the adjacent claim; completeness asks whether claims that need evidence are covered; source quality asks whether the source is fit for the claim. These axes should not be replaced by raw citation count. Claim segmentation must occur before scoring, because one sentence may contain both supported and unsupported predicates.

Absorption is stronger than display

Mention records whether an entity appears. Citation records whether a source is visibly attributed. Entailment records whether the source supports a claim. Absorption asks whether source-distinctive information is present in the generated answer in a supportable way. Establishing absorption is difficult when many sources contain the same fact, so the label must include an alternative-source check and an uncertainty state. Interface post-processing can also show or suppress citations without changing generation, making surface behavior another explicit stage.

Inspect the mechanism or evidence structure

Diagnostic matrixAttribution and absorption diagnostic
Attribution and absorption diagnosticFour citation-support cells distinguish supported citation, unsupported citation, supported but uncited claim, and unsupported uncited claim. Beneath them, absorption is labeled absent, plausible, or source-distinctive with required evidence.Cited + supportedclaim-level decisionCited + unsupportedclaim-level decisionUncited + supportedclaim-level decisionUncited + unsupportedclaim-level decision
Figure design. A cited/not-cited × supported/not-supported matrix with a separate three-level absorption strip and examples in every cell. Absorption is a separate, source-distinctive axis.
Long description

Four citation-support cells distinguish supported citation, unsupported citation, supported but uncited claim, and unsupported uncited claim. Beneath them, absorption is labeled absent, plausible, or source-distinctive with required evidence.

One action, one feedback state

Action

Annotate claims in five frozen responses and compare with a rule-based draft.

Feedback

Disagreements expose segmentation, entailment, source-quality, and alternative-source assumptions.

Accessible alternative

A CSV annotation sheet and adjudication guide reproduce the complete task.

Open the claim–evidence mapper

Produce a reviewable intermediate file

Task
Double-code a subset of frozen responses and adjudicate every disagreement.
Inputs
Responses, source passages, CES schema, annotation guide, adjudication log.
Timebox
70 minutes
Intermediate file
annotation-v1.csv and disagreement-log.md
Stop condition
Mark unresolved rather than force agreement when the source or claim boundary is insufficient.

Why each controlled source is here

Complete, retrieve, and revise

Checkpoint
L04 CES graph review.
Reflection
Describe a case where citation count improves while evidence quality does not.
Revision
Re-run claim coverage after adjudication and report the denominator.
Low-compute route
All responses and sources are frozen; manual annotation is the core route.

Five retrieval questions

01How do correctness and completeness differ?

Correctness tests whether attached citations support claims; completeness tests whether all evidence-requiring claims receive adequate support.

02What is a source-distinctive claim?

Information whose wording or content can be traced to the source more specifically than to common background knowledge or many interchangeable sources.

03Why can UI behavior confound citation analysis?

A platform may add, reorder, or hide citations after generation, so visible attribution and generated content can change independently.

04What is the evidence boundary for W05?

Benchmark definitions and findings are task- and model-bounded; course labels must be operationalized and tested for agreement.

05What must you submit or revise after this week?

L04 CES graph review. Re-run claim coverage after adjudication and report the denominator.

Continue in the practice package