# W09 Transcript Equivalent — Evidence-Rich Content Architecture

## Status and use

No W09 recording exists. This document is a planned readable or recordable equivalent, not a transcript recovered from audio or video. The durations below are **instructional planning budgets** rather than timestamps or observed speaking times. A future recording requires verified duration, synchronized captions, reviewed transcript, audio description, chapter markers, and accessibility testing before it can replace this text route.

## Planned chapter budget

| Chapter | Planned duration | Focus |
|---|---:|---|
| 1 | 4 minutes | Structure without ranking superstition |
| 2 | 4 minutes | Governed claims and factual locks |
| 3 | 4 minutes | Evidence-page anatomy |
| 4 | 4 minutes | Co-located scope, methods, dates, and limitations |
| 5 | 4 minutes | Visible content and schema consistency |
| 6 | 4 minutes | Factual equivalence and intervention cards |
| 7 | 4 minutes | One-factor local comparison and deterministic audit |
| 8 | 4 minutes | Accessibility and validity release gates |
| 9 | 4 minutes | Factorial, sequential, and bundled fallbacks |
| 10 | 4 minutes | Rollback, deviations, and evidence routes |
| 11 | 2 minutes | Bounded conclusion and handoff |

**Total planned route: 42 minutes.** The route is designed to fit a 35–45 minute no-video equivalent, but has not been rehearsed or piloted.

## Chapter 1 — Structure without ranking superstition

W09 begins with a useful but easily distorted idea: structure can make verified information easier to use. A reader can locate a scoped answer, compare values, inspect methods, see dates, find limitations, resolve sources, and understand what changed. A declared parser can recover stable claim and source IDs. A page can expose correction routes and retain qualifications near the claims they constrain.

None of those benefits is the same as guaranteed retrieval, ranking, or citation. A heading is not a universal signal. A table is not inherently authoritative. Structured data do not become factual merely because a validator accepts their syntax. The value of architecture depends on the task and the evidence available for the stage being discussed.

We will keep four evidence objects separate. First, a structure can be easier for a person to verify. Second, it can be more consistent for a declared parser. Third, it can produce a stage-specific contrast inside an open, reconstructed pipeline. Fourth, a live system can display a time-bounded outcome under recorded conditions. Each transition needs a new study.

The local L05 fixture directly addresses part of the second object. It parses static control and treatment pages, compares their claim and source records, verifies selected HTML invariants, isolates a declared CSS difference, and produces hashes and rollback evidence. It contains a hypothesis about human comprehension, but no human participates. It contacts no external system.

This distinction changes the success criterion. A strong W09 result may be that the page is easier to audit, factually stable, accessible, versioned, and reversible, even when a later target metric is unchanged. A null system response does not erase the editorial value. Conversely, a favorable metric cannot excuse a changed claim, hidden limitation, inaccessible interaction, or failed rollback.

Think of evidence architecture as a verification mechanism before you think of it as an optimization. The primary questions are: Can a reviewer find the right proposition? Does its evidence resolve? Are method, date, and scope visible? Can another representation silently say something stronger? Can we prove what changed? Can we return safely?

## Chapter 2 — Governed claims and factual locks

A page should be a view of a governed fact system, not the sole place where facts live. An internal claim record can include stable ID, canonical entity, proposition, value and unit, population, locale, validity interval, evidence span, source identity, independence, owner, review status, conflicts, approved wording, and correction history. Visible HTML, metadata, structured data, feeds, and downloadable records should derive from compatible versions of that system.

Why atomic claims? One polished sentence can assert certification, performance, causality, population coverage, and business effect at once. A single citation may support only one part. We split independently challengeable propositions so each receives a scope and evidence edge. Atomic does not mean context-free. A percentage keeps its denominator. A certification keeps its product version and validity period. A comparison keeps its comparator and method. A limitation that changes interpretation remains attached.

Before creating treatment, freeze a claim lock. For L05, the lock uses exact approved strings and source-ID sets. CL-001 says that a visible citation does not disclose a hidden retrieval stage and links to S-001. CL-002 separates citation presence from claim support and links to S-004. CL-003 defines a reproducible surface record through system, locale, time, query, and response artifact and links to S-008.

Exact equality is appropriate because the factor is layout. If the intervention intentionally rewrites expression, the equivalence contract becomes more complex. It must compare normalized proposition, quantities, units, denominators, conditions, dates, source identity, attribution, disclosure, and epistemic force. “May improve” is not equivalent to “improves.” “Among returned responses” cannot disappear. A source swap is not harmless when visible wording stays the same.

