# W12 Slide Production Script — Optimization under Competing Queries **Production status:** Text-first specification; no deck, recording, or classroom pilot exists. **Visual contract:** All diagrams are newly designed for W12. Do not import, trace, or restyle paper figures, blog graphics, platform screenshots, or attached-TeX artwork. Use labels, geometry, patterns, and position in addition to color. Every quantitative graphic requires the equivalent table in speaker materials. ## Slide 01 — One page, five intents, no universal win **On-screen text** `Which gain survives a portfolio of intents?` Definition · Compare · Select · Verify · Counterfactual One revision may help and harm at once. **Visual specification** Draw one central document silhouette feeding five unequal, labeled lanes. Each lane terminates in a different icon: dictionary, balance, decision fork, magnifier, and reversed arrow. Two lanes rise, one stays level, and two bend downward. Use solid, dotted, and striped paths so direction remains legible in grayscale. Original vector artwork only. **Speaker notes** Open with the essential question, not a tactic. A page serves multiple information needs, and those needs can prefer incompatible revisions. W12 is a governed selection problem. We will preserve adverse cells and make the policy behind “overall gain” inspectable. The goal is not to discover a universally effective layout. **Teaching check** Ask: “If four lanes rise and one falls, what information is still missing before selection?” Expect population, weights, loss magnitude, hard constraints, uncertainty, and decision rule. **Alt text** A single document sends five differently styled paths toward five intent icons; some paths rise and others fall, illustrating competing query effects. ## Slide 02 — The decision object comes before the score **On-screen text** Decision = candidate set + portfolio + objectives + gates + budget + rule Valid actions: release one · retain control · inconclusive **Visual specification** Create a horizontal assembly line of six labeled boxes, ending in a three-way action switch. Place an empty score badge above the line with a lock icon, showing that scoring cannot begin until the six boxes are complete. Number every box and repeat the sequence in a text footer. **Speaker notes** Define optimization as selection over frozen, permitted source states. Candidate identity includes component, claim/evidence scope, owner, cost, authorization, and rollback. A no-change action is essential. If a class can only select a revision, it is not testing whether revision is warranted. **Teaching check** Learners rewrite “optimize this content” as one decision sentence containing all six objects and a no-release action. **Alt text** Six prerequisite boxes precede a three-way switch for release, control, or inconclusive; scoring remains locked until the specification is complete. ## Slide 03 — Freeze the five-stratum portfolio **On-screen text** Definition 0.30 · Compare 0.25 · Select 0.20 Verify 0.15 · Counterfactual 0.10 Total = 1.00; weights frozen before outcomes. **Visual specification** Use a five-row ledger, not a pie chart. A proportional bar appears beside each numerical weight, with a final checksum row showing `1.00`. Add a timestamp seal reading `policy input`. Put a crossed-out hand beside an after-results weight slider. **Speaker notes** Weights describe a policy allocation, not a natural truth or verified user distribution. Every stratum needs a definition and provenance. W06 supplies the sampling-frame discipline: paraphrases stay with their intent cluster, desired-answer prompts are excluded, and a convenience list is not silently called a population. **Teaching check** Ask who has authority to set the weights and what evidence would justify changing them. A change after outcomes must be a separately labeled sensitivity analysis. **Alt text** A five-row weight ledger sums to one; a policy seal appears before outcomes, while post-result weight adjustment is crossed out. ## Slide 04 — Vector first, scalar later **On-screen text** `J = [g_definition, g_compare, g_select, g_verify, g_counterfactual, −cost]` Higher is better in every displayed direction. Keep the trade-off pattern visible. **Visual specification** Show a six-cell vector as aligned columns. Place upward arrows over five gain columns and a downward raw-cost arrow under the last, then show why `−cost` aligns the display direction. Use no radar chart because area can obscure exact comparisons. **Speaker notes** An objective vector prevents an early average from hiding losers. Units and scales still matter. The W12 gains share a fictional normalized scale, while cost is an authored point budget. This vector is for comparison among feasible candidates; it does not contain factuality, safety, or accessibility rewards. **Teaching check** Give a raw vector with positive cost and ask learners to mark every preferred direction before asserting dominance. **Alt text** Six aligned coordinates display five gains to maximize and negated cost to maximize, with directional arrows above each column. ## Slide 05 — Hard gates are