# W05 Planned Transcript — Generation, Source Use, Citation, and Absorption

## Status and use

No W05 recording exists. This document is a planned no-video equivalent, not a transcription of a delivered lecture. The chapter durations below are instructional planning budgets. W05 has not been taught, rehearsed, timed, or piloted, so the stated duration is not an observed media length. The transcript contains every concept assessed in the core route. A later recording may depart editorially, but it must preserve the evidence boundaries, receive checked captions, and be reconciled with this text.

Use this route with the slide alt text or without slides. All visual relationships are repeated verbally. When I ask for a learner check, pause and write an answer before continuing. The route uses the public identifiers `PAPER-24`, `PAPER-33`, `PAPER-13`, and `PLAT-04`; internal catalog keys are not needed.

## Planned chapter budget

| Chapter | Topic | Planned minutes |
|---:|---|---:|
| 1 | One citation, six propositions | 4 |
| 2 | Generation conditions and stage boundaries | 4 |
| 3 | Atomic claims and resolved citation objects | 5 |
| 4 | Entailment, completeness, and correctness | 5 |
| 5 | Presence, absorption, and source use | 5 |
| 6 | Four-cell diagnosis and source fit | 4 |
| 7 | L04 fixture and reproduced counts | 4 |
| 8 | C-018 and C-020 disagreement cases | 4 |
| 9 | Reading papers and platform documentation by design | 3 |
| 10 | Bounded conclusion and handoff | 4 |
|  | **Total planned route** | **42** |

## Chapter 1 — One citation, six propositions

Begin with a modest observation. A response appears on a named generative-search surface. Beside one sentence is a citation that resolves to source S. What can we say?

We can say the recorded interface displayed a citation object associated with S under the conditions we preserved. That is already useful. We can store the query, answer, anchor, target, time, locale, and surface. But the visible link does not complete the mechanism story.

The first proposition is source presence. Even this needs a set. Was S present in a retrieval candidate list, in the packed context, or only in a visible citation panel? Those are three different memberships. If the only artifact is the final interface, then “present in the visible citation list” is the strongest direct statement.

The second proposition is source use. Use means that S contributed to a declared property of the answer under a justified comparison. A citation is compatible with use, but it is also compatible with a later citation-attachment process, redundant evidence, or a wrong-source link. Use needs a controlled contribution test or another clearly argued method.

The third proposition is absorption. A verified item distinctive to S appears faithfully in the answer. This is stronger than generic overlap. It requires a preregistered fingerprint and an alternative-source check. It still does not automatically identify causal use.

The fourth proposition is displayed attribution. This is the visible citation event. It is the one our opening screenshot directly supports.

The fifth proposition is entailment. Does the resolved passage support the atomic answer claim to which the citation is attached? Entailment is a claim–passage relation, not a property of a URL in isolation.

The sixth proposition is answer correctness. Is the claim defensible in scope, time, and context against the appropriate evidence? A page can entail a stale or false claim. Entailment to one source is therefore not a truth certificate.

Here is the learner check. Complete six sentences, each beginning with “S is…”. Use the phrases visible in the citation list, present in the effective context, causally contributory, source-distinctively absorbed, supportive of claim A, and sufficient for current correctness. Then write the artifact needed for each. You should find that the artifacts differ.

The rule for the rest of W05 is simple: report the edge you observed. Do not let one edge silently stand for the whole graph.

## Chapter 2 — Generation conditions and stage boundaries

Before interpreting an answer, record its condition. The user query is only one component. Conversation history can change references and constraints. Search tools can issue new queries. Retrieval can produce candidates. Reranking and deduplication can change order. Context packing can remove long or redundant passages. System and tool instructions can affect the requested form. Model version and decoding state can alter the output. Citation attachment may occur during or after sentence generation.

In an open local system, we may record most of these components. In a public interface, we may see only a small part. The correct response to an unavailable field is not to imagine it. Write “unobserved.” This separates an honest observational record from a speculative architecture diagram.

A useful condition card has an observation ID, named surface and exposed configuration, exact query and prior state, ordered effective-context passage IDs and hashes if available, raw output and structured citations, time, locale, account or session state, and known exclusions. If the effective context is inaccessible, record that state. If a dynamic alias names a model family without an immutable version, record the caller-visible alias and the absence of a stable backend identity.

