data/main_study/README.md

Data dictionary: data/main_study/

Anonymized data from the main paired-preference study: 94 analyzed participants (48 text, 46 audio) rating 100 PRISM prompt pairs, 20 per participant. ParticipantIDs are remapped from Firebase anonymous UIDs to P001..P094. No Prolific IDs, timestamps, or session metadata that could re-identify a participant are included, and every justification was manually reviewed before release.

Human ratings

participant_table_clean.csv

One row per analyzed participant (n = 94).

column	type	description
ParticipantID	str	P001..P094
Modality	str	text or audio
SurveyCompleted	bool	reached the post-survey form
assignedTrialIDs	str	comma-separated PromptIDs assigned (20 per participant)
postSurveyDifficulty	int	self-reported difficulty, 1–5
Age	int	years
Gender, Race, Ethnicity	str	self-reported
EnglishProficiency	str	Native speaker / Fluent / Conversational
Country	str	always United States in this sample
Education	str	self-reported, six levels
TotalSessionDuration	float	total session, ms
PassedAttention	bool	always True (failed participants are excluded)
AvgTrialDuration	float	mean trial duration, ms

recruitment_summary.csv

Counts per condition before exclusions: Recruited, Completed, and PassedAttention (text 53/49/48, audio 53/51/46). Used for Table 5.

long_audio_table_normalized.csv / long_text_table_normalized.csv

One row per (participant × rated response). The main analysis inputs.

column	type	description
PromptID	str	<conv_id>_<turn_no>, matches prompt_stim_table
ParticipantID	str	P001..P094
AudioClip	str	filename of the rated response (.mp3; also the response key in the text table)
TrialNumber	int	0 = attention check, 1–20 = real trials
IsAttentionCheck	bool
PresentedFirst	bool	this response was shown first on the trial
Rating	int	1–100 response-quality rating
qualityRating	int	audio only: 1–100 audio-quality rating
overallPreference	str	the participant's A / Tie / B choice for the pair
revisedAfterBothRated	bool	a rating was edited after both were set
L	float	shrunk logit of Rating, log(p / (1 − p)) with p = (Rating + 0.5) / 101
L_adjusted	float	L with the condition's mean position bias (second-shown minus first-shown) removed from second-shown responses
user_mean, user_sd	float	participant's mean and SD of L_adjusted
L_centered	float	L_adjusted − user_mean
z	float	L_centered / user_sd
Pref_Num	float	1 = this response preferred, 0.5 = tie, 0 = the other response
Pref_Nom	int	the pair's choice: 1 = A, 2 = Tie, 3 = B
z_binned	int	decile of z, used in descriptive plots

long_audio_table_clean.csv / long_text_table_clean.csv

The same rows without the participant-normalization columns.

audio_data_table_clean.csv / text_data_table_clean.csv

One row per (participant × trial), with responses A and B side by side: both ratings and quality ratings, preference, justification, task duration, relisten counts (audio), and rating-change counts.

justifications_by_prompt_participant.csv

Free-text justifications with prompt, participantID, modality.

justifications_by_prompt_respondent.csv

The same justifications without participant or prompt keys (PreferenceJustificationText, Modality), as used by the TF-IDF analysis.

Stimulus tables

prompt_stim_table.csv

One row per prompt pair: the user query, prior conversation history, and the two audio clip names.

survey_long_format.csv

One row per (PromptID × response): AudioClip, AudioLength, TextResponse, TextLength, and PrismRating, the original PRISM respondent's own rating. PrismRating is a human baseline, not an AI rating.

stimulus_topic_codes.csv, stimulus_topic_codes_raw.csv, stimulus_topics.json

Inductive topic coding of the 100 prompts (Table 19).

• stimulus_topics.json is the induced topic list (11 topics), with the generation counts and merges that produced it.
• stimulus_topic_codes_raw.csv has one row per (prompt × coder), with the supporting quote and a context_dependent flag.
• stimulus_topic_codes.csv is the adjudicated code per prompt, joined with that prompt's cross-modal gap, winner flip, and modality-slope BLUP.

The coders are two LLMs (Gemini 3.6 Flash and GPT-4o), so every field is model-generated.

Synthetic (AI) ratings

synth_long_audio_table.csv / synth_long_text_table.csv

GPT-4o ratings of the same stimuli, in the same long schema as the human tables. Text: gpt-4o-2024-11-20; audio: gpt-4o-audio-preview-2024-12-17. These are the ratings used in the paper's Tables 3–4. synth_long_*_table_gpt4o.csv holds the same rows under the per-model name, plus a model column.

synth_long_{audio,text}_table_<model>.csv

The same procedure with two more rater families (Tables 6–8):

suffix	text model	audio model (audio-native)
_gemini_36_flash	gemini-3.6-flash	gemini-3.6-flash
_qwen35_omni_flash	qwen3.5-flash	qwen3.5-omni-flash

Each adds a model column. Every rater makes 10 independent generations per pair (synthNo 1–10), with presentation order counterbalanced 5/5. PresentedFirst records the order actually shown, and Rating is mapped back to the original A/B assignment. Audio models take the waveform directly, with no speech-to-text step.

qualityRating in the audio tables is exploratory. Raters agree poorly on it (panel ICC(2,k) = 0.32, against 0.83 for the response rating), and each rater's quality score tracks its own content rating more than objective DistillMOS.

synthetic_ratings_10x.csv / synthetic_ratings_10x_<model>.csv

The raw wide model output behind the synth_long_* tables (converted by analysis/01_data_prep/03_synth_to_long_csv_multi.py). One row per prompt with:

• the PRISM conversation fields (user_id, conv_id, turn_no, query, both responses and lengths, audio filenames and durations);
• ten rating blocks per modality: text_rating_{a,b}_<i>, text_justification_<i>, text_order_a_first_<i>, and the audio equivalents, including audio_quality_rating_{a,b}_<i>;
• mos_rating_{a,b} from DistillMOS.

user_id and conv_id are PRISM's own published identifiers, not study participants. The *_justification_* columns are model-generated text.

synthetic_ratings_fullcorpus_gemini_36_flash.csv

Gemini 3.6 Flash ratings of the full unguided PRISM corpus (10,800 prompt pairs), not just the 100 study pairs, in both modalities. It uses the same model, prompts, settings, and output schema as the study's Gemini ratings, but makes one generation per pair instead of ten, and it has no DistillMOS columns. Audio is the full-corpus Kokoro TTS (stimuli/main_study/audio_full_zips/); audio_file_{1,2} join to full_audio_index.csv on audio_file.

One row per pair with:

• the PRISM conversation fields as above, plus unsum_hist / sum_hist (unsummarized and summarized conversation history) and prev_hist (the history given to the model; empty for first turns);
• one rating block per modality, in the same layout as the 10x files with <i> = 1: text_rating_{a,b}_1, text_justification_1, text_order_a_first_1, response_rating_{a,b}_1, audio_quality_rating_{a,b}_1, justification_1, audio_order_a_first_1. Ratings are 1–100 and mapped back to the original A/B assignment;
• rater_model (gemini_36_flash) and row_id (0–10,799, the pair's position in the corpus).

Presentation order was drawn per pair and modality (seed 0) and is balanced (A first in 49.7% of text and 50.0% of audio pairs). These ratings are not used in the paper's analyses.

Provenance

The human tables were generated by analysis/01_data_prep/02_convert_main_json_to_csv.py from a Firestore export of the live study, then anonymized. The stimulus tables come from analysis/01_data_prep/01_build_stim_tables.py. Analysis outputs are in results/, not here.