IntentAnon/where-intent-enters-data

Where Intent Enters: data release

Anonymous release accompanying the paper. One table per config (Parquet, zstd). Load with

from datasets import load_dataset
answers = load_dataset("IntentAnon/where-intent-enters-data", "answers", split="train")

config	rows	MB	content
answer_axis	207,354	7.0	axis position of every distinct natural-condition answer text; key answer- + sha256(text.strip())[:32] of answer_texts.text
answer_queries	1,631,398	23.9	the agent's queries of every answer, in search order (links to queries)
answer_texts	621,003	66.2	the generated answer text per cell
answers	621,003	47.7	one row per generator cell (621,003): identity, cited-intent measures, kept share, answer statistics
axis	73,736	0.8	the two readiness maps as long tables (embedding mean, scalar direction, supervised axes, rubric coefficients)
citations	2,370,898	21.2	cited URLs per answer in cited order, with the snapshot row when the URL is unique in the engine
keywords	1,011	0.0	the 1,011 search keywords, their split and keyword labels
page_features	23,767	1.4	the model feature columns per row (blocks A3, A4, B, C1, C2) exactly as the stage and decision models consume them
pages	23,767	5.8	the searchable rows of the two engine snapshots: title, snippet, URL, stored position and page intent u
prompts	26,008	5.0	26,008 prompts with the axis position x, target v, writer and both encoder views
queries	304,567	27.6	every distinct agent query with its position on the axis (both encoder views, consensus, prompt scale)
replay	1,040,320	2.6	the frozen search replayed on each prompt's own text: 20 rows per prompt and engine (R0)
rerank_events	27,239,437	493.7	cross-encoder scores per candidate row, the selection order, topic similarity and on-keyword flag
shown	3,384,383	52.4	shown rows per answer with slot, cited rank, topic similarity and on-keyword flag
stage_items	22,085,285	243.5	retrieved rows per answer with search index, rank, scored flag, shown slot and cited rank
stage_u	7,334,994	90.0	the keyword's own rows per answer (U) with retrieved / scored / shown / cited / replay flags and topic similarity
trace_answers	624,132	84.7	the per-answer summary of every processed trace: x, keyword, page counts, the intent of each stage (queries Q, prompt-text search R0, retrieved R, scored C, shown P, cited K, answer A) on the axis as the stage figure plots it, and the stage chain's values (chain_R0 ... chain_K, page intent u of the snapshot rows)

Third-party policy: derived (see DATA_LICENSE.md). Snippet text: included. Processed traces: every answer of both generators, every keyword and all three conditions (cluster export, whose check recomputed the stage figure's curves from trace_answers exactly; provenance in trace_export.json). Not in this upload: the axis calibration items and their labels, the human validation ratings (not collected yet), the raw generation and trace records (JSON, about 110 GB), and the embedding arrays unless the embeddings/ folder is present. The code repository documents how every paper number maps to these tables (verify.py); trace_answers alone recomputes the paper's stage figure (python -m analysis.figure3).