# W16 Transcript / No-Video Equivalent — Capstone synthesis and defense ## Status and use This is the planned transcript and no-video equivalent for W16, “Capstone synthesis and defense,” in the Validate phase. Its essential question is: **What would make the result worth trusting?** It uses the synthetic L08 fixture, `CAPSTONE-DEMO-001`, to show how a capstone moves from a collection of polished files to a bounded, inspectable evidence release. No W16 recording exists. This text must not be described as a transcript of a delivered lecture, a substitute for instructor judgment, or evidence that anyone completed the course. The chapter timings below are instructional planning budgets for a future forty-two-minute teaching session. They are not timestamps from recorded media. A local automated PASS means only that specified files, hashes, relationships, and deterministic calculations satisfied the package validator. It is not human review, independent reproduction, accessibility conformance, a pilot, grading calibration, external review, or a recording. It is also not scientific acceptance, legal authorization, publication approval, or deployment readiness. Those distinctions are part of the lesson, not footnotes to be hidden at the end. ## Planned chapter budget The planned total is forty-two minutes: - Chapter 1, four minutes: trust as an inspectable chain. - Chapter 2, four minutes: freeze the question, estimand, and claim boundaries. - Chapter 3, four minutes: reproduce the twelve-event arithmetic. - Chapter 4, four minutes: build the claim-to-artifact traceability matrix. - Chapter 5, four minutes: retain negative results, rejected claims, and deviations. - Chapter 6, four minutes: distinguish redistributable and non-redistributable dependencies. - Chapter 7, four minutes: apply authorization, privacy, license, safety, and accessibility gates. - Chapter 8, four minutes: use checksums, versions, and release notes. - Chapter 9, four minutes: design the eight-minute presentation. - Chapter 10, three minutes: conduct the seven-minute defense and answer reviewers. - Chapter 11, three minutes: revise within forty-eight hours and issue a bounded release. These allocations are planning constraints. An instructor may slow down for arithmetic, traceability, or gate decisions, but should not silently delete the defense or revision workflow to save time. The capstone itself is estimated at eighteen to thirty hours. The forty-two-minute lesson introduces the synthesis and defense logic; it does not compress the capstone work into one class meeting. ## Chapter 1 Let us begin with the essential question: what would make the result worth trusting? A visually impressive chart is not enough. A confident presenter is not enough. A repository with many files is not enough. Trust begins when a reviewer can move from a bounded claim to the artifacts that support it, inspect every important transformation, identify unresolved gates, and determine where the evidence stops. In W16, we will describe that movement as a chain: question, estimand, sources, query or data manifests, code, observations, analysis, figures, claims, limitations, and decisions. Every arrow matters. If a figure cannot be traced to a result, or a result cannot be traced to code and observations, the chain is broken. A beautiful object at the end cannot repair a missing edge near the beginning. The L08 fixture `CAPSTONE-DEMO-001` is deliberately small. It contains twelve synthetic events, a deterministic analysis, a figure, a claim ledger, limitations, negative results, protocol deviations, an artifact manifest, and checksums. Because the arithmetic is simple, we can focus on the evidence architecture. Small does not mean casual. In fact, the simplicity makes vague provenance harder to excuse. Think of trust as a reviewable property, not a mood. The reviewer should be able to ask, “Which exact observation supports this number? Which version of the code produced it? Which limitation narrows the sentence? Which gate remains open?” A trustworthy release answers those questions with stable identities and explicit boundaries. When it cannot answer, it reports the gap rather than decorating around it. ## Chapter 2 Before packaging artifacts, freeze what the study is trying to say. Begin with the approved question and the estimand. In the L08 teaching fixture, the primary arithmetic compares the proportion of synthetic outcomes in a treatment arm with the proportion in a control arm. That is a narrow descriptive estimand over twelve events. It is not a production estimate, not a population causal effect, and not proof that a GEO intervention improves visibility. Next, turn prose claims into stable ledger entries. `CAP-001` is the supported headline claim. It states that, in this fixture, four of six treatment events have the outcome, compared with two of six control events, producing a difference in proportions of one third. Its ceiling is descriptive arithmetic for this fixed synthetic dataset. The identifier lets every artifact refer to the same claim even when wording changes. `CAP-002` is supported but is not the headline result. It records an adverse secondary finding: citation correctness is four of six in treatment and five of six in control, a treatment-minus-control difference of negative one sixth. `CAP-003` is rejected. It represents the tempting but unsupported statement that the intervention improves production GEO visibility. Retaining the rejected claim shows where the evidence