# W16 Slide Script — Capstone synthesis and defense ## Visual production contract All diagrams are newly designed for W16 as original instructional visual specifications. Do not import, trace, or restyle figures, screenshots, diagrams, or layouts from papers, the attached TeX source, websites, or L08. Use a warm white ground, ink-black type, cobalt for declared evidence, amber for pending review, and vermilion for blocked edges. Color must be redundant with labels, shape, position, or line pattern. Every data-bearing visual must ship with its text-table equivalent. Minimum body text is 28 px in a 16:9 frame; title is 44 px or larger. ## Slide 01 — Trust begins with a traversable claim **On-screen text** W16 · Validate Capstone synthesis and defense **What would make the result worth trusting?** 18–30 h capstone · linked L08 **Visual specification** Draw one large claim card at left and a release box at right. Between them, place ten small labeled stepping-stones: question, estimand, data, code, observation, result, figure, limitation, review, checksum. A continuous solid cobalt line reaches the release; a second dotted vermilion line breaks at “data.” Add a small caption: “Polish cannot bridge a missing edge.” No decorative science icons. **Speaker notes** Set the threshold for the session. Trust does not mean certainty or agreement. It means a skeptical reviewer can traverse a bounded claim through authorized evidence and transformations, see what failed, and understand the release decision. The capstone combines eighteen to thirty hours of prior design, lab, analysis, packaging, defense, and revision. Today’s L08 fixture is tiny on purpose: every edge is inspectable. A local validator can check identities and arithmetic, but human judgment remains necessary. **Teaching check** Ask learners to name the first edge they would inspect in a result they distrust. Accept different answers only if they identify a testable relationship rather than “the visual style.” **Alt text** A claim reaches a release through ten labeled evidence steps; a second path breaks at data, showing that a missing edge blocks release regardless of polish. ## Slide 02 — Local PASS has a narrow meaning **On-screen text** `W16 VALIDATION PASS` means: - package structure matches the contract; - frozen hashes and arithmetic match; - L08 routes to human review. It is **not** human review, independent reproduction, accessibility conformance, pilot, grading calibration, external review, or recording. **Visual specification** Create a two-zone gate diagram. The left zone, “automated integrity,” contains checked boxes for structure, identity, edges, arithmetic, and output inventory. A narrow arrow ends at an amber door labeled `READY_FOR_HUMAN_REVIEW`. Beyond it are six closed human gates, each with an empty checkbox. Use a heavy vertical boundary and the words “necessary, insufficient.” **Speaker notes** Machine language often sounds more conclusive than it is. We therefore begin by naming exclusions. The validator can decide whether this authored package and the frozen synthetic fixture are internally consistent. It cannot inspect scientific importance, legal authority, contextual safety, assistive-technology experience, instructor calibration, or learner outcomes. Nor does running the same command twice create independent reproduction; independence concerns people, environments, and access routes as well as process separation. **Teaching check** Give the statement “the script passes, so the capstone passed.” Learners must rewrite it as a bounded routing sentence. **Alt text** Automated checks lead only to an amber human-review door; six closed human gates remain beyond it. ## Slide 03 — Freeze five sentences before packaging **On-screen text** 1. Question 2. Estimand 3. Claim 4. Audience 5. Decision If one changes, inspect every downstream edge. **Visual specification** Arrange five index cards in a horizontal row. Each card contains its label plus one blank ruled line. From each card, a thin line descends to a common bracket labeled “versioned release premise.” Place a revision stamp above the estimand card and show concentric ripples touching later cards. The ripple pattern, not color, carries change propagation. **Speaker notes** These five sentences prevent a release from becoming a collection of artifacts without a decision. The estimand is the exact quantity, not a topic. The claim is the repeatable sentence the evidence permits. Audience controls interpretation and risk. The decision states what happens under favorable, null, adverse, or unreproduced outcomes. Freezing does not prohibit change; it forces a documented version transition. A new outcome or population may require a new analysis, not merely an edited headline. **Teaching