not tiny rewards **On-screen text** Must pass: factual · legal · authorized · safety · privacy · accessibility Budget ≤ 8 Filter first. Score survivors second. **Visual specification** Design seven vertical turnstiles before a scoring table. A candidate with a large star is stopped at the factual turnstile. No bypass arrow exists. Use thick double borders for gates and thin single borders for reward cells. **Speaker notes** Non-compensable means exactly that: no portfolio gain can purchase permission to be false, unauthorized, unsafe, unlawful, privacy-violating, or inaccessible in its essential information. Context-specific legal and accessibility review remains human work. The fixture only freezes pass/fail fields. **Teaching check** Ask learners to repair `0.8 visibility + 0.1 factuality + 0.1 accessibility`. The correction is a constrained problem, not different weights. **Alt text** Seven mandatory turnstiles stand before the scoring area; a high-scoring candidate is blocked by factual-integrity failure with no bypass. ## Slide 06 — Seven frozen candidates **On-screen text** C0 control · C1 proximity · C2 answer-first module C3 bundle · C4 persuasive append · C5 decorative headings · C6 anchors **Visual specification** Arrange seven index cards in a two-row tray. Each card shows ID, one component icon, cost badge, and rollback availability. C3 visibly links C1 and C2 with a brace. C4 carries a warning triangle but its outcomes remain concealed at this stage. **Speaker notes** Candidate identity is frozen before outcomes. C1 shares the name of the L05 layout factor, but the W12 outcome values are authored separately. C3 tests a bundle. C4 is retained to demonstrate why gate failures must remain visible rather than disappearing from the optimization trace. **Teaching check** Ask which candidates permit a single-component ablation. Expected: compare C3 with C1 and C2, while noting the bundle may interact. **Alt text** Seven labeled candidate cards show component and cost; a brace marks C3 as the combination of C1 and C2. ## Slide 07 — Reveal gates before outcomes **On-screen text** Hard-feasible: C0, C1, C2, C5, C6 C3: cost 9 > 8 C4: factual + safety fail **Visual specification** Create a candidate-by-gate matrix with check marks and patterned failure cells. The outcome columns remain covered by a gray curtain labeled `not yet inspected`. Use an explicit bottom sentence: “A failed gate cannot be compensated.” **Speaker notes** The reveal order prevents raw outcome excitement from weakening the rules. C3 fails budget despite passing integrity fields. C4 fails factual and safety gates. Keep both in the audit trace, but exclude both from weighted selection and Pareto comparison among feasible options. **Teaching check** Ask whether C3 could re-enter if its weighted mean is twice C1’s. The answer is no unless the authorized budget policy is changed in a new analysis. **Alt text** A gate matrix admits five candidates and excludes C3 for cost and C4 for factual and safety failure; outcomes remain hidden. ## Slide 08 — The signed query-conflict heatmap **On-screen text** Positive, zero, and negative cells all remain. C1 counterfactual = −0.01 C2 select = −0.02 **Visual specification** Draw the seven-by-five matrix as a heatmap with plus signs, zeros, and minus signs printed inside every cell. Magnitude uses lightness, while border hatching marks infeasible C3 and C4 rows. A complete numeric table appears in the handout. **Speaker notes** Do not crop the inconvenient columns. C1’s counterfactual loss and C2’s selection loss are part of their identities. C4’s large comparison and selection values remain visible only to show how an inadmissible revision can look attractive on a proxy. **Teaching check** Ask learners to name one candidate with a conflicting outcome and one with no negative strata. Expect C1 or C2 for conflict, C6 for all positive changes. **Alt text** A seven-row, five-column signed matrix prints every positive, zero, and negative value; infeasible rows are crosshatched. ## Slide 09 — One dot product in full **On-screen text** C1 = `.30(.05)+.25(.04)+.20(.02)+.15(.06)+.10(−.01)` `= .015+.010+.004+.009−.001 = .0370` **Visual specification** Use five vertical multiplication tiles feeding a sum bar. Align decimals and retain four places. The negative final contribution is drawn below the baseline with a striped fill. Include the weights checksum beside the formula. **Speaker notes** This is transparent aggregation, not proof of value. The arithmetic is deterministic because weights and cells are frozen. Never round each product prematurely. The single mean is reported beside, not instead of, the five stratum outcomes. **Teaching check** Have learners reproduce C6’s mean: `0.0200`. Require the five products, not just a calculator result. **Alt text** Five weighted products for C1 feed a sum of 0.0370; the counterfactual contribution is visibly negative. ## Slide 10 — Raw score leaders can be inadmissible **On-screen text** C3 raw mean = 0.0565 → budget fail C4 raw mean = 0.0410 → factual and safety fail C1 = 0.0370 → eligible for later rules **Visual specification** Show