boundary was enforced. A frozen claim ledger does not forbid revision. It makes revision legible. If a reviewer demonstrates that the headline is overstated, preserve the original identity and record whether the wording was strengthened, weakened, or removed. Do not quietly rewrite history. A mature capstone lets the reviewer distinguish the planned question, the computed observation, the proposed interpretation, and the final bounded claim. ## Chapter 3 Now reproduce the arithmetic before discussing the design of the release. There are six control events. Two have `outcome = 1`, so the control outcome rate is two divided by six, or approximately `0.333333`. There are six treatment events. Four have `outcome = 1`, so the treatment rate is four divided by six, or approximately `0.666667`. Subtract control from treatment. The result is `0.666667 - 0.333333 = 0.333333`, allowing for displayed rounding. That value supports `CAP-001`, but only with the stated ceiling. The arithmetic does not establish that treatment caused the difference. The twelve rows are synthetic, and there is no production sampling frame, randomization argument, uncertainty model, or external-validity evidence. Repeat the process for citation correctness. Five of six control events are marked correct, giving approximately `0.833333`. Four of six treatment events are correct, giving approximately `0.666667`. The treatment-minus-control difference is approximately `-0.166667`. This is `CAP-002`, the adverse secondary result. It must remain visible even though it complicates a success narrative. Why perform such elementary calculations openly? Because independent checking begins with exact inputs and explicit operations. A reviewer should be able to count the twelve rows manually, run `analysis.py`, and compare the generated result with `expected_result.json`. Agreement across those routes increases confidence in arithmetic integrity. It does not independently validate the scientific interpretation, the authorization conditions, or the social consequences of using the result. The deterministic L08 validator can rerun the analysis and confirm that the generated values match the expected values. It can also confirm that the required files and declared hashes agree. If it prints PASS, the appropriate decision is “ready for human review,” not “research proven.” Automated consistency is one layer in a larger defense, and W16 is designed to keep those layers from being confused. There is another reason to do the manual calculation: it creates an analysis route that does not depend on trusting the same program being tested. If both the expected result and the new output came from one unnoticed coding error, exact textual agreement would be reassuring but circular. Counting event tiles, reducing the fractions, and comparing signs supplies a small heterogeneous check. For a larger capstone, the analogous check might be a second implementation, a hand-audited sample, an invariant, or an independently specified query. The check must be proportionate, documented, and distinct enough to expose plausible errors. Now imagine that one treatment row is duplicated. The script’s uniqueness rule should stop because event IDs must be unique. Imagine instead that a row silently changes arms. The six-per-arm rule may detect it, but if two rows swap arms the count rule alone will not. Identity hashes and a frozen manifest become relevant. These examples show why no single control carries the release: schema validation, identity, manual invariants, semantic comparison, and human design review cover different failure modes. ## Chapter 4 We can now build the claim-to-artifact traceability matrix. Put claims in rows. Use columns for source or data, code, result, figure, limitation, reviewer response, and checksum or release identity. A filled cell should contain a stable artifact identifier, not a vague phrase such as “see analysis folder.” The purpose is to make each evidentiary edge auditable. For `CAP-001`, the supporting path is `DATA-001`, `CODE-001`, `RESULT-001`, `FIG-001`, and `LIMIT-001`. `DATA-001` identifies the observations. `CODE-001` identifies the deterministic analysis. `RESULT-001` identifies the computed output. `FIG-001` identifies the primary visual. `LIMIT-001` constrains interpretation. If any one of these references resolves to the wrong file or an unverified version, the headline path is broken. For `CAP-002`, the path is `DATA-001`, `CODE-001`, `RESULT-001`, `NEG-001`, and `LIMIT-001`. The first three artifacts are shared with the primary analysis. `NEG-001` preserves the adverse result rather than hiding it in an appendix. `LIMIT-001` again prevents a descriptive contrast from being promoted into a broad causal conclusion. For rejected `CAP-003`, the path is `LIMIT-001` and `NEG-001`. This may initially look unusual because the row does not point to a supportive figure. That is the point. The matrix should represent reasons for rejection as carefully as reasons for acceptance. A reviewer can see that the production-visibility statement exceeds the outcome, setting, and design represented in the evidence. Traceability is more than file presence. A folder may contain every named artifact and still fail if the relationships are false. The figure might have been drawn from an earlier result. The claim ledger might cite the wrong artifact identifier. A release note might omit a post-review change. W16 therefore asks reviewers to inspect edges, not merely count nodes. A useful defense