check** Ask which of the five sentences “we study GEO performance” satisfies. The answer is none without further specification. **Alt text** Five versioned premise cards—question, estimand, claim, audience, and decision—share a bracket; a change ripple from estimand reaches downstream cards. ## Slide 04 — The claim-to-artifact matrix is the defense map **On-screen text** Rows: claims Columns: data · code · result · figure · limitation · reviewer · release · checksum Cells: artifact ID, `N/A + reason`, or **BLOCK** **Visual specification** Draw a three-row by eight-column matrix. Label rows `CAP-001`, `CAP-002`, `CAP-003`. Populate only representative cells: `DATA-001`, `NEG-001`, `LIMIT-001`, and one vermilion `BLOCK`. Use solid cell borders for required edges, dashed borders for contextual edges, and a key below. Keep all text visible; include the same matrix as a speaker-handout table. **Speaker notes** The matrix is not a checklist where every cell should be yes. Claim status determines required edges. A supported public headline needs data, code, result, representation, and limitation. A rejected claim may correctly have no supportive figure, but it must have a durable rejection rationale and scope ceiling. “Not applicable” is a reasoned state. A broken mandatory reference is a block, not a completeness percentage that can be averaged away. **Teaching check** Ask whether `CAP-003`, a rejected production claim, should receive a green result cell. Learners should say no and route it to rejection evidence and limitations. **Alt text** A three-claim traceability matrix uses artifact IDs, reasoned not-applicable cells, and a blocked cell rather than undifferentiated checkmarks. ## Slide 05 — Stable identifiers resist narrative drift **On-screen text** `CAP-001` → `DATA-001` · `CODE-001` · `RESULT-001` · `FIG-001` · `LIMIT-001` `CAP-002` → `DATA-001` · `CODE-001` · `RESULT-001` · `NEG-001` · `LIMIT-001` `CAP-003` → `NEG-001` · `LIMIT-001` **Visual specification** Use a bipartite graph with three rectangular claim nodes on the left and six circular artifact nodes on the right. Draw exactly twelve labeled lines. Solid lines indicate support; dotted lines to `CAP-003` indicate rejection/context. Add small count badges: 3 claims, 6 displayed artifact nodes, 12 edges. Note that the full dossier contains 16 artifacts. **Speaker notes** Titles change during editing. Stable identifiers give reviewers durable coordinates. The L08 ledger declares twelve edges across three claims. Some artifacts are reused: the same data and code support the primary and secondary arithmetic, while `LIMIT-001` bounds all claims. The dossier contains sixteen artifacts because packaging, planning, environment, response, and governance records also matter. Traceability is many-to-many, not a folder tree. **Teaching check** Have learners count the edges from the visible list before revealing the badge. Then ask why reuse of `DATA-001` does not make the claims identical. **Alt text** Three claim rectangles connect by twelve lines to six visible artifact circles; supported and rejected relations use different line styles. ## Slide 06 — L08 freezes twelve synthetic observations **On-screen text** Synthetic fixture `CAPSTONE-DEMO-001` | Arm | Events | Outcome positives | Citation-correct positives | |---|---:|---:|---:| | Control | 6 | 2 | 5 | | Treatment | 6 | 4 | 4 | **Visual specification** Create twelve equal tiles in two rows. Top row is control E1–E6; bottom row is treatment E7–E12. Each tile contains two glyphs: a filled or empty square for outcome and a filled or empty circle for citation correctness. Labels and glyph shapes encode values without relying on color. Place the exact four totals at right. **Speaker notes** The fixture is authored synthetic data, not users, queries, or platform observations. Each of twelve events has an arm and two binary measures. This makes manual verification possible. We can count every positive before executing code. The scale is not evidence of external validity; it is an instructional control. The full release must continue to say “deterministic synthetic fixture only.” **Teaching check** Ask learners to verify the four totals using the tiles. If a count differs, stop before discussing claims. **Alt text** Twelve labeled event tiles in control and treatment rows encode two binary measures with squares and circles; totals are control two and five, treatment four and four. ## Slide 07 — Reproduce the primary arithmetic **On-screen text** Outcome rate: - Control: (2/6=0.333333) - Treatment: (4/6=0.666667) - Difference: (4/6-2/6=0.333333) Claim ceiling: descriptive fixture arithmetic only. **Visual specification** Draw two fraction bars with six compartments each, two filled for control and four filled for