# W09 Slide Script — Evidence-Rich Content Architecture ## Production conventions This script defines exactly 24 slides. Use a restrained academic palette: deep purple for governed objects, gold for declared decisions, charcoal for frozen artifacts, blue-gray for pending review, and red hatching for a failed gate or prohibited transfer. Labels, patterns, shape, and position carry every distinction that color reinforces. All diagrams are newly designed for W09. Do not import, trace, or restyle figures, screenshots, page layouts, or TeX artwork from papers, platforms, websites, or other courses. Each data visual must have a text table. Each diagram needs the supplied alt text and a matching long-description route in production. ## Slide 01 — Structure can improve use without changing facts **On-screen text** > W09 · Evidence-Rich Content Architecture > How can structure improve use without distorting facts? > Evidence interface ≠ ranking recipe **Visual specification** Place one frozen fact ledger at the top. Two arrows lead to a control page and a treatment page with visibly different layout but identical claim/source seals. A large bracket encloses both pages and reads “same factual message.” Below, separate four possible evaluations: deterministic equivalence, human comprehension, open-pipeline behavior, and live observation. Only the first is filled. **Speaker notes** Open with the distinction that drives W09. Structure is valuable when it helps a person or declared system recover the right claim, condition, method, date, limitation, and source. That value does not require a ranking effect. When we test structure, we must protect the factual message and name the stage being studied. L05 verifies local equivalence and one changed factor. It does not measure comprehension, retrieval, citation, or production behavior. The four evaluation boxes prevent a software check from being narrated as a field result. **Teaching check** Ask which box L05 fills. Expected answer: deterministic local structural and factual-equivalence evidence only. **Alt text** A frozen fact ledger produces differently arranged control and treatment pages with identical claim seals; four evaluation levels show only local equivalence as completed. ## Slide 02 — Outcomes and release artifacts **On-screen text** - Reconstruct evidence-page anatomy. - Lock atomic claims and sources. - Align visible content and schema. - Isolate one factor. - Gate accessibility and validity. - Reproduce diff and rollback. **Visual specification** Create a six-segment ring around a central release package. Each segment names one outcome. Around the outer rim, list claim lock, control/treatment, intervention card, equivalence review, complete diff, audit report, deviation log, and rollback manifest. A lock icon marks the transition from design to build. **Speaker notes** The unit's output is an auditable intervention package, not a redesigned screenshot. The claim lock states what cannot change. Control and treatment preserve the two arms. The intervention card names the permitted difference. The equivalence review records independent checks. The complete diff exposes every changed line. The audit report checks the contract. The deviation log prevents silent repairs. The rollback manifest identifies how to restore the control and verify it. Together they support a bounded conclusion about what was changed and preserved. **Teaching check** Ask which artifact tells a reviewer why the factor was expected to matter. Expected: the intervention card, including hypothesized stage, expected signal, alternatives, and falsification condition. **Alt text** Six learning outcomes surround a versioned release package whose outer artifacts preserve design, comparison, deviations, and rollback. ## Slide 03 — Four evidence claims must remain separate **On-screen text** 1. Easier for a person to verify 2. More consistent for a declared parser 3. Stage effect in an open pipeline 4. Time-bounded live-surface outcome Evidence does not automatically climb the ladder. **Visual specification** Draw four platforms rising from left to right, but replace the staircase arrows with gated bridges. Each bridge requires a new study label: user study, parser validation, pipeline experiment, or live design. Place L05 on platform two with an arrow backward to platform one labeled “hypothesis only.” **Speaker notes** Structural work often overclaims by sliding between these objects. A parser test may show that claim IDs and sources are recoverable. It does not show that readers understand them better. A reconstructed retrieval result does not describe a commercial system. A live display change may be time-bounded and mechanism-opaque. Each bridge requires new data, assignment, outcomes, and validity assumptions. W09 keeps the ladder visible so a local PASS cannot be promoted into ranking language. **Teaching check** Ask what new evidence is needed