Now consider a citation mismatch. The answer says Plan K retains logs for thirty days, but the attached source discusses seven days. Several explanations are possible. The wrong source may have been attached. The correct source may have changed after the response. A citation anchor may resolve incorrectly. The generator may have used a different context. Parametric knowledge may have supplied the claim. The claim may be fabricated. These alternatives occupy different stages.

The reference decomposition retrieval, reranking, packing, generation, attachment, and display helps us ask for the right artifact. It is not a claim that every platform implements those steps in that order. A closed product may combine them. Our diagram is a diagnostic map, not a blueprint of proprietary internals.

Learner check: you observe a correct answer sentence with a wrong visible citation. Write three competing hypotheses. For each, name one record that would discriminate it if you controlled the system. Then rewrite your favorite hypothesis without using the word “because.” The rewrite should sound like “one possible explanation is…” until evidence selects among alternatives.

This condition discipline is also why a later platform dashboard count cannot reveal a whole generation trace. A dashboard may define an event precisely for its surface. It still does not expose every upstream or downstream stage.

## Chapter 3 — Atomic claims and resolved citation objects

We now move from the response level to the claim level. A fluent paragraph can hide several propositions. Imagine the sentence: “Plan K retains records for thirty days except for legacy accounts and is certified under Standard X.” We need at least three assessable units: the general duration, the exception, and the certification statement. The exception remains linked to the duration because dropping it changes meaning.

An atomic claim is not merely a short clause. It is the smallest proposition that can be judged while preserving important qualifiers. Record the exact span and normalized proposition. Preserve quantities, units, entity scope, population, date, condition, comparator, exception, and modality. “May,” “must,” “causes,” and “is associated with” are not stylistic decorations. They change what evidence must support.

Why does segmentation matter numerically? Completeness counts material claims. If one compound sentence becomes one claim, one citation may appear to cover it. If it becomes three claims, we may see that only the duration is supported. Neither segmentation is automatically right. The codebook must say when to split and when to preserve a linked qualification.

Use three tests. Can one predicate be true while another is false? Could different sources support them? Would separating the qualification mislead the reader? These questions usually locate a stable unit. Freeze the spans before comparing systems. If coders change segmentation after seeing citations, the denominator becomes outcome-dependent.

Next resolve the citation. A visible number such as bracket four is an anchor, not an evidence passage. Follow its target URL and redirects. Record a canonical identity, page version or access time, exact passage if available, and provenance origin. A page can change. A redirect can point elsewhere. Three sites can mirror one press release. These facts affect the evidence object.

Keep source and passage separate. The passage may entail a claim, while the source type remains unfit to establish it as true. A marketing page can support “the marketing page says X.” It cannot necessarily support X as an external performance fact. A national standard can support a requirement within its scope but cannot demonstrate that a content edit improved citation behavior.

Learner check: segment the Plan K sentence. For each claim, write the passage fields required for support. Then imagine that the citation resolves to a page that says the organization “aligns with a framework related to Standard X.” Does that entail certification? No. Topic overlap and related terminology are not the certification proposition.

Finally, treat unresolved resolution as data. A blocked or missing page is neither automatically supportive nor automatically wrong. Keep an unresolved state and show its denominator treatment.

## Chapter 4 — Entailment, completeness, and correctness

We are ready for three distinct measures. First, citation entailment. Take the displayed claim–citation edges. For each edge, ask whether the resolved passage supports the attached claim under the codebook. If claims have weights, add supported edge weight and divide by total eligible displayed-edge weight.

Second, completeness. Take the material claims. For each claim, ask whether at least one attached citation adequately supports it. Add the weight of covered claims and divide by total eligible claim weight. Multiple citations on one already covered claim do not cover a different claim.

The denominator shift is the lesson. Suppose an answer has five equal-weight material claims. Only one claim has a citation, and that citation is perfectly supportive. Entailment is one out of one. Completeness is one out of five. Reporting “citation quality equals one hundred percent” without naming the denominator would hide four unsupported claims.

Now reverse the case. All five claims have citations, but two links are topically related and do not support their propositions. Completeness under a support requirement is three fifths, not five fifths. Entailment across five equal edges is also three fifths. Citation count is five, yet count alone hides the two mismatches.

