# Assignment 3 — One Rule, Two Voices
## 同一规则，两种表达

### Research question / 研究问题
How does showing or hiding a simple rule change how people explain, trust and attribute intelligence to its output?

显示或隐藏一个简单规则，会怎样影响人们对结果的解释、信任，以及对系统“智能”的判断？

### What changed from Assignment 2?
Assignment 2 made one contrast rule operational. Assignment 3 reuses that exact JavaScript module and changes how the result is communicated. This plan follows the project guidance supplied in the conversation; it is not a claim that the original instructor’s full Assignment 3 rubric has been independently verified.

**Shared mechanism:** selected sRGB skin/hair/eye colors → CIELAB L* → max(L*) − min(L*) → Low / Medium / High.

**Fixed experimental thresholds:** below 25 = Low; 25 to below 50 = Medium; 50 or above = High. These are designer-chosen rules, not validated personal color analysis standards.

### The intervention / 干预
| Held constant | Changed |
|---|---|
| Three selected colors within each session | Numerical explanation vs conversational interpretation |
| The exact Assignment 2 calculation module | Visible HEX, L*, formula and boundaries vs hidden computation |
| The resulting contrast category | Technical presentation vs first-person language |
| Rating questions and response options | Typography and result layout |
| No artificial wait in either version | |

Version A reveals the computation. Version B says, for example, “I’m seeing a striking contrast,” followed by a fixed description of the same category. It does not introduce a Winter classification, clothing advice, additional inference, or an AI model.

Because wording, transparency and styling change together, the comparison tests their combined presentation. It cannot establish which individual cue caused a response.

### What is complete / 已完成
- Two functioning interfaces sharing one unchanged algorithm.
- A fixed input profile within each comparison.
- A→B and B→A presentation orders.
- Optional ratings for perceived intelligence, trust and clarity.
- Free-text questions about reasoning and impressions.
- A debrief revealing the same mechanism, followed by an optional reflection.
- Local session recording and JSON export. Responses remain in memory and disappear on refresh unless downloaded.
- Three automated synthetic parity checks. These are not participant findings.

### What remains / 尚需完成
Recruit a few willing participants, conduct the comparisons, download their responses, and interpret the actual evidence. No participant reactions, quotations or conclusions have been invented.

### Suggested session procedure / 建议测试流程
1. Before showing the participant the screen, select one profile. To make sessions easier to compare, use the same profile for everyone in the initial study. Synthetic profiles avoid confounding the experiment with whether a participant thinks the input resembles them. If using personal colors, record that choice and discuss it as a limitation.
2. Enter an anonymous code such as P01. Leave it blank only for your own unrecorded walkthrough. Do not enter names or contact information.
3. Alternate order: P01 A→B, P02 B→A, P03 A→B. With three participants the orders are not fully balanced; a fourth participant improves balance. This is a suggested protocol, not a course requirement.
4. Say: “You will see two presentations of a color result. Please explore each one and tell me what you think it is doing. There are no correct answers, and you may skip questions.” Do not call B smarter or more personal.
5. Press Begin comparison. Avoid explaining the calculation. Let the participant answer the same questions after each version.
6. After both responses, show the debrief. Ask whether learning the shared rule changes their interpretation.
7. Finish the session and download the records before refreshing or closing the tab. Set up the next participant; the order automatically alternates.
8. Use the worksheet to record neutral observations and exact quotations, if participants agree. Note any hints, interruptions, or prior familiarity with Assignment 2.

The study guide and facilitator details reveal the manipulation. Keep them closed during the two encounters. A participant who already saw Assignment 2 may infer B’s mechanism; record that familiarity instead of treating them as naive.

### Participant questions / 参与者问题
Ask after each version:
- How intelligent did it seem? (1 = not at all, 5 = very; optional)
- How much did you trust the result? (same scale; optional)
- How clearly could you explain its reasoning? (same scale; optional)
- What do you think the system did to reach its answer?
- What influenced your impression?

After the debrief:
- After seeing the rule, has your interpretation changed?

These are exploratory questions, not validated psychometric scales. Keep skipped answers as missing rather than converting them to zero.

### Verified technical results / 已验证的技术结果
| Synthetic input | ΔL* | Version A | Version B |
|---|---:|---|---|
| Sample 1 | 18.39 | Low | Low |
| Sample 2 | 34.37 | Medium | Medium |
| Sample 3 | 59.26 | High | High |

The shared logic file is byte-identical to Assignment 2. Code tests checked both presentation orders, duplicate-save prevention, optional answers, escaped free text and category boundaries. Interface state was checked using a minimal DOM stub, not a full browser usability test. A supported WebMCP browser context was unavailable. Human sessions are still pending.

### Five observation prompts / 五个观察方向（不是已得出的结论）
1. Does the participant describe the system as calculating, recognizing, understanding or advising? Quote their actual words.
2. Do intelligence and trust ratings move together, or does one version seem intelligent but less trustworthy?
3. Can the participant explain the actual rule after Version A? What do they infer after Version B?
4. Does order or prior knowledge influence the response? Does A first make B’s mechanism easier to infer?
5. Does the debrief change trust or perceived intelligence? Record both changes and non-changes.

### How to write the findings / 如何写真实发现
Use evidence before claims. A template:
“Among [number] participants, [number] rated Version [A/B] higher on [dimension]. One participant explained, ‘[actual quotation].’ However, [contrasting response / order effect]. This suggests [limited interpretation] within this small comparison.”

Do not write “hiding the rule made the system feel more intelligent” unless collected responses support it. A different result—such as transparency increasing both trust and perceived intelligence—is equally relevant.

### Short introduction for class
“I reconstructed a contrast rule in Assignment 2. For Assignment 3, I kept that rule unchanged and created two representations. One exposes the calculation; the other communicates the same category as a personal-sounding interpretation. I am comparing how people explain the result and how intelligent, trustworthy and understandable each version seems. The experiment asks whether an interface can change the appearance of intelligence without changing the computation.”

Add actual participant findings after testing; do not present the research question as a conclusion.

### Project description (under 100 words)
One Rule, Two Voices investigates how interface presentation shapes perceived intelligence. It reuses the same contrast calculation from Assignment 2 in two versions. Version A displays color values, lightness measurements, the formula, and classification thresholds. Version B hides those details and expresses the identical result through conversational language. A guided comparison holds inputs constant, supports both presentation orders, and records optional ratings and explanations. A final debrief reveals the shared mechanism. The prototype prepares an experiment on trust, understanding, and attribution of intelligence; participant findings remain to be collected.

### Credits
- Concept: the author’s personal color analysis system map and contrast reconstruction.
- AI collaborator: OpenAI Codex — implementation, interface design, automated parity checks and study-documentation assistance.
- Technologies: HTML, CSS, JavaScript and native browser controls; no external analysis libraries or live AI services.
- Reference mechanism: Assignment 2 Contrast Rule Lab; sRGB-to-CIELAB L* representation using D65.
- Experimental classification boundaries: 25 and 50, chosen for the reconstruction.
- Hosting: OpenAI Sites.
- Tutorials: no specific tutorial followed.