# W05 Slide Script — Generation, Source Use, Citation, and Absorption Production note: exactly twenty-four slides are specified. All visuals are original instructional diagrams. Purple, gold, grey, and red may support hierarchy, but labels, geometry, line style, and position must carry every distinction without color. ## Slide 01 — One link, six different questions **On-screen text** `source presence · source use · absorption · displayed attribution · entailment · answer correctness` Essential question: *Was a source merely displayed, or did it shape the answer?* **Visual specification** A single citation card enters a six-branch diagnostic fan. Each branch ends in a differently shaped test card: membership box, ablation switch, fingerprint strip, interface frame, claim–passage bracket, and evidence balance. No branch connects directly to another. **Speaker notes** Open with a generated sentence and one visible link. Ask learners what the link proves. Accept only “the recorded interface displayed a resolvable citation object” until more evidence is supplied. Introduce the six questions as separate research objects. A strong answer may require several cards, but no card substitutes for another. State that W05 concerns generation and claim-level attribution, not a hidden-platform ranking formula. **Teaching check** Learners choose the one event directly observable from an ordinary screenshot and name two events that need additional records. **Alt text** “One visible citation branches into six independent diagnostic tests for presence, use, absorption, attribution, entailment, and correctness.” ## Slide 02 — The condition card comes first **On-screen text** Record: surface · version · query/history · effective context · instructions · output/citations · time/locale Unknown is a valid value. **Visual specification** Seven stacked ledger rows form a condition card. Inspectable fields use solid borders; unavailable closed-surface fields use hatched borders and the word `unobserved`. A hash tag hangs from input and output rows. **Speaker notes** An answer depends on the query, conversation, context, system state, and generation policy. In a local pipeline, many fields can be logged. In a public interface, effective context or model version may remain unknown. Recording `unobserved` preserves the evidence ceiling. It is better than reconstructing an imagined prompt from the citation list. The card defines which event another analyst could attempt to reproduce. **Teaching check** Give a screenshot with date and query only. Learners list four missing condition fields and distinguish “missing from record” from “did not occur.” **Alt text** “A seven-row condition ledger separates recorded solid fields from unavailable hatched fields, with unknown preserved explicitly.” ## Slide 03 — A reference pipeline, not a platform blueprint **On-screen text** Retrieve → rerank → pack → generate → attach/resolve → display Each arrow is conditional. Closed implementations may differ. **Visual specification** Six labelled stations run left to right. A dashed outer bracket reads “reference decomposition.” Above every arrow is a small question mark; below are example artifacts: list, scores, context IDs, raw answer, citation graph, screenshot. **Speaker notes** Use the sequence to locate evidence, not to claim that every product follows it. Citation attachment can be joint with generation or a later process. A source may be retrieved but never packed. A packed source may not affect the answer. A correct answer may precede citation attachment. When a mismatch appears, diagnose the narrowest stage supported by artifacts. **Teaching check** Ask where a correct sentence with a wrong link could fail. Accept generation, attachment, resolution, or source-change hypotheses only when phrased as alternatives, not conclusions. **Alt text** “Six conditional stations connect retrieval to visible display, with a distinct inspectable artifact under every stage.” ## Slide 04 — Atomize claims, preserve qualifications **On-screen text** “Plan K retains logs for 30 days, except legacy contracts, and is certified under X.” Claim 1: duration. Claim 2: exception. Claim 3: certification. **Visual specification** The sentence appears as three interlocking strips. The exception strip remains hooked to the duration strip; the certification strip is separate. Span offsets and units sit underneath. Scissors appear only between independently judgeable predicates. **Speaker notes** Atomic does not mean deleting scope. Each claim should retain quantity, unit, population, date, comparator, and modality. Splitting changes the metric denominator; merging can let one supported clause launder another. Explain the truth-divergence and different-source tests. If separating an exception would distort the proposition, retain a linked claim–qualification structure. **Teaching check** Learners segment the sentence, then state which source passage fields must be present to support each unit. **Alt text** “A compound sentence becomes three claims, with the thirty-day rule still physically linked to its legacy-contract exception.” ## Slide 05 — Resolve the citation object **On-screen text** Anchor → target → redirects → canonical source → version/time → passage → origin cluster Resolution failure stays unresolved. **Visual specification** A chain of seven identity tags follows a displayed `[4]` anchor. One branch ends at a broken-link tray labelled `unresolved`; another ends in an origin cluster containing three mirrored URLs around one source node. **Speaker notes** A numbered anchor is not an evidence passage. URL