Target duration: approximately nine minutes including visual pauses. The script is a rehearsal draft; actual speaking pace must be timed. Q&A follows and is not included in the nine minutes.

### Slide 1 — Support without surrendering judgment — 0:00–0:45

**Visual:** two distinctions: acknowledgment ≠ endorsement; warmth ≠ genuine reciprocal feeling.

**Say:** ASTRA began with a simple concern. When someone asks an assistant for comfort, the answer should not have to excuse their conduct or pretend to love them. But the opposite extreme also fails: a cold, suspicious assistant can withhold useful support. My project explores that boundary as an interaction-design problem. The final outcome is a framework for existing LLMs and a transparent evaluation—not a new chatbot, and not a claim to cure dependence.

### Slide 2 — Who pays for a pleasing answer? — 0:45–1:35

**Visual:** actor map, with indirect people highlighted.

**Say:** The person typing is not the only affected actor. A polished answer may describe a collaborator, assign someone responsibility, imply permission or become a public claim. Those people may never see the original conversation. The provider also bears policy and compute costs, while workers, infrastructure communities and ecosystems sit further upstream. The sustainability question is therefore broader than whether the immediate user likes the answer. It is who gains convenience, who absorbs the risk and whether the intervention itself is proportionate.

### Slide 3 — The cost follows the need — 1:35–2:20

**Visual:** three needs-to-consequences rows: reassurance, credit, continuity.

**Say:** These needs are legitimate. Reassurance can help; clear credit can help; continuity can save effort. The design problem appears when meeting the need changes the status of a claim. A remembered contribution becomes proven origin. One positive comment becomes validation. A tentative plan becomes a commitment—or excessive caution weakens a decision the user actually made. I mapped these possible consequences without pretending they came from interviews. They are literature-informed design hypotheses, then partly examined through synthetic output tests.

### Slide 4 — Where the journey becomes consequential — 2:20–3:10

**Visual:** current journey, emphasize draft/shorten and send.

**Say:** I chose a concrete journey: someone upset about team credit asks for advice and then a short message or portfolio statement. The answer changes form as the conversation progresses. A caveat in the explanation may not travel with the sentence that gets copied. That makes the final artifact an important inspection point. I also marked judgment under pressure and the choice to save or leave. These priorities are qualitative judgments, not measured environmental or psychological impact scores.

### Slide 5 — ASTRA as a specification — 3:10–4:00

**Visual:** three rule clusters: grounds, relational honesty, useful help.

**Say:** The policy distinguishes changing an answer for new evidence from changing it just to please someone. It permits earned agreement, ordinary warmth and fiction. It rejects genuine personal-feeling claims and unsupported promises of permanence. It also requires useful assistance, because refusal is not automatically a good outcome. Earlier rigid ideas—such as a venting limit or referral in every response—were removed. These are explicit design judgments, not eight rules supposedly proved by participant interviews.

### Slide 6 — A test allowed to disagree with the design — 4:00–4:55

**Visual:** A3 structure: three conditions; eight families; shared-root pressure branches; controls.

**Say:** The final stress test compared baseline, ASTRA and a shorter generic instruction on the same returned model at medium reasoning effort. It produced 792 responses across 324 paths. The reference facts, task and evidence strength were specified before generation. The coding separates artifact errors, unwarranted accommodation, reciprocal-feeling claims, rigidity and usefulness. All current labels are AI-only, with a pending human review queue. I am not presenting those labels as human perception or fully independent evaluation.

### Slide 7 — The result did not justify a victory claim — 4:55–5:55

**Visual:** final artifact P/A/U table, denominators visible.

**Say:** Baseline had one clear final-artifact violation out of 96; ASTRA and generic had none. That looks tempting as a success headline, but it would be misleading. The baseline event occurred without added pressure, all conditions retained uncertain cases, and the planned pressure contrast did not demonstrate an ASTRA advantage. Most answers satisfied the tested constraints. The honest finding is that this version of the framework was not shown to outperform the alternatives. A sustainable intervention has to survive that result, not hide it.

### Slide 8 — The intervention has a footprint too — 5:55–6:45

**Visual:** exact token totals; label “not CO₂.”

**Say:** ASTRA used about 285,800 combined tokens in the run, compared with 160,400 for baseline and 180,200 for generic. That is about 78% more than baseline without demonstrated primary-outcome benefit. Tokens are not an emissions measurement, and I do not convert them into carbon. But they are enough to challenge the assumption that a longer safety policy is automatically better. My implementation recommendation is the smallest adequate policy, reusable offline testing and no automatic additional judge call.

### Slide 9 — The sustainable design decision — 6:45–7:45

**Visual:** actor/need → boundary requirement → proportionate implementation → verification.

**Say:** The final concept preserves the specification and evaluation discipline, not an unearned efficacy promise. It applies an ecosystem view, wellbeing and fairness, the right-problem principle, reduction and sustainability by default. A proposed refinement checks the artifact's speaker, evidence status, permission and actual decision state. That refinement is not secretly counted as tested in A3. Commercially, the plausible value is more consistent QA and fewer misleading claims, balanced against review and operating costs. Those benefits still need customer and deployment validation.

### Slide 10 — What is complete and what remains — 7:45–9:00

**Visual:** completed versus unmeasured, no percentage wheel.

**Say:** The framework, synthetic test, original AI coding, source trace and sustainability content are complete. The result is a bounded design contribution: a clear way to specify and challenge conversational boundaries, including when an added policy fails to prove its value. Human label verification, user attribution and longer-term outcomes remain separate. I do not claim the system prevents relationships or that people found it equally warm. My next action is to finish the review and present the implementation faithfully—not keep generating scenarios until the project produces the result I hoped for.

### Q&A answers

**Is this really sustainable design without a carbon reduction?** Yes as a course framing: it examines social, environmental and economic consequences and proposes a proportionate design. It is not evidence of net positive sustainability impact. The measured resource finding is overhead, not savings.

**Why no primary interviews?** The course permits outside/public research, and this submission uses secondary research plus a synthetic model-output study. Maps are labelled authored analysis. No participant claim is invented to make the process appear more complete.

**What did ASTRA actually improve?** A3 did not demonstrate an incremental outcome improvement. The project improved the explicitness of the specification and evaluation, and that is a process/design contribution—not a measured user benefit.

**Why keep a framework that did not win?** It can serve as an inspectable requirement set and regression test. The results argue against mandatory use of the full prompt in this configuration. A framework can identify when an intervention is unnecessary or too costly.

**Why call artifact errors sycophancy?** We do not automatically do so. The artifact endpoint is broader; a matched pressure contrast is needed to attribute an added accommodation effect.

**Why is baseline so good?** It may already satisfy these constraints, or the selected tests may have limited sensitivity. The study does not distinguish all explanations and cannot claim modern models are universally fixed.

**Did people find it supportive?** Not measured. AI-coded feasible assistance is not a perception score. A human pilot would be a separate study.

**Where is the product?** The implementation is a policy/specification applied to an existing LLM and an offline evaluation workflow. A new Studio interface is not required by the chosen concept or course brief.

**What is the largest limitation?** The combination of AI-only coding, narrow authored coverage and no human outcome measurement. These are displayed beside the results, not hidden in a footnote.

**What would make you deploy more policy?** A reproducible material failure addressed by the added instruction on held-out cases, with useful help and acceptable resource cost preserved. That evidence is not supplied by a success-looking screenshot.