Text similarity is useful for triage, not certification. Two passages can share almost every token while differing in one material date or modal. Embedding similarity can remain high after a limitation disappears. PAPER-28 reports similarity-related quantities, but W09 quarantines them with every other PAPER-28 percentage and threshold. They do not establish factual preservation.

The claim lock also protects the experiment's interpretation. If claims change along with layout, any response can arise from the new information rather than the structure. A factual-equivalence failure is therefore both an integrity failure and a design failure.

## Chapter 3 — Evidence-page anatomy

Now reconstruct an evidence page by function. Start with identity and version: what entity, product, owner, locale, publication state, and last substantive review does this page represent? Next place an answer-first scoped summary. It should state the central proposition and critical condition early without burying an exception elsewhere.

Then use atomic claim blocks. Each can carry a claim ID, proposition, status, date, evidence locator, source identity, and material limitation. An evidence table is helpful when values are genuinely comparable. Its cells should preserve unit, population, comparator, method, date, and source. If denominators differ, the page must show the difference rather than create a visually tidy false comparison.

Definitions clarify terms whose technical, legal, and commercial meanings diverge. The method section states how the result was collected, transformed, analyzed, and reviewed. Limitations and unknowns state where transfer fails. Source identities resolve exact origin, version, and relevant evidence rather than linking only to a homepage. A change history records corrections and supersession. A correction route names who can address a conflict and how evidence can be inspected.

These functions can be expressed differently on a short specification page and a long research report. The anatomy is not a mandatory visual template. The invariant is recoverability. A reviewer should not need to infer which product version a claim concerns or search another page to discover that the metric excludes failures.

An original diagram for this unit pairs a single-column page with a verification rail. Each region answers one question: what object, what claim, under which scope, based on what evidence, produced how and when, where it fails, from which origin, what changed, and who corrects it. The diagram carries a large note: anatomy, not outcome guarantee.

The answer-first principle does not justify removing nuance. The summary should include the most important boundary. If a result applies only to one locale and quarter, those conditions belong in the summary. Detail can expand below, but the high-level claim must not become stronger than the evidence.

Evidence tables deserve an extra warning. Native table markup can preserve the relationship between headers and cells when the information is genuinely tabular. W3C H51 is a useful bounded technique route for that implementation choice. It does not certify the whole page or establish WCAG conformance. A table can be semantically marked up and still compare incompatible populations, use unexplained abbreviations, overflow at zoom, or force an unusable reading sequence. The content model and the accessibility implementation must both be reviewed.

The correction route is equally important. An evidence page should tell a reader how to report a disputed claim, what identifier to cite, and who owns the review. The route does not promise that every challenge is correct. It creates an accountable intake and change process. When evidence changes, the correction should propagate through visible content, structured data, downloads, captions, and any controlled teaching or evaluation artifacts that reused the claim. The change log then records the old state rather than overwriting history.

## Chapter 4 — Co-located scope, methods, dates, and limitations

Information can be segmented, excerpted, skimmed, or quoted. A number separated from its unit and population can become misleading without any word being false in isolation. Co-located qualification keeps a claim and the conditions necessary to interpret it in one recoverable unit.

Imagine a train. The numerical claim is the engine. The unit, denominator, population, comparator, method, date, uncertainty, limitation, and source are cars. A well-designed evidence block lets a segmentation boundary fall before or after the train, not between the claim and the condition that defines it.

This does not mean repeating the same phrase throughout a page. It means preserving semantic completeness. Tables should hold values and conditions together. Definitions should sit near ambiguous terms. A material limitation should not live only in a collapsed footnote. A source link should resolve the relevant evidence object, not just a publisher.

Dates have different types. Publication date is not observation window. Observation window is not validity interval. Last reviewed date records an audit event; it does not refresh the underlying fact automatically. Next review date creates a governance trigger. A source can be newly published and describe old data, or remain valid while the page has not been recently reviewed.

Change history also needs types. A typo correction differs from a claim correction, evidence refresh, method change, structural intervention, or withdrawal. When W09 changes layout, that page state receives its own identity. If a number or scope changes at the same time, the intervention is no longer layout-only.

Limitations are part of the evidence, not an embarrassment. They reduce the risk that a local result is transported into another population or stage. They supply negative evidence against an overbroad interpretation. Removing a limitation for concision changes the factual message. Even moving it into an inaccessible or hard-to-discover interface can weaken equivalence.