to move from local equivalence to improved comprehension. Expected: an approved human task study with participants, allocation, outcomes, accessibility, uncertainty, and ethics. **Alt text** Four evidence platforms are separated by gated study bridges, showing that parser, human, open-pipeline, and live outcomes require distinct evidence. ## Slide 04 — A page is a view of a governed fact system **On-screen text** Claim record: ID · entity · proposition · value/unit · scope · date · evidence · source · owner · status · conflict · correction The page is not the only ledger. **Visual specification** Show a central claim ledger with twelve labeled fields. From it, generate four views: visible page, structured data, downloadable record, and correction queue. Cross-view arrows point back to one source-of-truth version. A broken red arrow shows the invalid pattern of editing schema separately. **Speaker notes** Organizations frequently maintain facts in scattered decks, pages, and campaign copy. W09 places a governed claim system upstream. Every public representation should resolve to reviewed claims and evidence. This design reduces stale prices, expired certifications, unqualified outcomes, and contradictory metadata. It does not require publishing the internal ledger. It requires traceable generation and correction. The version field lets a reviewer distinguish a structural treatment from an underlying claim change. **Teaching check** Ask why the structured-data file should not be edited as a separate promotional surface. Expected: it can create stronger or conflicting claims that no longer match the governed visible content. **Alt text** A twelve-field claim ledger generates four public and governance views, all linked to one version; a separate schema edit is marked invalid. ## Slide 05 — Atomic claims preserve challengeable propositions **On-screen text** Compound statement → claim IDs - capability and version - certification and validity - outcome and denominator - comparator and method - population, locale, and date Atomic does not mean context-free. **Visual specification** Display one fictional compound sentence as a long capsule. Slice it into five numbered claim cards, but connect all material conditions with braces so they cannot be dropped. Each card links to a source/evidence node. Use a red empty placeholder where evidence is unresolved. **Speaker notes** Atomicity makes each proposition independently reviewable. It prevents one strong source from laundering neighboring claims. But naive splitting can strip conditions. If “among completed responses” belongs to a percentage, it travels with that claim. If a certification applies only to one product version and date, those fields remain local. A claim lock can use exact approved text for a layout-only intervention or normalized propositions plus a rigorous equivalence protocol for expression-preserving edits. **Teaching check** Ask whether a limitation should become a separate footnote when it changes the denominator. Expected: no; the material condition must stay attached to the claim, though a detailed explanation may also appear elsewhere. **Alt text** A compound sentence is divided into five evidence-linked claims while braces keep version, method, denominator, population, locale, and date attached. ## Slide 06 — Evidence-page anatomy by verification function **On-screen text** Identity/version · answer/scope · claim blocks · evidence table · definitions · method · limitations · sources · change history · correction route Anatomy, not outcome guarantee. **Visual specification** Create an original single-column page wireframe with ten labeled zones. Beside it, a “verification rail” asks ten questions: what object, what claim, under what scope, with which evidence, by what method, when, where it fails, from which origin, what changed, and who corrects it. Connect each zone to one question. Avoid product-card or search-result styling. **Speaker notes** The component list is functional, not visually prescriptive. A specification and a research report can instantiate the functions differently. A reviewer should be able to recover identity, scoped answer, atomic claims, comparable evidence, terms, method, limitations, source identity, revision history, and a correction route. The anatomy helps human verification and may support segmentation or parsing. It does not establish which design a hidden platform prefers. **Teaching check** Ask which component prevents a result from being treated as timeless. Expected: dates in the claim/evidence record plus version header and change history. **Alt text** A ten-zone evidence page connects each content region to a specific verification question and carries a warning that anatomy is not an outcome promise. ## Slide 07 — Co-located qualification survives segmentation **On-screen text** A value travels with: unit · population · comparator · method · date · uncertainty · limitation · source Do not separate the headline from its conditions. **Visual specification** Draw one evidence block as a train. The numerical claim is the engine and eight labeled condition cars follow it. A segmentation blade cuts before and after the entire train, not between cars. A crossed-out inset shows a headline card separated from a distant limitation footer. **Speaker notes** Pages can be chunked, excerpted, skimmed, or quoted. Co-location makes a semantic unit less likely to lose its qualification. It is not keyword repetition. It is scientific writing discipline. A comparison table should not align cells with incompatible populations or dates. A limitation that changes interpretation belongs beside the claim. A source link should resolve the exact origin, not merely a publisher homepage. **Teaching check** Ask which item must stay with “40% faster.” Expected: at least outcome definition, comparator, population, method, period, and uncertainty or limitation. **Alt text** A numerical claim and eight condition cars stay together when segmented; a detached headline and distant limitation are shown as invalid. ## Slide 08 — Date types and change history are evidence **On-screen text** Publication · observation window · valid-from/to · last reviewed · next review Change types: typo · evidence refresh · claim change · method change · structural intervention · withdrawal **Visual specification** Use two horizontal tracks. The first has five differently shaped date markers. The second is a version timeline with six change-type badges and hashes. A connector links every public claim version to the evidence and page state valid at that time. **Speaker notes** “Updated recently” is not enough for a scientific evidence page. Different dates answer different questions. A source can be published today but report an older observation. A certification may expire. A review date says when someone checked, not when the underlying fact was true. The change log distinguishes editorial correction from substantive mutation. During an intervention, structural changes receive their own version so they cannot hide claim changes. **Teaching check** Ask whether fixing a percentage denominator is a typo. Expected: normally no; it is a substantive claim or evidence correction requiring review and affected-artifact propagation. **Alt text** Five date types align with a versioned change-history track so each claim resolves to the evidence and page state valid at that time. ## Slide 09 — Visible content and schema form one claim surface **On-screen text** Consistency states: equivalent · absent by design · conflicting · stale · unreviewed Visible HTML · metadata · schema · links · captions/alt · downloads **Visual specification** Construct a matrix with atomic claim IDs as rows and six representations as columns. Use symbols plus words for the five states. One red cell shows a schema rating stronger than the visible rating. A gate below blocks release until conflict and stale cells are resolved. **Speaker notes** Not every claim belongs in every representation, so absence can be deliberate. Conflict is the problem. A machine-readable rating, region, price, award, or capability cannot exceed the visible and evidenced claim. Schema syntax validation and factual consistency are separate checks. Alt text and captions must also avoid stronger claims or promotional stuffing. The matrix makes hidden claim drift reviewable. **Teaching check** Ask how to mark a method limitation that has no schema property. Expected: absent by design, provided the visible evidence block retains it and no encoded claim contradicts it. **Alt text** A claim-by-representation matrix marks equivalent, intentionally absent, conflicting, stale, and unreviewed states; one stronger schema claim blocks release. ## Slide 10 — PLAT documentation has bounded authority **On-screen text** - **PLAT-02:** Google publishing guidance; not a scientific score - **PLAT-03:** valid markup may create eligibility; display not guaranteed - **PLAT-13:** shared vocabulary semantics; no visibility effect Documentation scope ≠ causal evidence **Visual specification** Create three document cards with a “can support” line and a “cannot establish” line. PLAT-02 has a policy border, PLAT-03 an eligibility gate, and PLAT-13 a vocabulary graph. All three stop before a separate outcome box labeled retrieval/ranking/citation. **Speaker notes** Official documentation is authoritative for the behavior or policy it actually declares, within version and surface. PLAT-02 informs responsible-publishing checks. PLAT-03 supports the eligibility-versus-display distinction. PLAT-13 defines vocabulary terms. None is an experiment showing that adding schema or following a formatting checklist causes visibility. Keep current documentation identities and access dates in the source record. **Teaching check** Ask what valid structured data guarantees