treatment. Beneath them, show a subtraction balance yielding two of six. Use numeric labels at every step. Add a thick horizontal ceiling line above the phrase “no causal or production claim.” Provide the three-row text table beside the visual. **Speaker notes** This is `CAP-001`. The analysis and expected JSON agree exactly within their numeric representation. Reproduction establishes the declared transformation from these bundled rows to these values. It does not establish randomization quality, population representation, measurement validity, inferential precision, or production effect. The claim ceiling travels with the number wherever it appears—in the brief, figure, talk, and release note. **Teaching check** Ask for one justified verb and one unjustified verb. “Shows a descriptive difference” is justified; “proves improvement” is not. **Alt text** Two six-part fraction bars show two control and four treatment positives; subtraction yields two-sixths or 0.333333 under a descriptive-only ceiling. ## Slide 08 — Preserve the adverse secondary result **On-screen text** Citation-correctness rate: - Control: (5/6=0.833333) - Treatment: (4/6=0.666667) - Difference: (-1/6=-0.166667) `CAP-002` routes through `NEG-001`. **Visual specification** Place the primary and secondary differences on a centered zero axis. The primary marker sits right at +0.333333; the secondary marker sits left at −0.166667. Use distinct shapes plus explicit signs. Make both labels equal size. A bracket connects the secondary marker to a document card labeled `NEG-001`. **Speaker notes** Negative means adverse relative to a favored direction, not statistically significant. This secondary measure prevents a selectively positive story. It may reveal a trade-off, noise, or a measurement issue; this fixture cannot adjudicate among them. The proper action is preservation and bounded discussion. Equal visual weight matters because hiding an adverse result in tiny text would undermine the evidentiary release even if the CSV remains intact. **Teaching check** Ask why preserving `CAP-002` increases trust even though it weakens a promotional interpretation. **Alt text** A zero-centered axis places the primary difference at positive one third and the secondary difference at negative one sixth, linked to the negative-results artifact. ## Slide 09 — A rejected claim belongs in the ledger **On-screen text** Rejected `CAP-003`: “The intervention improves production GEO visibility.” Missing: production context · visibility outcome · causal design · external validity **Visual specification** Show the rejected sentence on a paper strip with four punched holes, each labeled with a missing evidence class. Do not use a red strike-through alone; add a large “REJECTED — RETAINED FOR PROVENANCE” stamp. Two dotted arrows lead to `NEG-001` and `LIMIT-001`. Leave no arrow to a result or figure. **Speaker notes** Deletion erases learning. Retention tells future editors exactly which tempting sentence the evidence did not earn. The twelve-event fixture lacks the design and outcomes required for production and causal scope. The defense should not attempt rhetorical rescue. It should route the reviewer to the rejection and replace the sentence with `CAP-001`’s descriptive wording. **Teaching check** Invite one piece of evidence that would be necessary, but not sufficient, to reconsider `CAP-003`. **Alt text** A retained rejected claim has four labeled evidence holes and routes only to negative-results and limitations records. ## Slide 10 — Deviations are provenance, not embarrassment **On-screen text** Preserve separately: - exclusions; - protocol deviations; - failed reproductions; - resolutions and consequences. Never overwrite failure with “fixed.” **Visual specification** Design an append-only ledger as four dated rows. Each row has columns event, trigger, decision, affected claim, status. The third row says “validator added as packaging control; estimand unchanged.” A vertical spine at left shows that later rows do not cover earlier rows. Use version tags v0.1, v0.2, v0.3. **Speaker notes** L08 records that no row, outcome, or analysis changed after planning; its validator was added later as a packaging control. That is still worth documenting. Real failures may expose locale dependence, missing packages, unauthorized access, or unstable data. Preserve the original error, the response, and whether the claim survives. A materially changed analysis should receive a new version and perhaps a new estimand. **Teaching check** Ask learners to distinguish “bug fix with unchanged result” from “analysis change that alters the estimand.” Both need records; only the latter necessarily redefines the analysis. **Alt text** An append-only four-row provenance ledger preserves an added