# W02 Slide Script — Evidence Before Expression ## Production conventions - **Canvas:** 16:9, ivory background, deep purple structure, muted gold emphasis, charcoal text. Minimum 30 pt body and 44 pt title. - **Visual grammar:** claims are rounded rectangles, evidence passages are clipped-paper shapes, and source identities are hexagons. Solid, dotted, double, and broken lines duplicate edge meaning; color is never the only code. - **Evidence grammar:** every source use is paired with a visible “can support / cannot establish” boundary. Paper IDs are routes, not validity badges. - **Originality:** all diagrams are original instructional specifications. Do not reproduce a paper figure, commercial interface, company logo, or reference-TeX artwork. - **Delivery:** the deck is planned for a 90-minute seminar. Timing has not been observed in a rehearsal or class. ## Slide 01 — Evidence before expression **On-screen text** > W02 · Evidence before expression > Essential question: Which source can support which claim? Footer: “A citation beside a paragraph is not yet an evidence path.” **Visual specification** Center one polished headline card. Behind it, reveal six narrow claim cards as offset layers. Only the headline is filled; the six cards are outlined to imply hidden structure. Add a small key: sentence → atomic propositions. **Speaker notes** Open with the unit’s central inversion: do not begin by making prose more persuasive. Begin by making the factual substrate inspectable. Explain that one sentence can contain several claims requiring different evidence. Today’s output is a claim ledger and release decision, not a list of writing tactics. State that every case value is synthetic and that the two paper routes are bounded teaching sources. **Alt text** “A polished headline sits in front of six separate outlined claim cards, showing that fluent prose can conceal multiple testable propositions.” **Learning check** Ask learners to write what a bare URL cannot tell a reviewer. ## Slide 02 — Autopsy of a persuasive headline **On-screen text** > “After Northstar added FAQ schema, citations across all AI engines doubled and AI-generated sales rose 31% in six weeks.” Underline: **after · all · doubled · generated · 31% · six weeks** **Visual specification** Typeset the fictional headline across three lines. Use six numbered gold underlines, each connected to a margin tag: time order, population, comparison, causal construct, quantity, window. Include a “SYNTHETIC” stamp. **Speaker notes** Read the headline once without critique. Then ask what each underlined word commits the publisher to measuring. “After” supplies sequence but not cause. “All” defines a population. “Doubled” requires compatible denominators. “AI-generated sales” requires an operational definition and causal link. The sentence compresses deployment, observation, cause, transport, commercial measurement, and attribution. **Alt text** “A fictional vendor headline has six load-bearing phrases underlined and labeled by the type of evidence burden they create.” **Learning check** Thirty seconds: list at least four independently false propositions hidden in the sentence. ## Slide 03 — W02 outcomes **On-screen text** By the end, you can: 1. Atomize a compound claim. 2. Record source identity and dependence. 3. Assign a claim-relative ceiling. 4. Type an exact evidence edge. 5. Preserve contradiction and unknowns. 6. Release, narrow, escalate, or block. **Visual specification** Create a six-step horizontal workflow. Each step has a small output icon: split card, identity record, ceiling label, locator pin, conflict fork, release gate. The final gate has six labeled exits rather than a green/red binary. **Speaker notes** Frame the outcomes as production capabilities. Atomization is not an end in itself; it enables local support decisions. The release gate includes “unresolved” and “block” as successful analytical outcomes. A learner has completed W02 when another reviewer can reconstruct why each statement was approved or withheld. **Alt text** “Six sequential steps lead from splitting a claim to a release gate with multiple accountable actions.” **Learning check** Which outcome is impossible if the original wording is overwritten rather than preserved? ## Slide 04 — The paragraph is too large **On-screen text** **Unit of audit:** atomic claim Not: paragraph · page · source count · citation count One strong edge must not launder adjacent claims. **Visual specification** Show one paragraph box containing four differently shaped claim tokens. A single source hexagon connects solidly to claim 1. A translucent “credibility wash” incorrectly spills across claims 2–4; cross it out. Beside it, show the correct local edge. **Speaker notes** Explain citation laundering. A deployment log can strongly support one proposition while saying nothing about causal effect or cross-platform performance. Review at claim level so support remains local. Charts and captions also contain claims, so sentence counting is not enough. **Alt text** “One source correctly supports only the first of four claims; an invalid shaded spill toward the