a descending score list behind a transparent gate overlay. C3 and C4 appear at the top but are stamped `excluded`, not moved lower. C1 is the first row with an open path forward. Avoid winner podium imagery. **Speaker notes** Sorting first and filtering later is a dangerous implementation error. A correct dot product can still produce the wrong decision if infeasible candidates enter selection. “Excluded” is not the same as “low performing.” Preserve the exact reason. **Teaching check** Ask learners to describe C4 without saying “second best.” Expected: high raw authored mean, inadmissible due to factual and safety failures. **Alt text** A raw-score list places C3 and C4 above C1, but gate stamps exclude the first two before selection. ## Slide 11 — Add the loss floor **On-screen text** Portfolio non-inferiority floor: every stratum ≥ −0.02 C1 minimum −0.01 · C2 minimum −0.02 · C6 minimum 0.01 **Visual specification** Draw a horizontal threshold line at `−0.02` with one vertical minimum marker per feasible candidate. Use candidate labels on the markers and patterns to distinguish equality from strict margin. Add a note that the threshold is policy, not a statistical constant. **Speaker notes** A mean does not protect a material minority intent. The floor is an outcome guardrail applied after integrity gates. C2 sits exactly on it. If the rule required strict greater-than, C2 would fail; the operator therefore belongs in the preregistration. **Teaching check** Ask whether the floor can replace reporting each stratum. No: it only defines eligibility; the full pattern remains necessary. **Alt text** Candidate minimum outcomes are placed against a minus 0.02 threshold; C2 equals it, C1 exceeds it slightly, and C6 is positive. ## Slide 12 — Authored intervals and a meaningful threshold **On-screen text** Minimum worthwhile lower endpoint = 0.015 C1 lower = 0.019 ✓ · C2 = 0.013 · C5 = −0.0075 · C6 = 0.003 **Visual specification** Create a compact interval forest with a vertical zero line and a second patterned line at `0.015`. Label both endpoints numerically. Put `authored teaching intervals—not sampling estimates` in a bordered banner. **Speaker notes** The interval arithmetic is reproducible, but it has no empirical coverage because no observations generated it. C5 crosses zero and is null-compatible. C2 and C6 remain nondominated alternatives, yet they miss the primary meaningful-lower-bound rule. **Teaching check** Ask why “C5 has no effect” is too strong. Its authored interval includes zero and positive values; no empirical test was run. **Alt text** Four horizontal intervals are compared with zero and 0.015 lines; only C1 has a lower endpoint beyond the meaningful threshold. ## Slide 13 — Dominance is a coordinate proof **On-screen text** C1 dominates C5: all five gains ≥ · cost 3 < 4 · at least one strict Compare only hard-feasible candidates. **Visual specification** Use a six-row proof table with C1 and C5 side by side, preferred direction, and result symbols. Every row has an explicit inequality. A final logic bracket joins all rows to `dominates`. No scatterplot substitutes for the proof. **Speaker notes** Dominance depends on the declared coordinates and directions. C1 equals C5 on the negative counterfactual cell but exceeds it elsewhere and costs less. C6 is another witness. A failed hard gate is not a dominance relation, so do not use C3 or C4 in this proof set. **Teaching check** Give C1 and C6 and ask why neither dominates. C6 is cheaper and better on counterfactual; C1 is better on four gains. **Alt text** A row-by-row inequality table proves C1 is never worse than C5 on gains and costs less, establishing dominance. ## Slide 14 — The feasible Pareto frontier **On-screen text** Frontier = `{C0, C1, C2, C6}` C5 dominated C3/C4 infeasible, not frontier candidates **Visual specification** Draw an original two-panel visual: a simplified gain-versus-cost projection on the left and a text frontier table on the right. Connect frontier points with a solid line only as a guide. Crosshatch C3/C4 outside a bold feasibility frame. Footnote that the proof uses all six vector coordinates. **Speaker notes** A two-dimensional projection cannot prove six-dimensional dominance, hence the companion table. C0 remains on the frontier because it has zero cost. Frontier membership is not approval or universal optimality; it only means no feasible candidate is unambiguously better under this vector. **Teaching check** Ask why a zero-gain control can be nondominated. Any positive-gain revision costs more, so none is at least as good on cost. **Alt text** A gain-cost projection and table list C0, C1, C2, and C6 as feasible nondominated options; C5 is dominated and two rows are infeasible. ## Slide 15 — The primary decision rule is executable **On-screen text** 1 gates + budget → 2 loss floor → 3 lower bound 4 maximize weighted mean → 5 tie rule No survivor? retain C0. **Visual specification** Build a five-stage filter funnel ending in two action bins: `C1 for human local review` and `control/inconclusive`. At each stage print the surviving candidate IDs. Use rectangular