validator, failures, resolutions, and affected claims across versions. ## Slide 11 — Reproduction is a clean-room question **On-screen text** Can a reviewer start with the declared release and regenerate the primary result? Required: inputs · environment · command · code · expected output · identity checks **Visual specification** Split the frame into Author Workspace and Reviewer Temporary Directory. Only a sealed release bundle crosses the boundary. Inside the reviewer area, show input, command, and regenerated result. A comparator checks regenerated versus expected values. Do not draw a hidden channel between workspaces. Label the reviewer directory “clean temporary path.” **Speaker notes** Rerunning code in the author’s configured directory is useful debugging but weak evidence of portability. L08 executes analysis and validation in a temporary directory with no network requirement. It compares the generated result with the frozen expected JSON and inventories five audit outputs. Independence additionally concerns who performs the test and whether they possess undisclosed knowledge or access; this package’s local run is therefore not independent reproduction by another party. **Teaching check** Ask what hidden dependency a clean directory might reveal: environment variables, unlisted files, network state, locale, or cached outputs. **Alt text** A sealed bundle crosses from an author workspace into a clean reviewer directory where inputs and code regenerate and compare a result without hidden channels. ## Slide 12 — Non-redistributable does not mean unverifiable **On-screen text** Dependency card: ID · issuer · version · locator · lawful acquisition · license/access · privacy/security · local SHA-256 · authorized reviewer · verification method · unresolved scope Never bundle restricted bytes. **Visual specification** Draw a public release box separated from a controlled vault by a thick boundary. The public box contains only a dependency card and hash. An authorized reviewer icon follows a lawful retrieval arrow to the vault, then a hash-match arrow to a local analysis. Add a blocked branch for unauthorized access. All arrows have text labels. **Speaker notes** Many legitimate projects cannot publish source bytes. A precise boundary preserves legality and auditability. If an authorized reviewer lawfully obtains the exact version, matches the checksum, and reruns the analysis, reproduction may be possible. If the reviewer can inspect only supplied output, call it partial or output verification. If acquisition fails, do not substitute author assurance; mark the claim unreproduced and apply the preregistered release rule. **Teaching check** Ask which two fields “available under license” fails to specify. Expected answers include exact version, locator, terms, authorized role, and verification method. **Alt text** A public bundle holds only a dependency card and hash; an authorized reviewer lawfully retrieves controlled material and matches identity, while unauthorized access is blocked. ## Slide 13 — Five hard gates cannot be averaged **On-screen text** Authorization ∧ Privacy ∧ License ∧ Safety ∧ Accessibility Any material `blocked` or `pending` → **NO PUBLIC RELEASE** **Visual specification** Create a series circuit with five labeled switches. Four are closed; the accessibility switch is open and labeled pending human review. The release lamp at right remains off. Below, show a crossed-out weighted-score gauge to make clear that high values elsewhere cannot offset a failed gate. Use switch position and text, not color alone. **Speaker notes** These gates answer different questions. Authorization cannot compensate for an incompatible license. Correct arithmetic cannot compensate for privacy harm. Machine-detected alt text cannot settle whether the figure is understandable. Use explicit states and named reviewers. “Pending” is a meaningful block when the gate is material. A project may still circulate privately within a lawful review scope, but that is a different release decision. **Teaching check** Offer a 98-of-100 aggregate governance score with privacy failed. Learners should reject public release and explain why aggregation is invalid. **Alt text** Five switches form a release circuit; one pending accessibility switch keeps the release lamp off, and a weighted score is crossed out. ## Slide 14 — Accessibility is a human release gate **On-screen text** Automate: presence, structure, syntax Human review: meaning, order, interaction, comprehension Alt text present ≠ accessible figure **Visual specification** Use two stacked inspection trays. The automated tray contains heading tree, alt attribute, table headers, and contrast flag. The human tray contains reading-order arrows, a screen-reader speech