other claims is crossed out.” **Learning check** Ask for an example where one official source is decisive for one clause and irrelevant to the next. ## Slide 05 — One falsifiable predicate **On-screen text** Compound: “Schema was deployed **and** citations rose.” Atomic rows: - A. The bundle was deployed to the recorded page set. - B. Exported citation incidence changed between named windows. - C. The bundle caused that change. **Visual specification** Use a branching diagram. The compound sentence at left splits into three cards. Cards A and B have separate observation icons. Card C carries a comparator icon and a dotted question mark to show a different design burden. **Speaker notes** Define the working criterion: a row should contain one independently testable predicate. Deployment may occur without outcome change; outcome change may occur without deployment; and sequence does not identify cause. Atomization keeps the first two observations available even if the causal row fails. **Alt text** “A compound sentence branches into deployment, descriptive change, and causal-effect claims, with the causal claim requiring additional design evidence.” **Learning check** Which two rows could both be true while row C is false or unresolved? ## Slide 06 — Atomic claim grammar **On-screen text** `subject → predicate → object/value` Bound by: `population · unit · system/surface · time · locale · conditions · modality · uncertainty` **Visual specification** Design a sentence workbench. The subject, predicate, and object sit as three large blocks. Eight smaller boundary chips lock around them like a frame. An empty chip reads “unknown,” showing that absence is recorded rather than guessed. **Speaker notes** Walk through the fields. Modality distinguishes descriptive, comparative, causal, normative, predictive, and prescriptive claims. Uncertainty includes intervals, missingness, disagreement, and unknown states. The grammar need not appear in public prose, but it must exist behind any audited expression. **Alt text** “Three core claim blocks are surrounded by eight scope and uncertainty fields, including an explicit unknown value.” **Learning check** Give learners “citations increased” and ask which boundary fields are absent. ## Slide 07 — Modality changes the evidence burden **On-screen text** Observed after ≠ caused by | Modality | Example burden | |---|---| | Descriptive | defined observation | | Comparative | compatible units and baseline | | Causal | counterfactual design | | Normative | governing authority and scope | | Predictive | target, horizon, validation | **Visual specification** Five vertical lanes lead to different evidence gates. The descriptive lane ends at an observation card; the causal lane passes through comparator, confound, and uncertainty checkpoints. Keep lane lengths visibly different without implying importance. **Speaker notes** Emphasize that changing a verb changes the estimand. A before/after difference is evidence of sequence and descriptive change under a protocol. It is compatible with many explanations. A causal claim requires a design that addresses the relevant alternatives. **Alt text** “Five claim modalities travel through different evidence requirements; the causal path includes comparator and confounding checks.” **Learning check** Replace “drove” with a verb that matches a purely descriptive record. ## Slide 08 — Source identity before source use **On-screen text** Record: issuer · type · title · version · date · status · method · artifacts · license · canonical route · correction path · unknowns **Visual specification** Create a source passport with labeled fields. Three fields—method, license, correction path—are blank but visibly marked “unknown.” A DOI badge and institutional seal sit outside the required-field grid, labeled “identity cues, not evidence completion.” **Speaker notes** A URL is a location, not an identity record. Explain why stable version, method, artifacts, and correction path matter. A DOI can identify a paper but does not validate its design. An official logo can establish issuer but does not prove implementation or effect. Unknown fields remain unknown. **Alt text** “A source passport lists required identity fields and leaves three unknowns visible; DOI and logo cues do not fill those gaps.” **Learning check** What source change should automatically trigger claim re-review? ## Slide 09 — Independence is not source count **On-screen text** 9 source cards → 5 evidence families Common origin · syndication · shared data · overlapping authorship · common sponsor **Visual specification** Place nine source hexagons in five outlined family clusters. Four internal documents share one family trunk. Use identical hatch patterns for shared origins. A counter reads “9 documents / 5 families / independence still claim-specific.” **Speaker notes** Explain that repetition can be legitimate documentation yet not independent corroboration. A press release, partner blog, and webinar may repeat one dataset. Dependence matters when reviewers treat count as confidence. It does not make the sources useless; it changes what aggregation