question is: “Show me the shortest complete path from the headline sentence to the original observations.” The presenter should be able to traverse that path without improvising. If the path stops, the claim stops. That rule applies regardless of the presenter's confidence, the importance of the result, or the quality of the visual design. ## Chapter 5 A capstone becomes more trustworthy when it preserves inconvenient evidence. In `CAPSTONE-DEMO-001`, `CAP-002` is a negative secondary result: citation correctness is lower in treatment by about `0.166667`. Retaining it demonstrates that the package is not merely a highlight reel. It also gives the reviewer information that could change the interpretation of the primary outcome. The package also retains `CAP-003` as rejected. A rejected claim is not dead clutter. It documents an interpretive branch that reviewers may otherwise suspect was silently removed. It says, in effect: “We considered this production-level conclusion, tested it against the available design and artifacts, and determined that the evidence could not support it.” Protocol deviations deserve the same treatment. The L08 fixture states that no post-plan row, outcome, or analysis change occurred. It also records that the validator was added later as a packaging control without changing the estimand. That is a modest deviation, but documenting it teaches the correct habit. A deviation log should say what changed, when, why, who authorized it, and whether it affected claims. Failed reproduction attempts must also be preserved. If a reviewer runs the code and initially obtains a different result because of a missing file, an environment mismatch, or an undisclosed dependency, do not delete that event after the problem is fixed. Record the failure, diagnosis, corrective action, rerun result, and remaining uncertainty. A clean final run does not erase evidence about release fragility. The governing rule is simple: negative results, exclusions, deviations, and failed reproductions are part of the evidence system. They should be indexed, versioned, and discussed in the defense. Removing them to create a smoother story lowers trust precisely when polish appears to increase. ## Chapter 6 The L08 fixture is fully redistributable. Its synthetic observations, scripts, manifests, and figure can be included in the teaching release under the declared terms. A reviewer can receive the same bytes, verify their hashes, and rerun the analysis locally. This is the clearest route to independent reproduction, but it is not the only situation a capstone may face. Real projects often depend on material that cannot lawfully be redistributed. Examples include licensed databases, private platform exports, confidential records, account-bound query results, or copyrighted collections. The correct response is not to smuggle restricted bytes into the release. It is also not to pretend the dependency does not exist. The release must define a verifiable non-redistributable dependency boundary. Imagine an artifact called `DEPENDENCY-NR-01`. Its boundary record should identify the issuer, product or collection name, version or retrieval date, lawful acquisition route, license and access conditions, exact locator, role in the analysis, and a cryptographic hash computed by an authorized holder. It should explain which reviewers may access it and which verification procedure they can perform without the material being published. Two outcomes must be distinguished. If an independent reviewer lawfully obtains the same dependency, matches the declared version and hash, and reruns the pipeline, independent reproduction may be possible. If the reviewer cannot obtain it, the reviewer may still verify selected outputs, provenance statements, or controlled-room evidence. That is output verification, not full independent reproduction, and the report must label it accordingly. A hash does not legalize possession. Retrieval instructions do not grant authorization. A screenshot does not substitute for a dataset, and an account credential must never be packaged as reproducibility material. The dependency boundary must protect the restricted object while exposing enough identity and procedure to make the limitation auditable. In the defense, say exactly which route applies. For the bundled L08 fixture, the route is redistributable deterministic reproduction. For a hypothetical `DEPENDENCY-NR-01`, the route could be conditional reproduction by an authorized reviewer or partial verification only. Never blend those statuses into the phrase “reproducible enough.” Precision about what could not be shared is itself evidence of research control. The dependency card also needs a failure rule. Suppose the licensed collection vendor updates records in place, so an authorized reviewer can reach the same product but not the version used by the author. A matching title is not a matching dependency. If the original bytes cannot be lawfully recovered, the reviewer records the mismatch, identifies which claims rely on it, and follows the frozen decision rule. The outcome may be “historical output verified but source-level reproduction unavailable,” a weakened claim, or a blocked release. It must not be relabeled as success because the vendor is reputable. Similarly, a controlled-room demonstration can provide useful evidence without becoming redistribution. The reviewer may observe a checksum match, execute approved commands, and export a signed verification record while leaving protected