research-document
Applied validation cases
Applied validation cases
These are structural prototypes, not imitations. Each case compares a selected architecture with an alternative and records distortion risk.
| Task | Contract and risk | Selected structure | Prototype | Alternative and revision |
|---|---|---|---|---|
| Executive decision brief | Decide; expert but time-poor; high consequence | Brown-like narrow question + RFC-like definitions | Decision: choose A or B by Friday. Known: … Uncertain: … Tradeoff: … Recommendation: A because … Reversal trigger: … | Darwin-like cumulative narrative builds understanding but delays the choice. Retain one causal paragraph in an appendix. |
| General-audience science explanation | Explain unfamiliar mechanism; mixed priors | Darwin familiar bridge + objections | “Start with the familiar process… The mapping holds because… It does not imply… Three observations discriminate it from…” | Euclidean derivation is auditable but assumes primitives. Add a compact dependency diagram after orientation. |
| High-stakes medical summary | Inform/decide under distress; consent and accuracy paramount | Narrow question, explicit uncertainty, action closure | “What was found. What it may mean. What it does not establish. Options now. Warning signs. Who to call.” | Descartes-like journey is too self-focused; Darwin-like analogy may mislead. Use only validated patient-tested metaphors. |
| Legal argument | Persuade authorized decision-maker under precedent | Issue → rule → record → adverse authority → application → requested holding | “The question is… The governing rule… The strongest contrary authority… It differs because… Therefore…” | Narrative-first can reveal stakes but may weaken doctrinal traceability. Keep bounded facts before issue, not instead of rule. |
| Technical instruction | Correct action; error prevention; observable outcome | Euclidean dependencies + RFC terms, but task-linear | “Before you begin… 1… Verify… If X, stop… Success looks like…” | Pure RFC reference is precise but poor for first-time completion. Link canonical terms from steps. |
| Change announcement | Explain/coordinate; trust risk and power asymmetry | Descartes problem-before-method + Brown-like explicit decision | “What changes; why now; evidence considered; what does not change; impact by group; support; feedback and revision channel.” | Categorical institutional tone alone hides participation. Add uncertainties and recourse. |
| Persuasive public essay | Advocate truthfully amid resistance | Darwin cases/objections + transparent values | “Shared problem; concrete case; causal claim; strongest objection; evidence boundary; proposed action.” | Brown-like closure is memorable but communicator lacks adjudicative authority. Use invitation and evidence, not a holding voice. |
| UI warning | Prevent immediate error; seconds of attention | Euclidean condition/action/consequence, minimal | “Deleting this key permanently disconnects 14 devices. Export backup / Cancel / Delete key.” | Narrative and cumulative explanation fail the time constraint. Progressive disclosure can explain details. |
| Research handoff | Preserve state and enable continuation | RFC navigation + Euclidean dependencies | “Objective; current claim/evidence; exact artifacts; commands; unresolved discriminator; next action; stop condition.” | Personal narrative preserves context but obscures current state. Link journal for chronology. |
| Sensitive interpersonal message | Recognize/repair; low power or explicit asymmetry | Relational accuracy first; no writer persona | “I understand X affected you as Y. I’m sorry for Z. I will do A. You do not owe an immediate response.” | Judicial closure sounds coercive; explanatory defense centers the sender. Ask, listen, and leave autonomy. |
Framework revision produced by cases
The medical, warning, and interpersonal cases showed that writer selection is sometimes the wrong first abstraction. A safety or relational protocol can dominate writer-derived architecture. The framework now treats risk, power, accessibility, and outcome measures as gates before model selection.