A good checkpoint is to read an evidence block out of page context. Can you recover what was measured, for whom, compared with what, when, by which method, with what limitation and source? If not, the architecture has not yet achieved local recoverability.

## Chapter 5 — Visible content and schema consistency

Evidence pages often expose several representations: visible HTML, title and description metadata, structured data, link text, captions, alt text, feeds, and downloads. These form one claim surface. A hidden machine-readable layer cannot make claims that are stronger, newer, broader, or differently sourced than the visible evidence.

Build a consistency matrix. Each row is an atomic claim. Columns represent visible block, metadata, structured data, source link, caption or text alternative, downloadable record, and version. Mark each cell equivalent, absent by design, conflicting, stale, or unreviewed. Not every claim needs a schema property, so absence can be legitimate. Conflict and stale values block release.

PLAT-03, Google's general structured-data guidelines, supports two bounded rules: markup should represent visible or main page content, and correct markup does not guarantee a rich result. That is documentation about a Google-supported feature and policy surface. It is not a causal ranking study. PLAT-13, Schema.org for Developers, documents a shared vocabulary. It does not show that a platform uses a term in a hidden pipeline or that an outcome improves.

PLAT-02 provides Google-specific guidance about generative AI content and responsible publishing. The current official people-first guidance similarly supports original, sourced, non-manipulative material. These documents can inform a review checklist within their scope. They do not provide a scientific quality metric or guarantee.

Structured-data syntax validation and factual validation are different. A rating object can be syntactically valid and factually wrong. Suppose visible text says 4.2 out of 5 among 120 verified respondents in 2025, while schema says 4.8 and omits the count. The representation fails even if a validator accepts the JSON-LD. The fix must flow from the governed claim record, not from a desired display.

The same rule applies to alt text and captions. Accessibility text should communicate the image's purpose and evidence; it should not become a hidden promotional keyword channel. A downloadable PDF should not carry an older or stronger claim without a visible version warning.

Consistency reduces factual drift and makes the intervention auditable. It still does not create ranking evidence. A successful schema-visible review says the claim surfaces agree under the tested version.

## Chapter 6 — Factual equivalence and intervention cards

Factual equivalence is a field-by-field contract. For a layout treatment, lock claim IDs, exact text, quantities, units, denominators, population, entity and version, dates, qualifications, limitations, evidence, sources, attribution, disclosures, language, title, and reading order. The design also states what may change: the declared layout factor and only that factor.

The L05 checklist has six rows. Atomic claim text, numbers and qualifications, source identity, attribution, reading order, and rollback review all pass. The synthetic reviewer ID differs from the experimenter ID. This is a useful role invariant, but it does not prove a full expert or accessibility review occurred.

The intervention card freezes the scientific object. It needs identity and version, environment and authorization, factor, hypothesized stage, expected signal, falsification condition, non-claims, confounds, validation gates, stop conditions, and rollback. “Make it clearer” is not a factor. “Evidence-layout proximity through the `.evidence-block` style rule” is inspectable.

L05 names human comprehension as the hypothesized stage. The expected signal says that a later approved task may show fewer claim-to-source matching errors. It does not say that L05 observed such an effect. Non-claims explicitly exclude crawler, retrieval, ranking, citation, traffic, business, cross-engine, and production effects.

A mechanism hypothesis needs alternatives. Perhaps proximity makes matching easier. Perhaps the border attracts attention. Perhaps grid wrapping changes reading. Perhaps participants are familiar with cards. Perhaps assistive technology exposes the relationship differently. Confounds become design questions for a later study rather than excuses after the result.

Falsification matters. If matching errors do not change, or improve only at one viewport while accessibility worsens, the original mechanism is weakened or bounded. Do not add another factor after seeing a null. Record the result and decide under the predeclared rule.

## Chapter 7 — One-factor local comparison and deterministic audit

The L05 control and treatment pages are deliberately narrow. Both declare English, the same title, one `main`, one `h1`, one stylesheet, three claim blocks, and three source links. Their HTML differs only in the stylesheet filename. When that filename is replaced by a declared placeholder, the HTML is identical.

The stylesheets share root, body, main, anchor, focus-visible, and margin rules. The `.evidence-block` rule differs. Control uses only block margin. Treatment adds grid layout, two columns, gap, alignment, an inline-start border, and padding while keeping the margin. These declarations together constitute the conceptual factor `evidence_layout_proximity`. The fixture cannot estimate the separate effect of border, grid, gap, or padding.