bytes in place. The release should state what the reviewer was permitted to see, which transformations were executed, whether the reviewer chose the commands independently, and what could not be inspected. This creates a verifiable boundary, not an invisible exception. It also prevents a later reader from assuming that every artifact in the traceability graph was public. ## Chapter 7 Traceability and reproduction do not override authorization. A technically repeatable analysis can still be impermissible. W16 therefore treats authorization, privacy, license, safety, and accessibility as hard gates. A broken gate blocks release even when the arithmetic, manifests, and checksums are perfect. The authorization gate asks whether collection, access, processing, review, and release were permitted by the relevant owners or authorities. The privacy gate asks whether personal or sensitive information exists, whether minimization and consent conditions were met, and whether outputs could enable re-identification. `CAPSTONE-DEMO-001` uses synthetic course-owned data and declares no personal data, but external release still requires a human check of that assertion. The license gate asks whether each source, dataset, code component, and visual may be redistributed and under what terms. The safety gate asks whether the method or result could produce material harm, enable abuse, or be deployed beyond its tested boundary. A course fixture may be low risk, but the habit of checking cannot be postponed until a later “real” project. The accessibility gate applies to figures, documents, tables, navigation, contrast, reading order, captions, and alternative descriptions. The L08 figure manifest provides an alt description, caption, and source reference, and the SVG contains structural accessibility metadata. Those machine-checkable features are useful. They do not establish accessibility conformance across every assistive technology or user context. Human inspection remains pending. These gates should appear in the matrix, README, defense slides, and release notes. A reviewer must be able to see whether each gate passed, failed, or remains pending, who decided, and what evidence supports the status. “Not applicable” is a decision requiring justification, not an empty escape hatch. If any gate is unresolved, the capstone should state the consequence. It may be restricted to classroom review, withheld from publication, released without a protected dependency, or returned for remediation. Trust increases when the release makes such limits unmistakable. ## Chapter 8 Checksums answer a narrow but important question: are these the same bytes that the release claims to identify? The L08 dossier includes `checksums.sha256`, and its validation workflow can generate a computed checksum list. When a reviewer hashes the files and obtains matching values, silent file changes become less likely. Checksums do not prove that a claim is true, that code is correct, or that collection was authorized. They bind an assertion to a file version. That binding becomes powerful when combined with stable artifact identifiers, manifests, and a reproducible run. It allows the reviewer to say, “This result came from this code and these observations in this release,” rather than merely, “I found a similarly named file.” Every post-review change should produce a new bounded release. Update affected artifacts, rerun the analysis where necessary, regenerate figures, refresh the claim ledger, calculate new checksums, and increment the version. Do not overwrite the prior release and keep the same identity. The ability to compare versions is part of the response to review. Release notes should identify what changed, why it changed, which claims and artifacts were affected, whether conclusions were strengthened, weakened, or removed, and which gates remain pending. They should also record the validation result and its ceiling. A useful sentence is: “Automated dossier integrity passed; independent scientific review, legal approval, human accessibility review, and external release approval remain separate.” The validator may confirm nineteen required dossier files, sixteen declared artifacts, three claim records, twelve claim-to-artifact edges, and the deterministic twelve-event calculation. Those counts are useful control evidence. They are not a quality score. A release with many correct files can still contain a weak question or an unjustified interpretation. ## Chapter 9 The capstone presentation is limited to eight minutes. That constraint forces synthesis. Begin with the question and the exact estimand. State the headline claim with its ceiling. Show the twelve-event arithmetic. Then show the claim-to-artifact matrix and walk one complete path for `CAP-001`. Next, surface the evidence that resists the headline. Present `CAP-002` as the adverse secondary result and `CAP-003` as the rejected production-level claim. Explain any deviations and failed reproduction attempts. Identify whether the release is fully redistributable or depends on a controlled boundary. Report the status of every hard gate. Conclude with what the result can and cannot support. A defensible final sentence for the fixture is: “Within this deterministic synthetic twelve-event dataset, the treatment arm has a one-third higher descriptive outcome proportion, while citation correctness is one-sixth lower; the design does not establish production GEO improvement.” That sentence is narrower than a promotional claim and stronger as