under PLAT-03. Expected: no guaranteed display or ranking; at most eligibility for supported features under the documented conditions. **Alt text** Three bounded documentation cards stop at policy, eligibility, and vocabulary functions before an unmeasured outcome box. ## Slide 11 — Factual equivalence is a field-by-field contract **On-screen text** Lock: claim · number/unit · denominator · scope · entity · date · qualification · evidence · source · attribution · disclosure Similarity is not equivalence. **Visual specification** Place control and treatment claim records side by side. Eleven horizontal connectors compare locked fields. Ten are solid equal signs; one hypothetical missing qualification becomes a red broken connector. A decision gate reads “any material break → fail.” **Speaker notes** Layout can change when layout is the factor, but factual meaning cannot. Exact text equality is appropriate for L05. Expression-preserving edits require a richer comparison of propositions and conditions. A page can “feel the same” while changing a denominator, time range, source, or disclosure. One material break invalidates the single-factor interpretation. Independent review helps surface these changes but does not remove the need for a clear contract. **Teaching check** Ask whether moving a limitation into a collapsed element preserves equivalence. Expected: not automatically; visibility, accessibility, local recoverability, and reading order may materially change. **Alt text** Eleven field-level connectors compare control and treatment; a missing qualification breaks equivalence and fails the gate. ## Slide 12 — Semantic similarity cannot certify truth conditions **On-screen text** High overlap can hide: - removed denominator - changed modal verb - newer date - broader locale - different source - missing exception **Visual specification** Show two nearly identical text strips with a large “0.97 similarity” badge. Six small magnifiers reveal one changed material token or field in each example. A separate review seal says “proposition + scope + evidence,” not “embedding.” **Speaker notes** Text metrics can help triage a large diff. They do not validate facts. “May improve” and “improves” can have high similarity and different epistemic force. “Among returned responses” can disappear with a small edit and change the denominator. A source can be swapped while visible words stay identical. The equivalence review operates on governed claims, not only tokens or embeddings. PAPER-28 similarity values remain audit-only and cannot substitute for this review. **Teaching check** Ask what the similarity badge can legitimately do. Expected: flag or prioritize review; it cannot certify factual, attribution, or scope equivalence. **Alt text** Two highly similar text strips contain six material field changes revealed by magnifiers, showing why similarity cannot certify truth conditions. ## Slide 13 — The intervention card freezes the scientific object **On-screen text** ID/version · authorization · environment · factor · stage · expected signal · falsification · non-claims · confounds · gates · stop · rollback **Visual specification** Design a twelve-field intervention card as a laboratory specimen label, not a UI dashboard. Group fields into identity, mechanism, boundaries, and release. Place a checksum strip along the bottom. One blank field makes the card visibly unreleasable. **Speaker notes** “Make it clearer” is not a treatment. The card names exactly what changes and where the mechanism is hypothesized to operate. It states what observation would weaken the idea and which outcomes are not measured. Authorization and environment prevent live experimentation from being inferred. Confounds tell the reviewer what else could explain a result. Gates and stop conditions make integrity non-compensable. Rollback turns reversibility into an acceptance test. **Teaching check** Ask which field stops a local layout test from being described as a citation study. Expected: hypothesized stage plus explicit non-claims and claim ceiling. **Alt text** A twelve-field specimen-style intervention card groups identity, mechanism, boundaries, and release controls and cannot pass with a blank field. ## Slide 14 — A mechanism hypothesis needs alternatives **On-screen text** Factor: evidence-layout proximity Possible pathways: - easier claim-to-source matching - border attracts attention - grid changes wrapping - reader familiarity - assistive-technology difference **Visual specification** Place the factor node at left and five branching mechanism paths at right. Only the first is the target hypothesis; the others are dashed alternatives or adverse routes. Beside each path, list one discriminating measurement. Include a viewport and screen-reader branch. **Speaker notes** If treatment helps a comprehension task, spatial proximity is one explanation, not the only one. Decorative