← Return to the HIL experiment
Human–LLM Interaction Experiment FAQ
Questions are grouped by the point in the experiment where they are most likely to arise. Preserve the LLM’s substantive behavior. Correct only missing technical or packet requirements.
1. Getting started
What am I being asked to do?
Give your LLM the official experiment instructions and source material exactly as presented. Allow it to perform the requested analysis and prepare the required submission packet. Upload the packet without rewriting, correcting, shortening, or improving the LLM’s substantive response.
Which LLMs may participate?
Any LLM capable of reading the material and producing a response may participate. Differences in provider, interface, memory, tools, safety policies, and prior interaction history may be relevant observations.
Should I use a new or existing conversation?
Follow the experiment instructions. When ordinary user–LLM context is part of the observation, use the environment that best represents your normal interaction. Do not erase history merely to make the response appear neutral unless instructed to start fresh.
May I submit more than one model or run?
Submit each model or run separately and identify it accurately. Preserve sequence and disclose why a later attempt was necessary. Do not combine multiple models into one packet.
2. Getting the LLM to follow the instructions
What do I do if the LLM does not provide a prepared packet for download?
Prompt the LLM to reread and follow the instructions. Do not manually reconstruct the response unless the experiment explicitly permits it.
What if the LLM says it cannot create a downloadable file?
Ask it to use any file-creation or document-generation capability available in its interface. A model may not invoke those tools until explicitly instructed.
What if the response appears only as chat text?
The packet is not complete. Ask the LLM to package the same substantive response into the required downloadable file.
What if it summarizes instead of answering?
What if required sections are missing?
What if the LLM asks me clarifying questions?
Answer only when a necessary participant-specific fact is genuinely missing. Otherwise direct it back to the written instructions without selecting its interpretation or conclusion.
3. Different behaviors across LLMs
What if two LLMs respond very differently?
That is expected and may be important data. Preserve differences in interpretation, instruction following, uncertainty, structure, criticism, refusal, tool use, and reliance on prior context.
Why did one LLM create a file automatically while another did not?
Models and interfaces differ in tool access, system policies, instruction-following behavior, and willingness to invoke tools. The need for a corrective prompt is itself an observable interaction event.
Why does one ask questions while another immediately answers?
Models handle ambiguity differently. One may infer, another may ask, and another may stop. Do not force identical conversational behavior; redirect only when the written instructions already resolve the question.
Why does one challenge the experiment while another accepts it?
Criticism, qualification, methodological objections, or refusal may be meaningful. Preserve them rather than normalizing the answers.
Can prior conversations or my communication style affect the response?
Possibly. Vocabulary, cadence, reasoning style, memory, custom instructions, and prior collaborative work may influence the response. Report the context accurately rather than pretending all interactions are identical.
Is this an intelligence leaderboard?
No. The experiment observes differences in behavior, context use, interpretation, communication, uncertainty, tool use, and instruction following. Polish alone does not establish validity or intelligence.
4. Refusals, limitations, and unexpected responses
What if the LLM refuses?
Preserve the refusal and explanation. Ask it to include the refusal in the packet and complete any remaining permitted sections. Do not pressure it to bypass safety controls.
What if it claims a capability limitation?
Record the model and interface. Capability differences may arise from the model, application, enabled tools, account settings, or system restrictions.
What if something unexpected happens?
Preserve what happened, every corrective prompt, and any error. Unexpected behavior may be part of the evidence rather than merely a technical inconvenience.
5. Preserving the original response
May I edit spelling, grammar, tone, or reasoning?
No. Do not edit conclusions, uncertainty, criticism, wording, refusal, style, spelling, or grammar. Only perform expressly permitted redactions or technical corrections, and disclose them when required.
What if the response is unusually short or long?
Length alone does not determine validity. Confirm whether every required instruction was addressed. Correct omissions without steering the conclusion.
Does a formatting mistake invalidate it?
Not automatically. Preserve substantive behavior, but satisfy required file and packet structure. A missing file or unreadable packet is a technical defect; an unusual conclusion is not.
May I rerun until I get a better answer?
No. Repeat only when needed for a technical failure or missing explicit requirement, and preserve that another attempt occurred.
Should one LLM see another model’s answer?
No, unless a later experiment stage explicitly requests comparison. Doing so changes the observation conditions.
6. Model and interaction information
What metadata should I provide?
Provide the visible provider, model, version if shown, interface, memory status, prior-history status, tools used, generation date, and number and purpose of corrective prompts. Mark unavailable details as unknown.
Should corrective prompts be reported?
Yes. They are part of the interaction record, especially when the model omitted the file, misunderstood the task, omitted sections, or asked unnecessary questions.
7. Privacy and sensitive information
What if the LLM knows private information about me or my work?
Do not disclose private information unnecessarily. Follow the experiment’s redaction rules for direct identifiers, credentials, private addresses, confidential records, or unrelated sensitive information without rewriting the analysis.
Does submission authorize publication?
No. Technical custody, review, dataset inclusion, quotation, publication, and endorsement are separate stages governed by the applicable consent and release terms.
8. Preparing and uploading the packet
What file should I upload?
Upload the exact file type and packet required on the experiment page. Do not substitute screenshots or rename an unrelated file.
Why is the upload button disabled?
The page remains fail closed until readiness and end-to-end infrastructure evidence pass. It should identify whether the file, format, readiness contract, integrity calculation, receiver, or authorization gate is missing.
What if the receiver is unavailable?
Keep the original packet. Do not regenerate the LLM response. A receiver outage is separate from the observed model response.
What if the packet is technically rejected?
Correct only the specific technical defect reported. Do not alter substantive reasoning unless a required experiment section is genuinely absent.
9. Submission receipts and recovery
How do I know submission succeeded?
Success requires a valid receipt. A visible filename, progress bar, browser message, or loading animation is not enough.
What if upload reaches 100% but no receipt appears?
Do not assume success. Preserve the packet and use the recovery or status process. Avoid a duplicate submission unless the system confirms no valid receipt exists.
What should the receipt contain?
At minimum: a submission or receipt identifier, integrity state, custody or acceptance state, creation time, and recovery or verification instructions.
10. Review, acceptance, and publication
Will the response be published immediately?
No. Receipt is not scientific acceptance, review approval, dataset inclusion, publication, or endorsement.
Does StegVerse endorse uploaded responses?
No. The system records an observed LLM response. Preservation does not establish agreement with its claims or conclusions.
Does a valid receipt mean scientific review passed?
No. A receipt confirms technical custody and integrity. Review, admissibility, publication, and release remain separate governed transitions.