Weights add another decision. A high-impact medical or legal claim may receive greater weight than a minor descriptive sentence. If weights are chosen after seeing errors, they can manipulate the result. Declare materiality and weights before comparison, publish the rule, and show unweighted sensitivity.

Third, answer correctness. A passage may support the words while the underlying source is outdated, conflicted, or outside the relevant jurisdiction. Correctness requires an appropriate evidence set and scope analysis. It may need independent corroboration or an authoritative current record. Source quality should be recorded separately from entailment so the audit can say “the low-quality source entails the report-about-page claim, but does not establish external effectiveness.”

Learner check: in the five-claim, one-citation example, state the unit and denominator for entailment and completeness. Then add a stale source that supports an old value. Which relation remains high? The claim–passage entailment may remain high. Which assessment can fail? Current answer correctness.

Do not silently delete unresolved citation pages. Report displayed edges, resolved edges, supported edges, material claims, covered claims, unresolved items, exclusions, weighting, and qualification policy. A reader should be able to reconstruct every fraction.

## Chapter 5 — Presence, absorption, and source use

Presence is the easiest of these three concepts, but it still requires a named set. We can observe membership in a candidate list, an effective context, or a visible citation list. The claim should say which one. “S was present” without a set is an invitation to conflate stages.

Absorption asks whether content associated with S appears faithfully in the answer. Begin with a preregistered set of verified source claims. Good fingerprints retain unusual quantities, dates, qualifications, relationships, or formulations. Generic terms such as “improves efficiency” are weak identifiers. Search the frozen registry for duplicates and plausible alternatives. Then classify an answer as none, plausible, or source-distinctive for S.

None means no verified distinctive content is present. Plausible means aligned content appears but common knowledge, shared origins, or alternative pages prevent source-specific attribution. Source-distinctive means the operational fingerprint is present and the stated alternative check meaningfully narrows competing origins. Even this strongest label is not proof of training provenance or authorship.

Use is a contribution proposition. In a controlled environment, compare a full context C with C minus S. Hold query, history, instructions, model, decoding policy, context budget, other passages, ordering rule, and repetition plan fixed. Define the outcome in advance. It might be faithful claim coverage, qualification preservation, or answer correctness.

Removing S can produce accidental changes. If later passages shift positions, the treatment includes a position intervention. If the system inserts a replacement, the treatment includes new evidence. If the token budget shrinks, response length may change. A clean teaching design can leave an equal-length spacer where S was, while a production design must document what the actual assembler does.

Interpret null effects carefully. Another source may contain the same fact. The model may already know it. The outcome may be insensitive. Stochastic noise may dominate a small sample. Therefore “no observed difference” is not identical to “S was ignored.” A positive difference is also conditional on the constructed context and chosen utility.

Learner check: write four held-fixed variables for a source-removal comparison, then add two more. If you have fewer than six, the design is probably underspecified. Finally, write a null-result conclusion that lists redundancy and measurement sensitivity as alternatives.

The practical hierarchy is now clear. Citation display is an interface event. Absorption is a verified content relation with alternatives. Use is a controlled contribution claim. Do not promote one into the next without the required evidence.

## Chapter 6 — Four-cell diagnosis and source fit

Imagine a matrix with citation displayed on the rows and appropriate passage support on the columns. The upper-left cell is displayed and supported. This is adequate claim-level attribution under the declared source-fit and correctness checks. The upper-right cell is displayed but unsupported. It includes wrong-source attachments, topical pages, contradictions, and citations that support only part of a compound claim.

The lower-left cell is not displayed, but an auditor finds an appropriate supporting passage. This is an attribution or completeness gap. It does not prove the generator used the discovered passage. The lower-right cell has neither displayed attribution nor identified support. It is an evidence gap requiring correction, qualification, removal, or escalation.

Now add source fit. Suppose a marketing page says, “Our method always ranks first.” A response says, “The marketing page claims that its method always ranks first,” and cites that page. The passage entails the meta-claim. The source is fit to establish what its own page states. But if the response says, “The method always ranks first,” the same page is weak evidence for actual comparative effectiveness. The proposition changed.

Now add provenance. Five mirrored press releases can create five citations while representing one origin. The displayed count is real, but independent corroboration is one origin under the known dependency. A support matrix should therefore retain both URL identity and provenance-cluster identity.