can mean. **Alt text** “Nine source documents are grouped into five origin families, with four internal records connected to one shared trunk.” **Learning check** Why might four operational records still be valuable even though they are not independent? ## Slide 10 — Seven ceilings, not a prestige ladder **On-screen text** `L` local artifact · `N` normative/official · `A` primary research `B` pedagogy/implementation · `C` platform documentation `D` vendor observation · `E` practitioner synthesis Footer: “Ceiling = ordinary claim role within scope.” **Visual specification** Arrange the seven codes around a circle, not vertically. In the center place the question “What claim is being asked?” Draw a bidirectional spoke from each code. Avoid stars, scores, heights, or trophy imagery. **Speaker notes** The circle prevents a high-to-low interpretation. A platform help page can be the best source for its declared syntax. A controlled study can be the best source for its bounded effect. Neither substitutes for the other. Compound codes are possible, but averaging them into credibility is not. **Alt text** “Seven source-role codes surround the central claim question in a non-hierarchical circle.” **Learning check** Which ceiling is ordinarily strongest for the current documented syntax of one named platform? ## Slide 11 — Can support / cannot establish **On-screen text** For every source, write both: > This source can support… > This source cannot establish… Example: official eligibility documentation ≠ observed inclusion or effect. **Visual specification** Split the slide into two equal panels. The left contains a bounded green-outline proposition; the right contains three purple-outline excluded propositions. Use explicit words, not thumbs-up/down icons. **Speaker notes** The negative boundary is part of the evidence record. It prevents an official statement about eligibility from becoming proof of retrieval, citation, or ranking benefit. Ask learners to write the two sentences before they decide whether a source is “good.” **Alt text** “A source’s supported proposition is shown beside three clearly excluded propositions, emphasizing a bounded role.” **Learning check** Complete both sentences for a national standard used in a visibility project. ## Slide 12 — The Claim–Evidence–Source graph **On-screen text** Claim → exact evidence → versioned source Edges: support · partial · contradiction · qualification · non-informative · out of scope **Visual specification** Use three horizontal node bands. Claim cards sit above clipped evidence passages, which sit above source hexagons. Show one solid support edge, one dotted partial edge, one double contradiction edge, and one broken out-of-scope edge. Include a text legend. **Speaker notes** Explain why the middle layer matters. The exact passage or data slice bears on the claim; the source provides provenance and governance. A whole-document URL is too coarse. The graph preserves negative and qualifying relations instead of deleting them. **Alt text** “Atomic claims connect through typed lines to exact evidence passages and then to versioned source identities.” **Learning check** What information is lost when the evidence layer is removed? ## Slide 13 — Exact locator, compatible scope **On-screen text** Minimum resolvable edge: `claim ID + source version + exact locator + edge type + reviewer/date` Topical relevance is not entailment. **Visual specification** Show a source page with one highlighted paragraph and a locator pin labeled “§3 ¶2.” A nearby unhighlighted paragraph shares keywords but has a crossed-out edge to the claim. Display scope chips for platform, date, and unit on both claim and passage. **Speaker notes** A document can mention schema, citations, and leads without supporting a schema-to-sales causal proposition. Match object, population, system, time, and modality. The locator enables a second reviewer to reconstruct the decision and revisit it after a version change. **Alt text** “Only one exact paragraph has a valid edge to the claim; a keyword-similar passage is explicitly non-informative.” **Learning check** Name two acceptable locator types for a dataset or recording. ## Slide 14 — Five claim dispositions **On-screen text** Supported · Partially supported · Contradicted · Unresolved · Out of scope Footer: “No evidence found is not the same as false.” **Visual specification** Create five equal status cards with definition fragments and distinct shapes. Avoid traffic-light ordering. Under “unresolved,” show two incompatible edges retained. Under “out of scope,” show non-overlapping scope brackets. **Speaker notes** Define each status. Partially supported usually means the expression must narrow. Contradicted requires relevant incompatible evidence. Unresolved records conflict or insufficient discrimination. Out of scope means the evidence never observed or governed the named population or outcome. **Alt text** “Five non-ranked disposition cards distinguish support, partial support, contradiction, unresolved conflict, and scope mismatch.” **Learning check** Which status fits a cross-platform claim when only