salience, wrapping, familiarity, or assistive-technology interaction may drive the result. A later study should measure or control these routes. The L05 card records confounds but performs no human outcome collection. Mechanistic humility prevents one design preference from becoming doctrine. **Teaching check** Ask for a measurement that distinguishes proximity from border salience. Expected: a condition that changes grouping/proximity while holding border treatment constant, or a factorial design. **Alt text** One layout factor branches into a target comprehension pathway and four alternative or adverse explanations, each paired with a discriminating measure. ## Slide 15 — One-factor design rejects accidental polish **On-screen text** Declared difference: `.evidence-block` layout rule Locked: HTML except stylesheet name · claim text · source IDs · title · language · DOM order · other CSS **Visual specification** Create two transparent overlays of control and treatment files. Locked regions align in charcoal. The stylesheet filename and one CSS rule glow gold. A red stray line labeled “polish” falls outside the permitted mask and triggers fail. **Speaker notes** The treatment contract defines a mask of permissible change. Anything outside it is a co-intervention, even if aesthetically reasonable. In L05, HTML matches after normalizing the stylesheet reference, and CSS matches after normalizing the single evidence-block rule. Claim text and source identities are independently parsed against the lock. This is stronger than comparing rendered screenshots, which can omit metadata and off-screen behavior. **Teaching check** Ask whether changing link wording to sound clearer is allowed in the L05 factor. Expected: no; it changes text and potentially link purpose, so it needs another intervention or bundle definition. **Alt text** Control and treatment overlays match everywhere except a stylesheet reference and one permitted CSS rule; an extra polish line fails the mask. ## Slide 16 — The local HTML comparison is deliberately narrow **On-screen text** Control and treatment share: 3 claims · 3 sources · one `main` · one `h1` · language · title · DOM order Only local static HTML/CSS. No external endpoint. **Visual specification** Draw two document trees side by side. Identical nodes connect horizontally. The only divergent leaf is the stylesheet link, which leads to two CSS trees differing at one rule. Place a sealed perimeter around the diagram labeled “local fixture.” **Speaker notes** The parser extracts `lang`, title, `main`, `h1`, stylesheet, claim IDs/text, and source IDs. It verifies one stylesheet per page, unique sources, identical claim lock, and equal title/language. The HTML files differ only by the stylesheet reference. The CSS files differ only in one evidence-block rule. The sealed local perimeter matters: no crawler, account, production traffic, or remote dependency is involved. **Teaching check** Ask what the side-by-side tree cannot reveal. Expected: actual screen-reader usability, visual contrast in all states, human comprehension, or any external system response. **Alt text** Two identical document trees diverge only at the stylesheet link and one CSS rule inside a sealed local-fixture boundary. ## Slide 17 — Deterministic L05 audit: PASS with a narrow ceiling **On-screen text** | Locked claims | Source IDs | Review checks | Changed factors | Errors | |---:|---:|---:|---:|---:| | 3 | 3 | 6 | 1 | 0 | Factor: `evidence_layout_proximity` Stage hypothesis: `human_comprehension` **Visual specification** Use a five-column audit strip with exact counts. Below, a claim-ceiling bar stops at “local structural equivalence package.” Gray boxes beyond it read user response, system response, causal visibility. Add a link to the text output table, not a screenshot. **Speaker notes** Explain what PASS means. The inputs satisfy the schema and authorization boundary. Claims and source IDs match. The independent checklist has six passes. Exactly one declared factor appears in the diff. The run creates normalized claims, a complete diff, copied review, rollback manifest, and run manifest. PASS does not measure the hypothesized comprehension signal. The claim ceiling written by the tool explicitly excludes system response and causal visibility. **Teaching check** Ask learners to write one sentence beginning “PASS establishes…” and another beginning “PASS does not establish…”. Require exact stage language. **Alt text** An audit strip reports three claims, three sources, six checks, one changed factor, and zero errors; a ceiling bar blocks transfer to user or system outcomes. ## Slide 18 — Accessibility and validity are hard gates **On-screen text** Accessibility: language · headings · reading order · keyboard · focus · contrast · reflow · links · text alternatives Validity: claims · sources · schema · authorization · disclosure · privacy · non-deception Any core failure → block or