filters, not an opaque automation icon. **Speaker notes** The sequence is part of the estimand. Tie within `0.002`: lower cost, then fewer components, then control. If a later gate fails, rollback occurs without rescoring. Someone who did not author the preferred answer should be able to execute this rule. **Teaching check** Ask what happens if C1’s lower endpoint changes to `0.014`. No candidate meets all primary rules, so retain control and report inconclusive. **Alt text** A numbered filter funnel applies gates, loss floor, meaningful bound, mean, and tie rule, selecting C1 or retaining control if empty. ## Slide 16 — Weighted choice and maximin disagree **On-screen text** Primary weighted policy → C1 Maximin sensitivity policy → C6 Same matrix; different governance. **Visual specification** Split the canvas into two labeled policy lenses aimed at the same immutable matrix. The weighted lens highlights the dot product; the maximin lens highlights the lowest cell in each row. Both lead to separate decision cards, joined by a `policy sensitivity` bracket. **Speaker notes** Maximin chooses the largest worst-stratum outcome. C6’s minimum is `0.01`, compared with C1’s `−0.01`. The disagreement is not an arithmetic defect. It exposes whose losses the primary weights tolerate. Both results belong in the report. **Teaching check** Ask learners to defend one policy without claiming it is natural. The defense must name stakeholder authority and loss allocation. **Alt text** Two policy lenses read one matrix differently: weighted aggregation selects C1, while the best worst-stratum rule selects C6. ## Slide 17 — A null and a conflict are valid outcomes **On-screen text** C5: feasible, dominated, interval crosses zero C1: selected, but counterfactual −0.01 Do not erase nulls or losers. **Visual specification** Show two evidence cards. The C5 card carries three independent tags: feasible, dominated, null-compatible. The C1 card carries selected and harmed-stratum tags. Use separate shapes for status types so they cannot collapse into one traffic-light label. **Speaker notes** Optimization reporting needs more than winner and loser. Feasibility, dominance, uncertainty, and policy selection are different properties. A null can indicate a weak factor; a conflict can indicate real portfolio tension. Neither licenses post-hoc factor addition or stratum deletion. **Teaching check** Learners complete: “C2 is feasible and nondominated, but…” Expected: it misses the primary meaningful lower-endpoint rule and reaches the selection loss floor. **Alt text** Two multi-tag cards distinguish C5’s feasible, dominated, null-compatible status from C1’s selected but counterfactual-harming status. ## Slide 18 — Bundle ablation, two removals **On-screen text** C3−C2 = proximity conditional contribution = 0.0235 C3−C1 = module conditional contribution = 0.0195 C3 still fails budget. **Visual specification** Draw C3 as two interlocking blocks. In the top row remove the proximity block and land on C2; in the bottom remove the module and land on C1. Print subtraction traces beside both. A locked budget gate remains below the bundle. **Speaker notes** An ablation asks a conditional component question. The bundle’s budget failure cannot be offset by either contribution. These are authored differences, not causal estimates from assignment. A real factorial study would require support for interactions, measurement, and uncertainty at the correct unit. **Teaching check** Ask whether the larger `0.0235` proves proximity is universally more important. No; it is conditional on this bundle, matrix, policy, and authored construction. **Alt text** Two diagrams remove one block at a time from C3, yielding C2 and C1 with weighted differences, while the bundle remains over budget. ## Slide 19 — Negative interaction is not an inconvenience **On-screen text** Bundle − C1 − C2 = `−0.0135` Interaction vector = `(−.02, 0, −.01, −.03, −.01)` Report redundancy or interference; do not invent mechanism. **Visual specification** Use an additive expectation bar and an observed-bundle bar. The gap is a hatched negative segment labeled `descriptive interaction`. Below, print five signed cells matching the vector. Avoid gears or causal arrows. **Speaker notes** Components need not add. They may compete for attention, budget, or measurement space. The frozen values permit a descriptive interaction calculation but identify no external process. Preserve the gap and redesign a smaller bundle rather than editing expected values. **Teaching check** Ask which stratum has the largest negative interaction. Verify is `−0.03`. **Alt text** An additive expectation exceeds the bundle by 0.0135; five signed cells show the descriptive interaction is nonpositive in every stratum. ## Slide 20 — L05 proves local equivalence, not outcomes **On-screen text** `INT-001 · evidence_layout_proximity` `3 locked claims · 1 changed factor · 0 error(s)` No system response or causal visibility effect is measured. **Visual specification** Draw two local HTML page outlines connected to a deterministic audit box. The box outputs claim lock, source lock, diff, and