bubble, keyboard path, and the question “does the description convey the inference?” A final gate remains amber. Include a text-only ordered list of both trays. **Speaker notes** W16 visuals specify alt text, redundant encodings, large type, and text-table equivalents. Those are authoring controls. Formal conformance requires review against an applicable standard and actual outputs; user experience may require assistive-technology testing. The local validator deliberately refuses to convert design intentions into an accessibility-conformance claim. An inaccessible headline figure blocks release until repaired and reviewed. **Teaching check** Ask learners to improve “bar chart of results” as alt text. The response should state groups, values, comparison direction, and the descriptive-only ceiling. **Alt text** Automated checks inspect presence and syntax, while human review inspects meaning, order, interaction, and comprehension before an accessibility gate can close. ## Slide 15 — Checksums identify bytes, not truth **On-screen text** SHA-256 can show: - bytes match or differ; - release identity is stale or current. It cannot show truth, legality, validity, safety, or accessibility. **Visual specification** Place two identical document shapes on a balance with matching short hash prefixes. Next to them, place a third document differing by one visible dot and a nonmatching hash. Under the balance, add five sealed question boxes labeled truth, law, validity, safety, accessibility; no arrow connects the hash to them. **Speaker notes** L08 verifies eighteen required file checksums. The checksum manifest itself is handled separately rather than self-hashed in the same list. Identity matters because a reviewer must know which bytes produced a claim. After revision, recompute all dependent outputs and checksums. Merely updating a hash after changing data can conceal a stale figure or claim; inspect the dependency graph first. **Teaching check** Ask whether matching a downloaded corpus hash proves that the license permits redistribution. It does not. **Alt text** Two documents with matching hashes balance; a one-dot change has a different hash, while truth, law, validity, safety, and accessibility remain separate questions. ## Slide 16 — A release is a versioned evidence object **On-screen text** Release identity = version + artifacts + environment + commands + outputs + gates + notes + checksums Revision changes identity. **Visual specification** Design a circular release seal with eight labeled segments around a central version `0.1.0`. An arrow leads to `0.2.0`; changed segments—claim, figure, limitations, checksums—are shown with diagonal hatching. A change note bridges the two seals. Use no imitation of commercial certification marks. **Speaker notes** The bundle is more than a folder. It declares what belongs, what ran, what emerged, and what remains pending. Versioning allows comparison without overwriting history. A post-defense wording change may affect claims, figures, limitations, and reviewer responses even if the numeric result is unchanged. Release notes name that propagation. The seal is an identity metaphor only, not certification or approval. **Teaching check** Ask which artifacts must change if the headline is weakened but arithmetic stays fixed. At minimum: claim ledger, brief, perhaps figure/caption, limitations, response, release notes, and checksums. **Alt text** Two segmented release identities, versions 0.1.0 and 0.2.0, are connected by change notes; hatching marks changed narrative and checksum segments. ## Slide 17 — Machine outputs open review, not publication **On-screen text** L08 clean run: - 12 events · 2 metrics - 3 claims · 16 artifacts · 12 edges - 1 public headline · 1 figure - 19 required files · 18 verified checksums - 5 review outputs - decision: `READY_FOR_HUMAN_REVIEW` **Visual specification** Create a vertical audit receipt with six numbered lines and a perforated bottom edge. Beside it, draw an unopened door labeled Human Review. The receipt header says “automated dossier integrity.” At bottom, print the ceiling in a bordered box: “not scientific acceptance, publication, or deployment readiness.” **Speaker notes** These exact counts are useful because they make fixture drift visible. The five outputs are release audit, claim traceability CSV, computed checksums, reviewer instructions, and run manifest. Zero machine errors and reproduced arithmetic justify routing forward. They do not complete the defense, independent scientific review, privacy/license review, human accessibility review, or external approval. **Teaching check** Ask why nineteen required files yield eighteen verified checksums. The checksum list does not recursively hash itself. **Alt text** An automated audit receipt