rollback **Visual specification** Draw two parallel gate columns leading to release. Each column contains labeled test bars. The release path opens only when every non-exempt bar is passed with evidence. A target-metric arrow trying to bypass the gates is blocked. **Speaker notes** Accessibility is part of treatment integrity. A grid that visually groups claim and source may create a confusing reading order, narrow column, hidden overflow, or difficult focus route. Validity includes factual equivalence, source integrity, consistent schema, authorization, and honest disclosure. These are non-compensable. A favorable target result cannot pay for an inaccessible or misleading page. Record automated, manual, assistive-technology, and pending states separately. **Teaching check** Ask what to do if the treatment improves a comprehension proxy but breaks keyboard focus visibility. Expected: block or roll back; fix under a new version and redesign the study. **Alt text** Accessibility and validity gate columns must both pass before release; a target metric cannot bypass a failed gate. ## Slide 19 — Automated checks have a known ceiling **On-screen text** Automation can inspect selected invariants. Humans and assistive technology must review meaning and use. Automated · manual browser · screen reader · domain reviewer · release owner **Visual specification** Build a layered coverage map. Rows list tests; columns list five review modes. Filled cells show who can provide evidence. `html.parser` covers element counts, attributes, text, and IDs but leaves contrast, reflow, link comprehension, and screen-reader efficiency open. Use words and patterns, not a heat map alone. **Speaker notes** The deterministic auditor is intentionally small. It catches contract violations reliably, but it cannot establish all accessibility or factual questions. Manual browser work checks keyboard and responsive states. Assistive-technology review examines navigation and relationship recovery. Domain review validates claims and sources. The release owner verifies approvals and rollback. A single badge such as “accessibility score 100” hides these distinct responsibilities. **Teaching check** Ask which review can determine whether a source link makes sense when read out of context. Expected: human and assistive-technology review, supported but not replaced by automated link checks. **Alt text** A test-by-review-mode map shows automated coverage for structural invariants and open cells requiring browser, screen-reader, domain, and release review. ## Slide 20 — When two factors matter, use a factorial question **On-screen text** Two factors (A) and (B): control · (A) · (B) · (A+B) Estimate main effects and interaction—if independent units and review capacity support it. **Visual specification** Create a two-by-two design square. Factor A changes evidence proximity; Factor B changes heading labels. Every cell carries its own claim-lock and accessibility seal. At right, show main-effect arrows and an interaction bracket. Add a resource meter warning that four arms multiply review and sample needs. **Speaker notes** If proximity and labels must both change, a factorial design can separate them and estimate interaction. But it requires enough independently assigned units and full review of every arm. It is not automatically better than a narrow intervention. Underpowered cells or inconsistent gates can make the design less informative. The study question and resources determine the choice. **Teaching check** Ask what (A+B) provides that two separate one-factor tests may miss. Expected: evidence about interaction, where the joint effect differs from the sum or pattern of separate effects. **Alt text** A two-by-two factorial square compares control, two single factors, and their combination, with an interaction bracket and increased resource warning. ## Slide 21 — Sequential and bundled designs need honest estimands **On-screen text** Fallbacks: - reduce to one factor - predeclare sequence and stop rule - test factorial arms - call it a bundle Never hide a co-intervention as cleanup. **Visual specification** Show four parallel design routes from one problem. The sequential route has version gates and a predeclared stop sign. The bundle route encloses multiple changed elements in one treatment bracket. A deviation log runs beneath all routes, recording unexpected differences and dispositions. **Speaker notes** Complex redesigns are legitimate. The claim must match the treatment. A sequential design creates a new version and decision after each step under a frozen rule. A bundle design estimates the package, not individual components. Calling an unplanned wording change “cleanup” preserves neither design. Every deviation records discovery, affected files and claims, immediate containment, impact on the estimand, approval, and whether the analysis resets, becomes exploratory, or stops. **Teaching check** Ask