one surface was measured? ## Slide 15 — Contradiction is data **On-screen text** Do not delete the inconvenient edge. Compare: scope → denominator → method → version → date → dependence Then narrow, reject an edge, or remain unresolved. **Visual specification** Two evidence passages point with double lines to one claim. Between them, place a six-step adjudication ruler. The final state remains a fork, not a forced winner. **Speaker notes** Contradiction may dissolve after comparing denominators or windows, but that comparison must be recorded. Silent selection destroys the audit trail. If the record cannot distinguish explanations, “unresolved” with a next action is the correct output. **Alt text** “Two conflicting evidence passages remain connected to a claim while reviewers compare six scope and method dimensions.” **Learning check** What is the first comparison when two sources report different percentages? ## Slide 16 — Five transport tests **On-screen text** 1. Population 2. System 3. Time 4. Causal 5. Outcome Ask: “What bridge supports the move?” **Visual specification** Draw five short bridges, each spanning a labeled gap: sample→population, one surface→many, snapshot→present, association→effect, citation→sales. Three bridges are incomplete. Each has a small evidence requirement beneath it. **Speaker notes** Use the bridge metaphor carefully: this is an inference audit, not a mechanism diagram. For each generalization, name the measured object and the expressed object. If the bridge is absent, narrow the claim or define the study needed to build it. **Alt text** “Five inference bridges connect measured and expressed objects; incomplete bridges mark unsupported transport.” **Learning check** Which transport error turns assisted transactions into sales caused by schema? ## Slide 17 — Every percentage needs a noun **On-screen text** Synthetic export: - January: 24 / 100 response opportunities - February: 41 / 100 response opportunities - Change: +17 percentage points; +70.8% relative - Not: doubled **Visual specification** Use two 10×10 unit grids, with 24 and 41 cells outlined. Beside them show the absolute and relative calculations in equal-sized boxes. Draw a target marker at 48 labeled “required for doubling 24.” **Speaker notes** Name the metric: exported responses with at least one visible Northstar citation per recorded response opportunity. The export has parser and prompt comparability limits. Correct arithmetic does not repair measurement or identify causality. **Alt text** “Two hundred-cell grids show 24 and 41 events, with 17-point and 70.8-percent changes and a doubling threshold of 48.” **Learning check** Why can both +17 percentage points and +70.8% be correct? ## Slide 18 — Outcome substitution in the CRM **On-screen text** 13 → 17 assisted transactions ≈ +30.8% But the record does not contain: revenue · incremental sales · citation join · counterfactual **Visual specification** Show two database tables separated by a gap: citation monitor and CRM. Their key columns do not match. Above the gap, a proposed join arrow is crossed out. Below, define `assist-30d-v1` in plain text. **Speaker notes** The count change is a valid dashboard description under its rule. It is not “AI-generated sales.” The CRM includes any transaction with a tagged session in the prior 30 days, has no revenue field, and cannot join to a response or citation. Construct mismatch and causal mismatch are separate failures. **Alt text** “A citation-monitor table and CRM table have no common key, so a proposed causal join is crossed out.” **Learning check** Write the strongest descriptive sentence the CRM alone can support. ## Slide 19 — P20: model sensitivity is not truth **On-screen text** P20 route: two opposing passages → model stance adoption Useful for: bounded passage-feature sensitivity Not evidence of: truth · credibility · retrieval · citation · human persuasion **Visual specification** Create an original two-passage choice schematic. Both passages flow into a model box and one stance token exits. Around the diagram place five excluded-outcome labels outside a dashed experimental boundary. Do not recreate a paper figure. **Speaker notes** Explain the reviewed design boundary: controversial binary questions, opposing decontextualized passages, model-specific filtered samples, and dated model snapshots. Semantic relevance is the consistent instructional signal, but “winning” is stance adoption in this setup—not factual accuracy or source authority. **Alt text** “Two opposing passages feed a bounded model-choice experiment; truth, retrieval, citation, credibility, and human persuasion remain outside it.” **Learning check** What title-level overreach would occur if “convincing” were rewritten as “credible”? ## Slide 20 — P40: robustness is attack-specific **On-screen text** P40 route: replace 0 / 1 / 2 / 3 items in a three-evidence set Useful for: attack-specific degradation and aggregation risk Not evidence of: real poisoning prevalence or universal method ranking **Visual specification** Draw four rows of three evidence tiles. From top to bottom, zero, one, two, and three tiles