Each failure suggests a different action. A placement error calls for moving or resolving the citation. A partial-support error calls for splitting or qualifying the claim. A low-fit source calls for replacement or narrowing the proposition. Missing attribution calls for a suitable citation if the claim should remain. A false claim calls for correction even if its source is faithfully cited.

Learner check: place “a correct common fact with no visible citation” in the matrix. It is not displayed, and an external audit may find support. That does not show use of any named source. Then place “a false certification claim attached to a page about framework alignment.” It is displayed but unsupported, and source fit is inadequate for certification.

When you write the diagnosis, stop at the narrowest observed stage. “The engine ignored the source” is usually too broad. “The resolved passage does not entail the certification predicate” is precise and actionable.

## Chapter 7 — L04 fixture and reproduced counts

Let us apply the framework to L04. The fixture contains five synthetic responses, twenty-five primary claims, eight resolved source records with snapshots, and a second coder table for ten claims. The graph builder is offline and deterministic. It validates label combinations before producing the graph, summary, unsupported-claim report, disagreement log, relation matrix, and run manifest.

Unsupported claims have no source edge. Contradictions require a contradicted entailment state and false citation-correctness label. Support requires a source plus entailed or partial support. Merely cited requires a visible source edge with no entailment under this fixture’s rule. These invariants make the representation inspectable.

The reproduced relation counts are fourteen support, four contradiction, three merely cited, and four unsupported. Twenty-one claims have resolvable source edges. Four unsupported claims deliberately have none. Eleven claims appear in the review report because they are contradiction, merely cited, or unsupported.

We can calculate two illustrative fractions. Among the twenty-one claim–source edges, fourteen are marked supportive and citation-correct under coder A’s table. That edge fraction is fourteen divided by twenty-one, approximately zero point six six seven. Among all twenty-five claims, fourteen have a supportive edge. That claim coverage fraction is fourteen divided by twenty-five, or zero point five six.

Do not treat either number as the accuracy of a model or platform. The synthetic responses were designed to contain cases. There is no sampling frame representing real answers. The bounded statement is that the numbers summarize coder A’s authored L04 table under the declared rules.

The second-coder subset contains ten claims. The software emits six dimension-level disagreement rows. It prints: “L04 PASS: twenty-five claims, ten double-coded, six label disagreements.” Pass means schema and construction success. It does not mean the annotations are scientifically validated.

Learner check: explain why there are six disagreement rows but only two disputed claim IDs. One claim differs on five dimensions, and another differs on one. A row count is not a count of independent cases. Then state one reproducible software property and one remaining human-validation need.

This fixture is useful precisely because its ceiling is visible. It teaches how a measurement object is constructed while preventing a deterministic command from masquerading as external validity.

## Chapter 8 — C-018 and C-020 disagreement cases

C-018 states that all answer claims are completely cited. Coder A associates visible source S-004 and labels the relation merely cited. Coder B leaves the source empty and labels the claim unsupported. Because several labels depend on whether the edge exists, the coders also differ on citation correctness, source quality, and absorption. Five disagreement rows come from this one claim.

The substantive issue is placement. If S-004 is locally attached to C-018 by the interface or a documented resolver, then “merely cited” records a visible but non-supporting edge. If S-004 occurs only in a global response bibliography and no rule maps it to C-018, unsupported may be the better claim-level representation. The codebook must decide. The adjudicator should not choose whichever label makes completeness look better.

C-020 states that a marketing page says its method always ranks first. Both coders agree on S-006, support, entailment, citation correctness, and low source quality. They disagree only on absorption: source-distinctive versus plausible. The phrase may be distinctive inside the eight-source fixture, yet promotional superlatives are common. The annotation guide needs an explicit alternative-source and registry-scope rule.

These disagreements demonstrate why a single agreement score is insufficient. Across the ten-claim subset, raw agreement is one for entailment, zero point nine for source identity, relation, citation correctness, and source quality, and zero point eight for absorption. Kappa is descriptive and ranges from about zero point seven one for absorption to one for entailment. With only ten selected claims, none of these values establishes population reliability.