The auditor uses the Python standard library. It parses both pages, checks selected semantic fields, extracts claim text and source IDs, compares them with the lock, reads the independent checklist, normalizes only the declared HTML and CSS locations, and builds a unified diff. It writes normalized claims, the diff, equivalence review, rollback manifest, run manifest, and audit report.

The fresh result is: three locked claims, three source identities, six review checks, one changed factor, and zero errors. Status is PASS. The claim ceiling states that this is a local structural-equivalence package only and no system response or causal visibility effect is measured.

That language is essential. Determinism describes the relation between fixed files, script, and outputs. It does not imply deterministic user response or platform behavior. A screenshot would be insufficient because it can omit metadata, off-screen content, DOM order, and source differences. The complete source diff is the audit object.

The result also prepares reproducibility. Input hashes identify all seven source files. Any edit creates a new state and requires a new run. Keeping generated outputs together prevents a reviewer from seeing PASS without the evidence that supports it.

Reproduction is stronger when the reviewer predicts the expected difference before execution. In the studio, learners inspect the intervention card and source files, write down which HTML and CSS lines should differ, and name the expected counts. They then run the tool in a fresh temporary directory and reconcile every output. A surprising extra line is not ignored because the summary still says PASS; it triggers investigation of the script, inputs, or contract. Conversely, a byte-level difference can be legitimate only when it falls inside the declared factor and passes the semantic locks.

The output bundle should remain together. `intervention_audit.json` reports status and ceiling. `normalized_claims.json` makes extracted equality inspectable. `before_after.diff` exposes changed source. `equivalence_review.csv` preserves the reviewer record. `rollback_manifest.json` identifies restoration. `run_manifest.json` locks inputs and parameters. None of these alone is the result. The bundle is the evidence path.

## Chapter 8 — Accessibility and validity release gates

Accessibility is not a secondary style check. It is part of treatment integrity. A visual grid can appear clearer while producing narrow wrapping, an order mismatch, hidden overflow, confusing link reading, or poor keyboard focus. A page cannot pass release because a target metric improved if a core evidence path became inaccessible.

For both arms, review language, title, landmarks, headings, DOM and visual order, keyboard access, focus visibility, link purpose, contrast, zoom and reflow, responsive behavior, text alternatives, and assistive-technology navigation. Native data-table semantics can be supported by W3C H51 when evidence is truly tabular. Citing H51 does not establish WCAG conformance. The implementation and test evidence remain necessary.

The L05 parser checks a limited subset: language, title, one main, one h1, stylesheet presence, source uniqueness, claim extraction, and certain equality rules. It cannot determine contrast, focus behavior in a browser, screen-reader efficiency, or cognitive comprehensibility. Record each test as automated, manual browser, assistive technology, domain review, pending, or not applicable. Do not collapse them into a single score.

Validity gates include factual and attribution lock, visible-schema consistency, authorization, non-deception, disclosure, privacy, and absence of hidden user-specific variants or cloaking. A fabricated statistic or undisclosed stronger schema claim is a hard failure. A live environment with ambiguous authorization is a stop. A broken rollback is a stop.

If a treatment fails a gate, revert or redesign. Fixing the issue creates a new version. Do not silently repair the file and keep the original result. The deviation log records what failed and how the estimand changes.

Accessibility and validity are called non-compensable because a gain in one objective cannot cancel severe harm in another. That is a governance decision embedded in the release contract.

Consider a concrete adverse case. On a wide screen, two-column evidence cards may bring claims and links visually closer. At narrow zoom, the link column may wrap into short fragments, or CSS ordering could diverge from the DOM sequence. A sighted mouse user might report easier matching while a keyboard or screen-reader user encounters more navigation effort. The correct outcome is not an average clarity score that hides the subgroup harm. Preserve both observations, apply the release threshold, and revise or revert. If the design is rebuilt, it receives a new treatment hash and a new test.

The release matrix should show test method, version, artifact, reviewer, threshold, result, and retest trigger. “Automated pass” is not interchangeable with “manual pass.” A browser audit at one viewport does not cover reflow. A screen-reader session with one browser and assistive-technology version is bounded evidence, not universal conformance. A domain reviewer can assess claim support but may not be qualified to make an accessibility judgment. Assign each decision to the appropriate evidence and role.

## Chapter 9 — Factorial, sequential, and bundled fallbacks

Real redesigns often involve several components. Evidence cards may require grouping, borders, headings, link text, responsive rules, and updated source blocks. If the research question needs the separate effect of two factors, a factorial design can compare control, A, B, and A plus B. It can estimate main effects and interaction when enough independent units and review capacity exist.