research communication. Do not spend the eight minutes reading filenames or displaying decorative dashboards. The audience needs the logic of the evidence system. Visual design should help them locate claims, edges, breaks, and gates. If a figure cannot be explained, traced, and bounded during the allotted presentation, it should not carry the headline. ## Chapter 10 The presentation is followed by a seven-minute defense. The purpose is not to reward quick rhetoric. It is to test whether the presenter understands the chain, notices uncertainty, and changes claims when challenged. Reviewers should ask for artifact identities, arithmetic, exclusion logic, dependency boundaries, and gate evidence. Suppose Reviewer R1 asks why `CAP-001` is described as an improvement when the fixture has no causal design. The appropriate response is to accept the challenge, remove production or causal language, and bind the revised descriptive wording to `LIMIT-001`. Suppose Reviewer R2 asks why lower citation correctness was not prominent. The response is to elevate `CAP-002` in the presentation and ensure `NEG-001` remains part of the release. Reviewer R3 may ask whether checksums were regenerated after revisions. The response should cite the new release version and checksum manifest, not offer a verbal assurance. Reviewer R4 may ask whether the figure is accessible. If only automated structural checks exist, the honest answer is that human accessibility review remains pending. Do not convert partial evidence into a completed gate under pressure. Capture each issue in `review-findings.md` and answer it in `response-to-reviewers.md`. For every item, record the challenge, decision, affected claim and artifact identifiers, action taken, verification performed, and residual limitation. A response may accept, partially accept, or reject a suggestion, but rejection requires evidence-based reasoning. Let us model four concise exchanges. Reviewer R1 asks, “Your brief says improvement. Where is causal identification?” The author answers, “There is none; we accept the finding, remove that word, route the bounded sentence through `CAP-001`, and preserve `CAP-003` as rejected.” Reviewer R2 asks, “Why was citation correctness visually subordinate?” The author answers, “That prominence was misleading; we elevate `CAP-002` and `NEG-001` without adding an inference the fixture cannot support.” Reviewer R3 asks, “Which identity proves this is the post-revision bundle?” The author does not point to the date displayed in a document. The author names the new release version, the regenerated checksum manifest, and the release-note row linked to R1 and R2. Reviewer R4 asks, “Did a screen-reader user verify the SVG and surrounding document?” The author answers, “No. Structural controls are present, but human accessibility review remains pending and external release remains blocked.” These answers are stronger because they can concede limits without losing the evidence chain. A defense can also reveal a question that cannot be resolved within forty-eight hours. Perhaps the independent reviewer cannot obtain `DEPENDENCY-NR-01`, or the privacy reviewer requires a new authorization. The response category is not “we ran out of time.” It is “cannot resolve,” followed by the affected claim, the protective decision, and the condition for reopening review. The capstone can still demonstrate excellent research judgment by blocking its own headline. This text describes the intended defense procedure. It does not claim that an actual W16 defense, pilot, calibrated grading exercise, external review, or accessibility assessment has occurred. The local L08 response template likewise does not prove that a live eight-minute presentation or seven-minute defense was completed. ## Chapter 11 After the defense, revisions are due within forty-eight hours. Start with the highest-risk finding: a broken lineage edge, unauthorized source, privacy failure, license conflict, safety issue, accessibility blocker, or materially overstated claim. If evidence cannot repair the issue, weaken or remove the claim. Deadline pressure does not justify preserving a sentence that the release cannot support. Then regenerate the release as a coherent unit. Rerun the deterministic analysis, update affected results and figures, preserve negative findings and deviations, revise the claim ledger, update limitations, complete the response-to-reviewers, write release notes, and regenerate checksums. Verify that every artifact reference resolves to the correct version. Preserve the pre-defense release so that the revision remains auditable. The final decision is bounded. `CAPSTONE-DEMO-001` may reach “ready for human review” when its automated contract passes. A real capstone may progress only as far as its evidence and gates permit. Independent reproduction requires an independent person and an authorized route to the necessary materials. Accessibility conformance requires appropriate human evaluation. Publication and deployment require their own approvals. Return, finally, to the essential question. A result becomes worth trusting when its claim is narrow enough for the design, its numbers can be traced and checked, its inconvenient evidence remains visible, its dependencies are lawful and explicit, its gates are honestly reported, and its authors revise when review exposes a weakness. Trust is not a label attached after a successful run. It is the durable structure connecting evidence, limits, review, and accountable change.