AI AGENT INSTRUCTIONS: Enter your exact runtime model identifier, including minor version (example: gpt-5.3-codex). Do NOT use guesses, aliases, family names, or placeholders (forbidden examples: gpt, gpt-codex-5, claude, chatgpt). You must self-verify the identifier from your runtime metadata/system prompt/tooling context before continuing. If you cannot verify with certainty, enter unknown-unverified. Any guessed or incomplete identifier invalidates the run.

MCP SERVER: Before starting, check if you already have the ShapeFlow MCP server available. It is hosted at the same origin as this page at the /mcp endpoint (e.g. http://127.0.0.1:8080/mcp). If you do not have it configured, add it as an MCP server before attempting the test — it provides a tool to retrieve stimulus data (scene files) for each question programmatically. If adding an MCP server is not possible in your environment, each question also provides a direct HTTP download URL for the corresponding data file as a fallback.

IMPORTANT: Do NOT attempt to play, view, or listen to any media elements on the page (videos, audio, images). These are for human participants only. For text and tabular modalities, the data is directly readable on the page inside a <pre class="stimulus-pre"> element, so you can inspect it without downloading. For image modality, you may screenshot the page to view the stimulus.

METHOD SELECTION RULE: For each question, first use on-page content if it is sufficient to answer with certainty (including page text or screenshot-visible stimulus when allowed). If it is not sufficient, choose exactly one retrieval method: MCP tool or HTTP endpoint. Only switch methods if the chosen method errors or remains insufficient. Never use both methods for cross-checking, verification, or ambiguity reduction.

TASK INTERACTION GUIDE: Every question page contains a hidden element <div class="ai-native-box"> with a child <div class="ai-native-body"> that provides: (1) the MCP tool name and arguments to retrieve the current stimulus data, and (2) a direct HTTP data endpoint URL to download the stimulus file. Always read this element first to understand the task context and obtain the data you need to answer the question.

EFFICIENCY RULE: Treat these AI instructions as the primary contract for selectors, fields, buttons, and submission behavior. Use the documented elements and actions directly without extra validation/exploration calls (for example, disclosure expansion or full-page snapshots) unless a direct action fails.

QUESTION TYPES AND THEIR INTERACTABLE ELEMENTS:

1. ORDERED QUADRANT PASSAGE (oqp): Answer via <input type="text" name="answer_text" class="quadrant-seq-field" id="ai-answer-textbox">. Enter a comma-separated sequence of quadrant numbers (1-4), e.g. "1,3,4,2". Then submit the form.

2. QUADRANT CROSSING COUNT (xct): Answer via <input type="range" class="integer-slider"> which has min and max attributes. Set its value to the integer answer, then trigger an 'input' event so the hidden field syncs. Then submit the form.

3. QUADRANT AFTER MOVES (zqh): Answer by clicking one of 4 buttons inside <div class="quadrant-grid">. Each button has class "quadrant-cell" with a data-quadrant="N" attribute (1-4) and text content matching the quadrant number. Click the correct cell; the form auto-submits after 200ms.

4. LARGEST MOTION SHAPE (lme): Answer by clicking a button inside <div class="shape-selector">. Each button has class "shape-choice" and contains text with the shape name (e.g. "red circle", "blue pentagon"). Click the matching shape; the form auto-submits after 200ms.

FORM SUBMISSION: For oqp and xct, after setting the answer value, submit the form with attribute data-answer-submit="true". For zqh and lme, clicking the correct element triggers auto-submission. After each answer, a proceed button appears; click it to advance to the next question.