Factorial designs cost more. Every arm needs claim, schema, accessibility, and rollback review. Sparse or dependent units can make interactions unstable. The design should follow the decision, not the aesthetic ambition.

A sequential design is another option. Freeze the order of changes, decision rule, and stopping condition before observing results. Each step creates a new version and estimand. If the first result determines an unplanned second edit, the sequence becomes adaptive and needs appropriate exploratory language or a new preregistration.

Sometimes the honest answer is a bundle. If layout, wording, source links, and metadata change together and cannot be separated, report only the response to the complete package under the tested conditions. Do not assign the result to one favorite component.

The smallest-factor option remains valuable. Reduce the treatment until only one interpretable factor changes. This often produces a more useful diagnostic even if it does not look like a finished redesign.

Never hide co-intervention under the word polish. A link rewrite can change comprehension. A new heading can change segmentation. A source addition changes evidence, not just style. A responsive fix can alter reading order. Every difference belongs either inside the treatment definition, in a separated arm, or in the deviation log.

## Chapter 10 — Rollback, deviations, and evidence routes

Rollback is a tested transition between hash-addressed states. L05 retains the treatment HTML and CSS hashes and the control HTML and CSS hashes. Its instruction replaces the treatment pair with the control pair and verifies both. A rollback hash mismatch is a stop condition. The treatment, diff, and decision evidence remain archived after restoration.

The deviation log includes planned version, observed difference, detection time, detector, affected claim and files, immediate containment, cause if known, effect on the estimand, approval, and disposition. A null result, adverse wrapping, assistive-technology problem, unplanned wording difference, or failed restoration belongs in the record. The team decides whether to reset, downgrade to exploratory, redesign, or stop.

The PAPER route demonstrates why stage and validity ceilings matter. PAPER-12 rewrites a target already present in a fixed ten-item slate; it cannot establish upstream retrieval or live shopping outcomes. PAPER-21 uses controlled five-candidate contexts and includes utility regressions, so it does not support a universal writing preference. PAPER-23 reconstructs retrieval, reranking, and citation stages and retains negative body-only results; it supports stage-aware design, not production field weights.

PAPER-08 is an observational study of selected cited English B2B SaaS URLs without a causal page-edit assignment. PAPER-09 uses synthetic tourism targets and a small simulated visibility analysis. PAPER-28 remains audit-only because its percentages, thresholds, weighting, unit construction, product versions, factual-preservation route, and reproduction evidence do not support admission as a course result.

The PLAT route is also bounded. PLAT-02 contributes Google-specific responsible-publishing guidance. PLAT-03 contributes visible-content and eligibility rules for supported structured data and explicitly does not guarantee a rich result. PLAT-13 supplies vocabulary semantics. None proves a ranking, citation, or business effect.

The evidence routes are valuable because they help design tests and identify failure modes. They are not ingredient lists for guaranteed optimization.

Negative and adverse findings carry design information. PAPER-12 reports rewriting strategies that underperform inside its fixed slate. PAPER-21 contains utility cells that worsen. PAPER-23 reports that body-only changes can reduce stage outcomes and that gains do not peak at the same point. PAPER-09 describes a sample too small for a statistically established visibility conclusion. Preserving these results prevents W09 from becoming a list of favorable tactics. The lesson is to define the object, hold facts constant, measure the intended stage, and retain harms and nulls.

Official guidance has a different role. Google's people-first documentation can support an editorial control requiring original, sourced, non-manipulative content. PLAT-03 can support a test that markup reflects visible main content. W3C H51 can support a native-table implementation test. Each record still needs version, access date, applicability, implementation evidence, and an explicit ceiling. Documentation tells us what a named source declares; it does not prove that our page passed the test or produced an outcome.

## Chapter 11 — Bounded conclusion and handoff

Complete three sentences. First: “The declared factor changed evidence-layout proximity through the named CSS rule and stylesheet reference.” Second: “The local contract preserved three exact claims, three source identities, title, language, DOM order, other CSS, six review checks, and rollback evidence.” Third: “No human comprehension, accessibility conformance, retrieval, ranking, citation, traffic, or production effect was measured.”

Then choose retain, revise, revert, or inconclusive under the predeclared gate. For the fixture, PASS supports retaining the package as a teaching example, not deploying it as a proven optimization.

The W09 handoff is a versioned intervention that can be inspected and undone. The work succeeds when structure makes evidence easier to recover without making the public claim stronger than its source—and when every unknown remains visible.