Adjudication should preserve both coder files. Create a third decision record with the disputed dimension, both labels, the codebook clause, final or unresolved state, rationale, adjudicator, and date. If the rule changes, identify which other claims need re-review. Recalculate denominators after any edge or segmentation change.

Learner check: write an adjudication rule for C-018 that refers to citation placement, not topical relevance. Then write an alternative-source test for C-020. Finally, explain why agreement on support does not settle source-distinctive absorption. They are different constructs.

The key lesson is not that one coder is right. The pattern tells us where the instrument needs precision. Disagreement is retained evidence about the codebook.

## Chapter 9 — Reading papers and platform documentation by design

W05 routes three frozen papers. `PAPER-24`, titled *Diagnosing and Repairing Citation Failures in Generative Engine Optimization*, motivates a stage-aware failure taxonomy and targeted intervention within its reported setup. We borrow the research question: which stage failed, and what repair corresponds to that diagnosis? We do not import a universal improvement magnitude or assume that a repair transfers across platforms.

`PAPER-33`, *From Citation Selection to Citation Absorption*, motivates separating displayed-source selection from a measure of content influence or absorption. Its reported platform snapshot, fetch-conditioned page set, engineered features, and date define its claims. The portable idea is that link count and answer content are different objects. An engineered overlap or influence score remains a proxy, not hidden model attention.

`PAPER-13`, *Exposing Citation Vulnerabilities in Generative Engines*, adds an adversarial perspective on publisher attributes and citation sources. It encourages us to ask whether a cited origin is trustworthy and how source selection affects poisoning risk. Its empirical proportions remain tied to the topics, countries, systems, and period it studied.

`PLAT-04` is not a paper. It is official Microsoft/Bing documentation for an AI Performance public preview. It can define citations, cited pages, sampled grounding queries, and trends for the named surfaces at the checked date. Those counts do not automatically reveal rank, authority, claim entailment, absorption, or causal use. This difference between official interface documentation and research evidence is itself a learning object.

Learner check: for each of the four identifiers, write one statement the source type can directly support and one statement that would exceed its ceiling. Use “within the reported design” for papers and “for the named surface and checked date” for the platform document.

The reading method is more important than memorizing findings. Identify the measured object, environment, exclusions, denominator, system identities, and uncertainty. Then ask what new design would be required for the claim you actually want.

## Chapter 10 — Bounded conclusion and handoff

We can now construct a defensible W05 conclusion. It should contain six elements: condition, claim unit, source or citation state, measured relation, uncertainty, and boundary.

Here is a complete example. “In the frozen L04 synthetic fixture, coder A segments five responses into twenty-five claims. Fourteen claims carry support edges, four carry contradiction edges, three are merely cited, and four are explicitly unsupported under the authored codebook. Ten claims are double-coded, producing six dimension-level disagreement rows concentrated on placement for C-018 and absorption distinctiveness for C-020. The result describes these frozen annotation files; it does not establish hidden retrieval, causal source use, public-platform absorption, or general annotation reliability.”

Notice what the conclusion does. It says which table produced the counts. It preserves the synthetic boundary. It identifies where disagreement occurs. It separates absorption from use. It does not call the support fraction an engine-accuracy rate.

Your synchronous studio deliverables are a claim–citation graph excerpt, a disagreement memo, a metric card with separate entailment and completeness denominators, and a bounded conclusion. The asynchronous route uses the same assessment. You may run the L04 command or inspect the declared tables if execution is unavailable, but you must still reconstruct the units and evidence ceiling.

Final learner check. Correct this sentence: “Source S shaped the answer because it was cited.” A defensible rewrite is: “Under observation O, the interface displayed a citation resolved to S beside claim A. Whether S contributed to A remains a hypothesis requiring a controlled context comparison or another justified influence design.”

Now add support: “The resolved passage does not entail A under the frozen codebook.” This diagnoses an attribution failure. It still does not reveal whether S was in the effective context. Finally, add correctness only if you have the broader evidence needed to assess A.

W05 ends with a research habit: when a broad word appears, ask for its object. Present where? Used for which outcome? Absorbed according to which fingerprint? Cited at which claim? Entailed by which passage? Correct against which evidence and scope? Once those questions are answered, uncertainty becomes organized rather than embarrassing. It tells us exactly what to inspect, annotate, control, or leave unresolved next.