redirects, canonicalization, changing pages, and syndicated copies alter what “the source” means. Preserve access time or content hash. Cluster derivative pages when provenance is knowable. Do not count five mirrors as five independent confirmations. An unavailable page is not automatically a wrong citation; it is a resolution state requiring policy. **Teaching check** Which record distinguishes three URLs from three independent origins? Learners must answer with provenance or dependency evidence, not domain count. **Alt text** “A displayed anchor resolves through URL and passage identities, while a broken path stays unresolved and mirrored URLs share one origin.” ## Slide 06 — Build the claim–citation graph **On-screen text** Answer → contains → claim Claim → support / partial / contradict / merely cited → passage Passage → belongs to → source origin **Visual specification** Three horizontal bands contain answer nodes, atomic claim nodes, and source-passage nodes. Solid arrows mean support, half-dashed arrows partial support, barred arrows contradiction, and dotted arrows merely cited. A legend repeats each label in text. **Speaker notes** The graph prevents a global bibliography from appearing to support an entire answer. Every material claim needs an explicit state, including unsupported. Passage and source identity remain separate. In L04, unsupported claims deliberately have no source edge; a `merely_cited` edge records visible placement without support. The graph can preserve disagreement rather than forcing one confident label. **Teaching check** Ask learners to draw the edge for a visible page that discusses the topic but does not entail the attached claim. **Alt text** “Claims sit between answers and source passages, with labelled line styles for support, partial support, contradiction, and mere citation.” ## Slide 07 — Entailment asks about attached edges **On-screen text** `Entailment = supported weight on displayed claim–citation edges / total displayed-edge weight` Unit: claim–citation edge. **Visual specification** A denominator tray holds four edge tokens with weights 2, 1, 1, and 1. Three pass through a support gate; one enters a mismatch tray. A side card lists segmentation, passage, codebook, and adjudicator. **Speaker notes** Entailment evaluates whether a passage supports the claim attached to it. It does not tell us how many claims lack citations. State how partial support is scored and how unresolved pages enter the denominator. If an attached citation supports only one half of a compound sentence, segmentation or qualification handling determines the result. Report both counts and the formula. **Teaching check** Four equal-weight displayed edges contain three supported and one unsupported relation. Learners calculate 0.75 and name the missing completeness information. **Alt text** “Four weighted displayed edges enter an entailment denominator; three pass support and one fails, with protocol fields shown separately.” ## Slide 08 — Completeness asks about material claims **On-screen text** `Completeness = material-claim weight with ≥1 supporting citation / total material-claim weight` Unit: claim. **Visual specification** Five claim tiles form the denominator. Two have supported citation hooks, one has a wrong hook, and two have no hook. A materiality weight appears inside each tile. The previous slide’s edge tray is shown faintly and separated. **Speaker notes** Completeness changes the denominator from edges to claims. An answer can have perfect entailment for its only citation and poor completeness when four other material claims remain unsupported. Define which claims require evidence and how general conversational content is treated. Do not let multiple citations on one claim inflate coverage. Preserve qualifications as material where omission changes meaning. **Teaching check** If two of five equal-weight material claims have at least one supporting citation, what is completeness? Why can adding a second citation to one covered claim leave it unchanged? **Alt text** “Five claims form a completeness denominator; only two have adequate support, while extra edges cannot create coverage for uncovered claims.” ## Slide 09 — Correct citation does not guarantee correct answer **On-screen text** Passage says claim ≠ claim is true now and in scope. Check: source fit · currency · applicability · conflicting evidence. **Visual specification** A claim and passage align inside one bracket labelled `entailed`. Beside it, a calendar shows expiry, a scope stencil excludes the target population, and an evidence balance shows a contradiction. A second bracket labelled `answer correctness` spans all four objects. **Speaker notes** Entailment is relational: the passage says or supports the proposition. Answer correctness asks whether the proposition should be accepted against the relevant evidence. A stale official page can entail an obsolete value. A marketing page can entail “the page claims X” without proving X. Record source quality separately so an accurate support label does not become an authority or truth score. **Teaching check** Learners create one example of high entailment with low current correctness and name the evidence that exposes the difference. **Alt text** “Claim and passage match, but expiry, scope mismatch, and conflicting evidence keep answer correctness as a wider independent test.” ## Slide 10 — The four-cell diagnostic matrix **On-screen text** Displayed + supported: adequate attribution Displayed + unsupported: wrong or weak attachment Not displayed + support found: attribution gap Not displayed + unsupported: evidence gap **Visual specification** A two-by-two matrix uses column headers “appropriate passage supports / does not support” and row headers “citation displayed / not displayed.” Each cell contains a distinct icon and an action verb: accept, repair, attribute, correct. **Speaker notes** Explain that the lower-left cell means an auditor found support; it does not establish that the system used that source. Add source quality and answer correctness as annotations rather than collapsing the matrix. The matrix is useful because each cell suggests a different editorial or research action. A citation-count metric cannot distinguish the upper cells. **Teaching check** Place “correct common fact with no visible citation” in the matrix, then state why source use remains unknown. **Alt text** “A labelled two-by-two matrix separates supported attribution, wrong attachment, missing attribution, and missing evidence, each with a different action.” ## Slide 11 — Presence is a membership event **On-screen text** Presence may mean: candidate list · packed context · visible source list. Always name the set. **Visual specification** Three non-overlapping containers carry the same source card S: a candidate basket, a context window, and a browser citation panel. Membership checkmarks are scoped to each container; no arrows imply movement between them. **Speaker notes** The word “present” is incomplete unless its set is named. A source visible in the interface is not proven to have occupied the generator’s context. A source in an open context is not necessarily cited. In a closed system, only the visible list may be observable. The correct statement uses the set identity, time, and system condition. **Teaching check** Rewrite “S was present” three ways, one for each container, with a record that could verify the statement. **Alt text** “The same source occupies three separately named containers—candidate, context, and citation panel—without implied transfer between them.” ## Slide 12 — Absorption needs a preregistered fingerprint **On-screen text** `none · plausible · source-distinctive` Verify content. Preserve qualifications. Check common knowledge, duplicates, and alternatives. **Visual specification** A source fingerprint contains a number, date, uncommon qualification, and structural relation. It is compared with an answer strip. Three gates test fidelity, alternative sources, and duplicated origin before the strongest label. **Speaker notes** Generic word overlap is not absorption. Define verified source-distinctive claims before observing outputs. A shared public fact may be faithfully present but only plausibly connected to S. A distinctive figure plus its exception may justify the strongest operational label under the registry, while still not proving causal use or training provenance. **Teaching check** Learners explain why a shared definition appearing on twenty sites cannot identify one source, even if one of those sites is cited. **Alt text** “A source fingerprint passes fidelity, alternative-source, and duplicate-origin gates before receiving a source-distinctive absorption label.” ## Slide 13 — Citation and absorption can diverge **On-screen text** Cited, not absorbed. Absorbed, not cited. Cited and absorbed. Neither observed. **Visual specification** Two intersecting circles are explicitly labelled “displayed attribution” and “verified distinctive content.” Four example cards occupy the two exclusive areas, intersection, and outside region. A dashed boundary says `use still untested` around the whole diagram. **Speaker notes** Use this diagram to break the link-count habit. A citation can be attached to a claim that does not draw on the page. Distinctive content can appear with no displayed attribution. Co-occurrence of citation and absorption is stronger evidence of alignment but is not automatically a controlled contribution estimate. Interface post-processing and redundant origins remain alternatives. **Teaching check** Learners supply a diagnostic test for each exclusive region and identify which test requires access to source passages. **Alt text** “Citation display and distinctive-content absorption overlap only partly, and a larger dashed boundary states that causal use remains untested.” ## Slide 14 — Test use with a controlled removal **On-screen text** Compare `C` with `C − S`. Hold: query · instructions · model · budget · other passages · order policy · repetitions. **Visual specification** Two parallel context trays share identical slots. In the second, S is replaced by a labelled empty spacer rather than shifting every later card. Paired output cards feed a difference ledger with uncertainty bars. **Speaker notes** Define source use as contribution to a declared answer outcome under a constructed comparison. Removing S may otherwise change position, token budget, or replacement content. Declare whether empty space remains and how ordering is preserved. With stochastic decoding, use paired repetitions. A difference estimates conditional contribution in this context, not global importance, authorship, or a commercial mechanism. **Teaching check** What confound appears if all later passages shift upward after removal? Learners answer “position and context composition changed with the treatment.” **Alt text** “Full and source-removed contexts remain slot-aligned, then paired answer differences are summarized with uncertainty rather than a single deterministic arrow.” ## Slide 15 — Redundancy complicates a null result **On-screen text** No removal effect may mean: no use · redundant evidence · weak outcome · noisy decoding. Null is not a mechanism verdict. **Visual