lists exact L08 counts and ends at an unopened human-review door with a narrow claim ceiling. ## Slide 18 — The eight-minute presentation is an evidence index **On-screen text** 8-minute plan: 1 min question + decision 2 min design + authorized evidence 2 min primary + adverse/null result 2 min lineage + gates 1 min bounded conclusion + limitation **Visual specification** Show a horizontal eight-block strip grouped 1–2–2–2–1. Each group has a concise label and a matching geometric pattern. Beneath, draw an arrow from every group to a release artifact rather than to a decorative slide. Add “planning budget, not observed timing.” **Speaker notes** The talk should help a reviewer find evidence. Context is intentionally compressed. Results appear with uncertainty and negative evidence. The lineage section identifies the weakest edge and any non-redistributable boundary. The final minute states what the result does and does not support. These allocations are planning budgets; no W16 presentation has been recorded, timed, rehearsed, or piloted. **Teaching check** Ask which content should be cut first if context consumes three minutes. Learners should preserve result, lineage, gates, and limitations rather than squeezing them into unreadable slides. **Alt text** Eight time blocks form a 1–2–2–2–1 presentation plan, with each segment linked to evidence artifacts and labeled as unobserved planning time. ## Slide 19 — The seven-minute defense targets the weakest edge **On-screen text** Defense prompts: - What could reverse the conclusion? - What was excluded or changed? - Can another reviewer reproduce it? - Which gate remains open? - What evidence would remove the headline? **Visual specification** Place the traceability chain on a diagnostic bench. A magnifying lens enlarges the weakest dotted edge between result and claim. Five question tags hang from the lens. Author and reviewer are equal-size silhouettes on opposite sides; neither sits on a podium. A record button appears crossed out to show no recording claim. **Speaker notes** Defense is adversarial toward the evidence chain, not the person. Good answers route to artifacts, acknowledge uncertainty, and state a decision rule. “We will investigate” is incomplete without whether the claim remains releasable. The seven minutes are a planned questioning window, not proof that all projects can be reviewed adequately in that time. **Teaching check** Ask a learner to answer: “What would make you remove `CAP-001`?” A good answer names a data or computation mismatch, not embarrassment. **Alt text** A reviewer’s lens enlarges the weakest result-to-claim edge while five evidence questions surround equal-positioned author and reviewer figures. ## Slide 20 — Respond to reviewers with stable coordinates **On-screen text** Finding ID · challenged IDs · concern · decision · change · evidence · release version Decisions: accept · partly accept · decline with evidence · cannot resolve and block **Visual specification** Draw a four-row response table. Show sample row R1 challenging `CAP-003`, decision Accept, change Remove production claim, evidence `LIMIT-001`, version 0.2.0. Row R2 elevates `CAP-002`; row R3 regenerates checksums; row R4 leaves accessibility pending. Use icons plus text for decision states. **Speaker notes** A response note makes revision auditable. Declining a suggestion is legitimate when the evidence supports the decision, but authority or preference is not an evidentiary reason. “Cannot resolve” is also legitimate; it should weaken or block the affected claim. This teaching table models likely responses. The current L08 fixture contains no completed live defense or forty-eight-hour revision, so do not present the examples as observed outcomes. **Teaching check** Ask why “decline: outside scope” is insufficient unless the frozen question, estimand, or gate demonstrates that scope. **Alt text** A response-to-reviewers table links four findings to claim and artifact IDs, decisions, changes, evidence, and a new release version. ## Slide 21 — Revise within forty-eight hours, then regenerate **On-screen text** Defense → classify findings → revise → rerun → inspect → re-hash → release notes → re-review Due within 48 h. Preserve the prior version. **Visual specification** Create an eight-stage clockwise loop with a visible branch from inspect back to revise. A small archive box outside the loop holds v0.1. A new v0.2 bundle appears only after re-hash and release notes. Place a forty-eight-hour label on the loop without a clock face or exact timestamps. **Speaker notes** Revision is not a prose-only activity. Changing a claim can require updated figures, limitations, trace exports, manifests, reviewer responses, and checksums. If rerun results differ, stop and