change hatch pattern. Four verification-method arrows lead to an outcome matrix with different winners by column. Add “one-run point estimates” beside the boundary. **Speaker notes** Use the route to challenge naïve source counting. Controlled replacement can degrade verification differently across attacks and procedures. The reviewed version has one run per fixed configuration and incomplete release details; it does not tell us how often production sources are compromised or which method is universally robust. **Alt text** “A three-evidence set is progressively replaced, and different verification methods perform differently across attack columns.” **Learning check** Why does good clean performance fail to settle poisoned robustness? ## Slide 21 — Northstar source map **On-screen text** 9 cards · 5 families · 6 claims Strong for one fact ≠ strong for the headline **Visual specification** Place the six `CLM` cards above nine `SC` hexagons grouped in five family bands. Draw only three example edges: SC-02 supports recorded deployment counts, SC-03 partially supports exported incidence, and SC-09 makes the cross-platform claim out of scope. Leave remaining edges for live completion. **Speaker notes** Transition to the worked case. Remind learners that the source cards have different roles. The deployment manifest is stronger than the blog for page state; the platform help card is stronger for declared eligibility; the methods memo is decisive about local measurement scope. None supports the whole headline. **Alt text** “Six claims sit above nine sources grouped into five families, with three illustrative typed edges and the rest intentionally blank.” **Learning check** Which card is credible yet unsuitable for proving a Northstar effect? ## Slide 22 — Model adjudication **On-screen text** | Claim | Disposition | |---|---| | 120 pages | narrow to 118 completed / 116 valid | | doubled | contradicted | | schema caused change | unresolved | | all engines | out of scope | | assisted count +31% | supported within rule | | sales caused by schema | unsupported / block | **Visual specification** Use six equal rows. Put original fragments at left, dispositions in the middle, and release actions at right. Use shape-coded disposition badges and explicit words. Do not use a single green score. **Speaker notes** Walk through why the resolution preserves some facts and blocks others. The result is not “the campaign failed.” The result is that the evidence supports a documented deployment state, an uncorrected monitor change, and an assisted-count change under a stated rule—while not identifying a schema effect, cross-platform result, or sales effect. **Alt text** “Six claim fragments receive different dispositions and release actions rather than one overall verdict.” **Learning check** Which row requires a new measurement definition before it even reaches causal design? ## Slide 23 — Bounded rewrite **On-screen text** > In one synthetic exploratory monitor, exported visible-citation incidence changed from 24/100 to 41/100 opportunities. Prompt wording, interface, parser context, content, links, latency, and publicity also changed; the record does not identify a schema effect or transport to other surfaces. **Visual specification** Place the rewritten paragraph in an ivory evidence box. In the margin, align six small chips: surface, numerator/denominator, windows, co-interventions, causal boundary, transport boundary. A separate CRM chip is visibly detached. **Speaker notes** Compare the rewrite with the opening headline. The new version is longer because it restores measured objects and boundaries. It is also more reusable: another analyst can locate the protocol and decide what follow-up evidence is needed. The CRM statement remains separate because there is no valid join. **Alt text** “A bounded rewrite is annotated with six scope fields, while the unrelated CRM outcome remains separate.” **Learning check** Ask learners to shorten the rewrite without reintroducing cause, universality, or outcome substitution. ## Slide 24 — Release gate and exit ticket **On-screen text** Before release, can a second reviewer reconstruct: - the atomic claim? - the exact evidence and source version? - the ceiling and disposition? - the contradiction or unknown? - the release action and next evidence step? **Exit:** name one credible source that is unsuitable for your claim—and why. **Visual specification** Show a five-field checklist feeding a six-exit gate: approve, narrow, caveat, seek evidence, escalate, block. Under the gate, place an empty two-sentence exit-ticket card. **Speaker notes** Close with the reconstructability standard. The objective is not maximum positivity. It is the strongest expression the evidence can carry. Remind learners that a block or unresolved status is a completed analytical result. Preview L04: the seed ledger becomes a Claim–Evidence–Source graph with adjudication. **Alt text** “Five reconstructability checks lead to six accountable release actions and a written exit ticket.” **Learning check** Collect the two-sentence answer and one atomic row before learners leave.