specification** Three sources feed the same verified claim through parallel pipes. Removing S leaves two pipes. A separate noisy meter and insensitive outcome gauge appear beside the pipeline. Four labelled explanations connect to a zero-difference card with dotted lines. **Speaker notes** When multiple sources provide the same information, removing one may leave the answer unchanged. The result is compatible with non-use and redundancy. An outcome that ignores qualifications can also miss a meaningful effect. Repeated sampling may be too small. State these alternatives and design follow-up source-cluster or claim-level ablations when ethical and feasible. **Teaching check** Ask why a near-zero leave-one-source-out contrast cannot prove that S was ignored. **Alt text** “Three redundant evidence pipes survive removal of one source, while noisy and insensitive measurements provide additional explanations for no observed difference.” ## Slide 16 — Diagnose the narrowest failure **On-screen text** segment → resolve → place → entail → qualify → fit → complete → absorb → test use Stop where evidence stops. **Visual specification** A vertical diagnostic ladder contains nine labelled rungs. A movable stop flag sits at `place`; higher rungs are hatched and marked “not yet tested.” Each rung has one required artifact beside it. **Speaker notes** A broad statement such as “the engine ignored the page” often skips several untested stages. Begin at the claim and move only when the necessary artifact exists. A misplaced citation requires an attachment diagnosis. An unresolved page blocks passage entailment. An absorption label without an alternative-source check stays plausible. The narrow failure points to a specific repair or next study. **Teaching check** Given a broken redirect, where should diagnosis stop? At resolution; support and use remain unassessed. **Alt text** “A nine-rung diagnostic ladder stops at citation placement, leaving unsupported higher-stage inferences visibly hatched.” ## Slide 17 — L04 is a deterministic teaching fixture **On-screen text** 5 synthetic responses · 25 claims · 8 sources 10 double-coded claims · offline · standard-library graph builder **Visual specification** Five response folders feed twenty-five claim slips and eight source cards into a transparent graph-builder box. Ten claim slips carry paired coder tabs. A closed network icon and synthetic stamp define the boundary. **Speaker notes** L04 provides inspectable artifacts without accounts or live model calls. Its software path is deterministic. Its annotation labels are authored teaching judgments, not ground truth validated across a population. The fixture supports learning segmentation, invariants, agreement, and bounded conclusions. It does not estimate performance of an external platform. **Teaching check** Name one property made reproducible by the code and one scientific property that still needs human validation. **Alt text** “Five synthetic responses, twenty-five claims, eight sources, and ten paired annotations enter an offline transparent graph builder.” ## Slide 18 — Read the L04 relation matrix **On-screen text** Support 14 · Contradiction 4 · Merely cited 3 · Unsupported 4 These are authored fixture labels, not a population estimate. **Visual specification** Four horizontal bars use distinct fill patterns and print the counts at both ends. Beneath them, a denominator plaque reads `25 frozen claims`. No percentages or confidence intervals appear because no sampling population is asserted. **Speaker notes** Interpret each count as a property of coder A’s frozen table under the codebook. Twenty-one claims have a resolvable source edge; four unsupported claims deliberately have none. Eleven claims require review across contradiction, merely cited, and unsupported states. Avoid translating the distribution into “system accuracy.” The responses were authored to exercise failure modes. **Teaching check** Why is fourteen divided by twenty-five not a production support rate? Learners name synthetic construction and absence of a sampling frame. **Alt text** “Four patterned bars report fourteen support, four contradiction, three merely cited, and four unsupported labels out of twenty-five frozen claims.” ## Slide 19 — Disagreement locates codebook pressure **On-screen text** 6 disagreement rows. C-018: five dimensions. C-020: absorption only. Preserve both coders before adjudication. **Visual specification** A disagreement heatmap has claims as rows and six label dimensions as columns. Five marked cells line up on C-018; one marked absorption cell appears on C-020. An immutable two-layer archive sits beneath an adjudication overlay. **Speaker notes** The ten-claim subset has six dimension-level disagreement records, not six independent claims. C-018 differs on whether an S-004 edge exists and on the dependent relation labels. C-020 agrees on support and source quality but differs between source-distinctive and plausible absorption. This pattern suggests where the codebook needs examples. Do not erase original coder states. **Teaching check** Which disagreement should trigger an alternative-source rule example? C-020’s absorption distinction. **Alt text** “A heatmap concentrates five disagreements on C-018 and one absorption disagreement on C-020 above a preserved pre-adjudication archive.” ## Slide 20 — Agreement is not validity **On-screen text** Report: subset `n` · class prevalence · raw agreement · dimension-specific kappa · severe cases. High agreement can reproduce a bad rule. **Visual specification** Two gauges sit side by side: reproducibility of labels and