9.8 KiB
Presentation Evaluation And Closing Design
Purpose
Scenes 12 and 13 close the defense with evidence, claim boundaries, and a clear statement of the thesis contribution. They must not read as a model leaderboard or as a generic list of limitations. The ending should move from bounded evaluation evidence, through supported future work, back to the stable planner/runtime boundary, and finally into examiner questions.
This design replaces the current three-number evaluation row and generic conclusion narrative with two dedicated presentation scenes.
Design Principles
- Treat the 36-trial campaign as bounded longitudinal engineering evidence, not a controlled model comparison.
- Prefer one continuous visual artifact per scene over a collection of cards.
- Preserve exact thesis counts and non-claims.
- Keep chat hidden. The final scenes belong to the thesis argument and its evidence, not to the prepared demo agent.
- Reuse the editorial canvas, stage caption, discussion panel, evidence text, and existing presentation tokens.
- Keep all primary content readable at 1280 by 720 and 1024 by 768, including browser zoom with internal stage scrolling where necessary.
- Use motion only to shift emphasis or reveal a new evidence layer. Reduced motion must preserve the complete meaning.
Scene 13: Evaluation Evidence Board
Scene 13 uses one persistent evaluation board. Each beat changes emphasis inside the same board so the audience does not have to reconstruct the campaign from unrelated layouts.
Beat: cohort
The cohort equation is the dominant artifact:
2 challenges x 2 hosted models x 3 instruction profiles x n=3 = 36 trials
An audited outcome rail beneath the equation shows the exact counts:
- 27 clean product-path passes
- 8 invalid evaluation samples
- 1 failure
The interface must say audited outcomes, not success rate. It must not show
percentages or rank models or profiles.
Beat: validity
The cohort equation remains visible as context but recedes. The dominant visual becomes an audit reconciliation:
automatic result manual evidence audit
7 automatic successes -> invalid as clean evidence
3 automatic failures -> accepted from saved evidence
The board states the claim boundary verbatim:
Bounded longitudinal engineering evidence, not a controlled model comparison.
Supporting copy explains that manual audit separates task completion from valid product-surface evidence. It must not imply that automatic grading is generally unreliable; the claim is limited to this campaign.
Beat: findings
The board resolves into observed engineering findings. Use a connected ledger or annotated rail, not six generic cards. Display these findings:
- Schema discovery
- Repair hints
- Binding commands
- Output schemas
- Shell assumptions
- Source contamination
The framing is UX gaps exposed by trials. The scene may say that individual
trials revealed these gaps. It must not claim statistical significance,
controlled profile effects, token savings, model superiority, or general model
performance.
Evidence And Discussion
The scene keeps the existing evidence pointer to the thesis Evaluation and
Appendix C. The existing evaluation-validity discussion branch remains
available from the scene discussion rail.
Scene 13: Boundary, Future Layers, And Questions
Scene 13 uses a stable contribution line as its central visual:
External planner -> typed workflow substrate -> deterministic runtime
|
v
persisted, inspectable evidence
The substrate remains in the center across all beats. Surrounding content may appear or recede, but the center must not move enough to make the audience relearn the diagram.
Beat: limits
The contribution line dominates. Three explicit non-claims appear below it:
- Not a production sandbox
- Not a scheduler
- Not a broad agent benchmark
The copy must not imply that these are partially delivered features.
Beat: future
Five supported future-work branches appear around the stable substrate:
- External agent interface and planner loop
- Production security and credentials
- Scheduling and hosted operations
- Stronger controlled evaluation
- Runtime and provider expansion
Each branch uses a short label and one concrete example. The branches are surrounding layers, not requirements that belong inside the deterministic core. The visual must preserve the planner/runtime boundary.
Beat: conclusion
The future branches recede. The stable contribution line becomes the only dominant artifact. The closing statement is:
Planner proposes; runtime executes.
Supporting copy may state that the typed workflow substrate makes reusable agent-operated automation inspectable. It must not describe the submitted prototype as a bundled autonomous agent.
Beat: questions
Arrow Right from the conclusion enters a full-stage defense discussion index. This is the transition from the prepared talk to examiner questions.