rollback. A thick boundary stops before a separate W12 matrix card labeled `authored teaching input`. **Speaker notes** L05 aligns with C1’s factor name and verifies exact local properties. It does not generate query gains. This boundary prevents a software-equivalence PASS from being narrated as user comprehension, retrieval, ranking, citation, or production evidence. **Teaching check** Ask which evidence would be required to claim a human-comprehension effect. Expect an authorized human study with a defined population, assignment, measures, and analysis. **Alt text** A local control-treatment audit outputs equivalence and rollback records, then stops at a boundary before the separately authored W12 outcome matrix. ## Slide 21 — Paper routes are bounded, not a tactic menu **On-screen text** Core: PAPER-03 · PAPER-31 Extend: PAPER-04 · PAPER-27 · PAPER-30 Extract design objects; retain benchmark and system ceilings. **Visual specification** Create five source-route cards with two columns: `use here` and `cannot establish`. Use identical card size to avoid implying a leaderboard. PAPER-03 maps to query conflict; PAPER-31 to feature objectives; extensions map to output control critique, confidence/agent critique, and adaptive proposal audit. **Speaker notes** Do not import paper figures or prose. Each route informs a methodological question. Reported results stay attached to the paper’s query set, candidate construction, metrics, systems, and versions. No paper validates HarborGuide or supplies universal weights. **Teaching check** Ask learners to state one ceiling for PAPER-27: a proposed confidence or routing model does not reveal hidden state in a named closed platform. **Alt text** Five equal source cards pair a bounded methodological use with a cannot-establish boundary for each PAPER identifier. ## Slide 22 — Preregister deviations and rollback **On-screen text** Freeze: weights · gates · budget · candidates · matrix version · rule Deviation = new analysis, never overwritten primary Gate failure after choice → restore C0 **Visual specification** Design an append-only ledger with three rows: preregistration, decision, and later deviation. A rollback arrow points to a hash-addressed C0 vault. Show an edit pencil adding a row, not erasing an earlier row. **Speaker notes** Post-hoc changes are sometimes necessary, but transparency requires a new version and rationale. Record who changed what, when, why, and how the selection changed. Rollback is triggered by integrity and protocol failures without bargaining against the score. **Teaching check** Scenario: decision-maker doubles compare weight after seeing C1. Learners label it policy sensitivity or deviation, recompute, and preserve the original. **Alt text** An append-only ledger keeps the original policy and later deviation; a rollback arrow restores a hash-locked control without erasing history. ## Slide 23 — Deterministic reproduction has a boundary **On-screen text** Validator reproduces: hashes · gates · means · intervals · frontier · rules · L05 It does not reproduce an empirical effect, accessibility, legality, or learning. **Visual specification** Create a two-column contract. The left column has seven checkable software objects; the right has six human or empirical questions behind an open review gate. A heavy vertical rule separates them. Put `fail closed` above the left column. **Speaker notes** The standard-library validator rejects missing files, unexpected files, malformed identities, changed narrative hashes, and decision drift. It reruns L05 in a temporary directory. A PASS proves only that the declared local calculation and package were preserved. **Teaching check** Ask whether a deterministic PASS establishes WCAG conformance or a platform effect. Both answers are no; dedicated evidence is still required. **Alt text** A contract separates reproducible package and arithmetic checks from unresolved human, legal, accessibility, and empirical review questions. ## Slide 24 — The bounded W12 decision **On-screen text** C1 selected by primary policy; C6 by maximin. C1 counterfactual loss = −0.01. C5 dominated/null-compatible; C3/C4 inadmissible. Retain C0 rollback. **Visual specification** End with a signed decision card containing five fields: primary choice, sensitivity choice, harmed stratum, exclusions, and evidence ceiling. Place the C0 hash vault beneath it. No trophy, upward-only arrow, brand mark, or platform logo appears. **Speaker notes** Read the conclusion without embellishment. In this frozen synthetic matrix, C1 survives the weighted rules and C6 survives maximin. That is a policy-sensitive local arithmetic result. It does not establish a universal content strategy, ranking or citation guarantee, user benefit, or production effect. Human review remains pending. **Teaching check** Exit prompt: state the selected candidate, one loser, one policy sensitivity, one rollback trigger, and the strongest claim the evidence cannot support. **Alt text** A five-field decision card records C1, C6, the counterfactual loss, excluded candidates, and the synthetic evidence ceiling above a control rollback vault.