explain before release. Keep the prior version so a reviewer can compare. Forty-eight hours is the course response deadline, not evidence that a human defense has occurred in this package. **Teaching check** Ask what happens if re-hashing is done before regenerating a changed figure. The release becomes internally consistent only for the wrong dependency state; rerun order matters. **Alt text** An eight-stage revision loop runs from defense through re-review, preserves version 0.1, and emits version 0.2 only after inspection, hashing, and release notes. ## Slide 22 — Four reviewer findings reshape the teaching release **On-screen text** R1: causal production wording → remove R2: adverse secondary hidden → elevate R3: revision changed bytes → re-hash R4: accessibility human review pending → keep gate open **Visual specification** Show before and after claim cards side by side. Before has one broad headline and a tiny secondary note. After has bounded `CAP-001`, co-equal `CAP-002`, rejected `CAP-003`, and an amber accessibility gate. Four numbered arrows map each reviewer finding to one visible change. Add a footer: “authored worked example, not an observed defense.” **Speaker notes** This compact case demonstrates evidence-led revision. The story becomes less promotional and more faithful. The numerical result does not change, yet the public meaning, negative-result prominence, checksum identity, and release status do. The accessibility finding cannot be repaired by claiming that alt text exists; a named human review remains pending. **Teaching check** Ask which finding changes claim scope, which changes prominence, which changes identity, and which changes release readiness. **Alt text** Four reviewer findings transform a broad before card into bounded primary and adverse claims, a retained rejection, new hashes, and an open accessibility gate. ## Slide 23 — Stop conditions outrank presentation quality **On-screen text** Block on: - broken lineage or fabricated evidence; - unauthorized collection or access; - unresolved privacy, license, or safety failure; - inaccessible material output; - material overstatement; - unreproduced required result. **Visual specification** Draw six heavy stop bars across a path to release, each labeled with one condition. Above the path, a polished slide deck floats but cannot cross the bars. A side route labeled “revise, document, re-review” returns to the beginning. Use stop shapes, labels, and crosshatching rather than red alone. **Speaker notes** The capstone has noncompensatory failures. A beautiful eight-minute talk cannot offset fabricated provenance. A small descriptive result is acceptable if accurately bounded; a dramatic unsupported claim is not. When a required result cannot be reproduced, follow the preregistered rule: diagnose, preserve failure, weaken or remove the claim, and re-review. Never invent evidence to meet a deadline. **Teaching check** Present “all checks pass except the source license is unknown.” Learners must block public release rather than assign a high overall score. **Alt text** Six labeled stop bars block a polished deck from reaching release; a revision and re-review route loops back. ## Slide 24 — Ready for judgment, not guaranteed agreement **On-screen text** A trustworthy candidate lets an authorized skeptic: 1. traverse every public claim; 2. reproduce or delimit dependencies; 3. see negatives and deviations; 4. inspect hard gates; 5. observe evidence-led revision. Then—and only then—route to judgment. **Visual specification** Return to Slide 01’s stepping-stone path, now expanded into five parallel lanes matching the five criteria. Each lane reaches an amber forum labeled Human Judgment, not a trophy. One chair is empty to represent pending external review. Place `CAP-001`, `CAP-002`, and `CAP-003` along the relevant lanes. Add a small “no recording” label. **Speaker notes** Close by distinguishing readiness for judgment from success. The L08 machine route is complete locally; the human route is not. `CAP-001` can be questioned within its descriptive ceiling, `CAP-002` remains adverse and visible, and `CAP-003` remains rejected. Authorization, privacy/license, safety, accessibility, defense, independent review, and external approval still need named decisions. The objective is not invulnerability. It is a release whose claims, limits, and revisions remain inspectable under challenge. **Teaching check** Ask each learner for one sentence completing: “My capstone should be trusted only to the extent that…” The sentence must include an evidence route and a boundary. **Alt text** Five audit lanes—traceability, reproduction boundary, negative evidence, hard gates, and revision—lead to an amber human-judgment forum with pending seats, not to a trophy.