validity for the intended construct. A bridge labelled “codebook + representative sample + expert review” connects them but does not merge them. A small `n=10` tag limits the first gauge. **Speaker notes** Agreement measures consistency under an instrument; validity asks whether the instrument captures the intended relation. Both need evidence. Kappa is sensitive to prevalence and unstable in tiny samples. L04 values are descriptive on ten claims. Automated judges similarly require human calibration across relevant strata, not a single unexamined agreement number. **Teaching check** Give an example of perfect agreement on an invalid rule, such as labeling every visible citation supported. **Alt text** “Separate gauges represent label reproducibility and construct validity, joined by review evidence but constrained by a ten-item sample.” ## Slide 21 — Read PAPER-24, PAPER-33, and PAPER-13 by design **On-screen text** `PAPER-24`: stage diagnosis and targeted repair `PAPER-33`: selection versus absorption measurement `PAPER-13`: citation-source vulnerability and publisher attributes Portable: questions and design patterns. Non-portable: universal effects. **Visual specification** Three paper cards each contain columns for measured object, environment, reported result, and boundary. Their boundary cells extend downward into one shared “do not generalize without validation” rail. **Speaker notes** Use the public catalog identifiers, not internal Zotero keys. These papers motivate distinct aspects of W05. Their numerical findings remain tied to reported datasets, models, exclusions, and dates. The course borrows no figures. Learners should be able to state what each design directly measures and what additional evidence a cross-platform claim would require. **Teaching check** Assign each sentence—citation failure stage, engineered absorption proxy, publisher-risk attribute—to the relevant paper and add one boundary. **Alt text** “Three paper cards map diagnostic, absorption, and vulnerability designs to a shared boundary against universal transfer.” ## Slide 22 — PLAT-04 defines an interface, not an influence mechanism **On-screen text** `PLAT-04`: official Bing AI Performance public-preview documentation. Counts for named surfaces ≠ rank · authority · entailment · absorption · source use. **Visual specification** A platform dashboard frame exposes four labelled counters: citations, cited pages, sampled grounding queries, and trends. Five locked drawers below carry the excluded concepts. A date and surface badge is attached to the frame. **Speaker notes** Official documentation is appropriate evidence for what the named preview says its metrics represent at the checked date. It is not a peer-reviewed validation of every inferred construct. A citation count does not become a rank or authority score. This is a useful measurement-interface critique: begin with the platform’s definition, then add independent claim-level auditing where the data permit. **Teaching check** Rewrite “the dashboard proves our source influenced answers” as a bounded platform observation. **Alt text** “A dated dashboard exposes four named counts while rank, authority, entailment, absorption, and causal use remain in locked separate drawers.” ## Slide 23 — Two equivalent learning routes **On-screen text** Synchronous: planned 90-minute seminar + 75-minute studio. Asynchronous: notes + no-video transcript + frozen L04 route. Same assessed artifacts. No timing claim yet. **Visual specification** Two parallel paths begin at W05 concepts and converge on three identical outputs: claim–citation graph, disagreement memo, and bounded conclusion. A crossed-out camera indicates that no recording exists. Clock icons carry the label `authoring budget`. **Speaker notes** The two routes target equivalent outcomes, not identical elapsed time. The studio uses local synthetic material and no account. The transcript contains all assessed spoken content. State explicitly that the lesson has not been delivered, timed, or piloted. Accessibility provisions are design commitments awaiting human testing, not a conformance claim. **Teaching check** Learners name the three artifacts that must remain identical in assessment criteria across both routes. **Alt text** “Synchronous and asynchronous paths converge on the same graph, disagreement memo, and bounded conclusion; no recorded-media claim is made.” ## Slide 24 — Exit: report the edge, not the myth **On-screen text** Observation: “S was displayed beside claim A.” Diagnosis: “Passage P did not entail A.” Unknown: “causal source use.” Next test: resolve, adjudicate, or ablate under control. **Visual specification** A four-line release card contains observation, diagnosis, unknown, and next test. Behind it, the six opening test shapes return, now each holding a named artifact. A final boundary bar says “no hidden-platform mechanism inferred.” **Speaker notes** Close by converting broad language into an evidence chain. A good W05 conclusion names condition, claim, source passage, citation state, judgment, uncertainty, and boundary. It does not promise that one editing tactic controls a platform. Ask every learner to submit two sentences: one supported observation and one strongest non-established claim. The next research action should follow from the narrow diagnosis. **Teaching check** Final prompt: correct “the source shaped the answer because it was cited.” Require one observation and one controlled hypothesis. **Alt text** “A four-line card separates observation, diagnosis, unknown causal use, and next test, returning to the six independently evidenced questions.”