what to claim when layout, headings, wording, and sources all change together with no separated arms. Expected: only the effect or behavior of the complete bundle under the tested conditions. **Alt text** Four honest design routes—reduction, sequence, factorial, and bundle—share a deviation log; unplanned cleanup is excluded. ## Slide 22 — Rollback is a tested hash transition **On-screen text** Treatment pair → replace with retained control pair → verify HTML and CSS hashes → run acceptance checks Rollback failure is a stop condition. **Visual specification** Draw a state-transition diagram with treatment HTML/CSS hashes at left, an authorized replacement action in the middle, and control hashes at right. A verification gate compares both exact hashes and reruns core checks. Preserve a side branch to the archived treatment evidence, making clear rollback does not erase history. **Speaker notes** Rollback is more than retaining a backup. It needs exact identities, authority, instructions, execution, and acceptance evidence. L05 retains both arms and records all four HTML/CSS hashes. The local exercise replaces the treatment pair with the control pair and verifies both restored hashes. The treatment and audit remain archived for review. If a hash mismatches or the restored state fails checks, containment is incomplete. **Teaching check** Ask why deleting the treatment after rollback is wrong. Expected: it destroys evidence needed to inspect the change, deviation, decision, and possible incident. **Alt text** A hash-addressed transition restores control files, verifies both hashes and checks, and preserves the treatment evidence in a separate archive. ## Slide 23 — Read PAPER evidence through the intervention stage **On-screen text** - **PAPER-12:** fixed ten-item slate; no upstream inclusion claim - **PAPER-21:** fixed candidate contexts; utility regressions retained - **PAPER-23:** reconstructed stages; adverse body-only results retained - **PAPER-08:** cited-URL observational audit; no editing causality - **PAPER-09:** synthetic tourism simulation; no established production effect - **PAPER-28:** every percentage and threshold audit-only **Visual specification** Create six evidence cards arranged by stage: fixed candidate, reconstructed pipeline, and observational/simulated audit. Each card has “use,” “negative/null,” and “forbidden transfer.” PAPER-28 is covered by a diagonal audit-only pattern. Do not reproduce any source graphic or headline styling. **Speaker notes** The reading route is deliberately heterogeneous. PAPER-12 changes a target already inside a fixed slate. PAPER-21 learns rewriting preferences in controlled contexts and does not preserve every utility measure. PAPER-23 reruns a reconstructed pipeline and shows stage-specific gains and losses. PAPER-08 observes selected cited pages rather than randomizing edits. PAPER-09 uses synthetic targets and a small simulated study. PAPER-28 has unresolved unit, version, factual-preservation, and reproduction issues, so no reported percentage, weight, threshold, or platform effect enters W09 as a finding. **Teaching check** Ask which paper proves that evidence cards improve current commercial citation. Expected: none; each has a bounded design and none tests this exact L05 intervention on a current commercial surface. **Alt text** Six stage-labeled evidence cards retain legitimate uses, negative results, and forbidden transfers; PAPER-28 is visibly audit-only. ## Slide 24 — Exit ticket: what changed, what stayed fixed, what remains unknown **On-screen text** Complete three sentences: 1. “The declared factor changed ___.” 2. “Factual and accessibility gates preserved or blocked ___.” 3. “This local result does not establish ___.” Retain · revise · revert · inconclusive **Visual specification** Return to the control/treatment diagram from Slide 01. Add a diff lens between the pages, two release gates beneath, and a rollback arrow. A final decision diamond offers retain, revise, revert, or inconclusive. Beyond the local boundary, user and system outcome boxes remain unfilled. **Speaker notes** For L05, the declared factor is evidence-layout proximity in one CSS rule. Three claim strings, three source identities, title, language, DOM order, and other CSS remain locked. Six review checks pass in the fixture, while broader accessibility review remains pending. The audit establishes local structural equivalence, not human benefit, retrieval, ranking, citation, or production transfer. A learner should choose a decision based on the predeclared gate, not on enthusiasm for the design. **Teaching check** Collect the three sentences. Reject any answer that uses guarantee language, omits the local-fixture scope, or treats PAPER-28 percentages as evidence. **Alt text** Control and treatment pass through a complete diff and two release gates toward a bounded decision, while user and system outcomes remain unmeasured.