The index gathers every existing discussion branch and groups it by defense topic:
- Thesis contribution
- Positioning and related systems
- Runtime and lifecycle
- Authoring and validation
- Demo integrity
- Evaluation
- Production readiness and future work
Each entry uses the branch's existing title and opens the existing
DiscussionPanel. Closing a discussion returns to the Questions beat, not to
the branch's original parent scene. The index must derive its entries from the
canonical discussion-branch catalog; it must not duplicate branch titles or
answers in a second data structure.
The Questions beat is a reusable presentation component but is only exposed as the final Scene 13 beat in this slice. A future presenter shortcut may open it globally without changing the component.
Component Boundaries
EvaluationEvidenceScene
Owns the evaluation-board composition and beat-specific emphasis. Its facts come from a small typed presentation model colocated with the scene rather than being embedded across JSX and CSS selectors.
ConclusionScene
Owns the stable contribution diagram, non-claim strip, future-work branches, and beat-specific emphasis. It does not own discussion navigation.
DefenseDiscussionIndex
Receives the canonical discussion branch definitions and an openDiscussion
callback. It groups branches through an explicit branch-to-topic projection.
The projection must be exhaustive so a newly added discussion branch cannot
silently disappear from the index.
SceneBody
Routes the evaluation and conclusion views to their dedicated components.
It passes openDiscussion to the discussion index. It should not gain the new
scene-specific rendering logic itself.
Interaction And Motion
- Arrow Left and Arrow Right continue to navigate beats through the existing presentation reducer and hash routes.
- Selecting a discussion index entry uses the existing discussion route and panel.
- Escape closes the discussion panel and returns to the Questions beat.
- Cohort, validity, and findings transitions may fade or shift emphasis over 150 to 250 milliseconds.
- Future branches may reveal progressively, then recede for the conclusion.
- Motion-disabled and
prefers-reduced-motionstates show the final layout without transitional transforms. - No autoplay is required for these scenes.
Accessibility
- The evaluation board has a labelled group and semantic lists for outcomes and findings.
- Counts and categories remain understandable without color.
- The contribution line and future branches use text labels in addition to connector styling.
- The discussion index is a labelled navigation region containing standard buttons or links with visible focus treatment.
- The active beat changes visual emphasis but never removes the factual context required to understand the scene.
Responsive Behavior
- At wide canvas sizes, the evaluation equation and audit reconciliation may use horizontal flow.
- At narrower 4:3 canvas sizes, the equation wraps into two rows and the audit reconciliation stacks without changing reading order.
- The future-work map may tighten around the substrate at 4:3 but must not overlap connectors or discussion controls.
- The discussion index uses two columns when space permits and one internally scrollable column when browser zoom or canvas width requires it.
- Scrollbars may be visually quiet, but all content must remain reachable by wheel, touchpad, and keyboard.
Testing And Verification
Implementation follows test-driven development.
Automated coverage must include:
- exact Scene 12 cohort and outcome counts;
- absence of percentages and model-ranking language;
- validity reconciliation counts and claim-boundary wording;
- all six observed UX findings;
- all three Scene 13 non-claims;
- all five supported future-work branches;
- stable closing contribution wording;
- the fourth
questionsbeat and direct hash route; - exhaustive inclusion of every canonical discussion branch in the index;
- discussion open and close returning to the Questions beat;
- semantic labels for the evaluation board, contribution diagram, and discussion index;
- responsive class or data-attribute contracts used by the CSS composition.
Verification includes focused presentation tests, the full console presentation suite, console typecheck, production build, and Playwright screenshots at 1280 by 720 and 1024 by 768 for every new beat. Browser review must check clipping, reading order, projector contrast, and the transition from conclusion into Questions.
Explicit Non-Goals
- No new charting dependency.
- No model leaderboard or statistical comparison.
- No changes to thesis evaluation data.
- No new discussion content or rewritten Q&A answers.
- No global presenter shortcut for the discussion index.
- No live LLM or chat work.
- No redesign of Scenes 1 through 12.