Pair of datasets compared column-by-column. curated has 279,196 rows; firehose has 125,645 rows. Divergence score combines normalised null drift, mean/length delta, entropy delta, top-value jaccard, and language-mix jaccard. Every term is capped at 1.0.
!pip install saturn-dissect
import subprocess
subprocess.run([
"saturn", "compare",
"hf://lukeslp/bluesky-alt-text:[train][source_mode=author_feed]", "hf://lukeslp/bluesky-alt-text:[train][source_mode=jetstream]",
"--label-a", "curated",
"--label-b", "firehose",
"--findings", "bluesky-alt-text-curated-vs-firehose.json",
"--llm", "anthropic:claude-opus-4-7",
])The curated dataset (279,196 rows) is roughly 2.2× the size of firehose (125,645 rows), and several columns differ structurally rather than just statistically. Four fields—`query`, `image_fullsize_url`, `image_thumb_url`, and `author_handle`—show a +100% null shift with zero top-value overlap, indicating they are populated on one side and entirely absent on the other. Text fields also diverge in shape: `alt_text` has a mean length +79 characters with low language overlap (jaccard 0.35) and almost no shared top values (0.03), while `raw_record_json` is +154 characters longer on average with jaccard 0.25. These patterns suggest the two sources capture different schemas/enrichments rather than just sampling noise. Confidence is high that the divergence is structural, though the evidence does not say which side carries the nulls.
citing: a_row_count · b_row_count · divergences
report.divergence_summary()
| column | kind | score | signals |
|---|---|---|---|
| alt_text | text | 2.01 | len_mean +79, lang-jaccard 0.35, top-val-jaccard 0.03 |
| query | categorical | 2.00 | null +100%, top-val-jaccard 0.00 |
| image_fullsize_url | text | 2.00 | null +100%, top-val-jaccard 0.00 |
| image_thumb_url | text | 2.00 | null +100%, top-val-jaccard 0.00 |
| author_handle | categorical | 2.00 | null +100%, top-val-jaccard 0.00 |
| raw_record_json | text | 1.89 | len_mean +154, lang-jaccard 0.25, top-val-jaccard 0.00 |
report.columns["image_mime_type"].delta
| stat | curated | firehose | Δ |
|---|---|---|---|
| n | 279,196 | 125,645 | — |
| null rate | 0.0% | 0.0% | +0.0% |
| unique | 5 | 8 | — |
| n_unique_delta | — | — | +3 |
| entropy_delta | — | — | +0.552 |
| top-value jaccard | — | 0.62 | |
report.columns["source_mode"].delta
| stat | curated | firehose | Δ |
|---|---|---|---|
| n | 279,196 | 125,645 | — |
| null rate | 0.0% | 0.0% | +0.0% |
| unique | 1 | 1 | — |
| n_unique_delta | — | — | +0 |
| entropy_delta | — | — | +0 |
| top-value jaccard | — | 0.00 | |
report.columns["image_ref"].delta
| stat | curated | firehose | Δ |
|---|---|---|---|
| n | 279,196 | 125,645 | — |
| null rate | 0.0% | 0.0% | -0.0% |
| unique | 260,595 | 114,610 | — |
| n_unique_delta | — | — | -1.46e+05 |
| len_mean_delta | — | — | +0 |
| len_median_delta | — | — | +0 |
| len_p95_delta | — | — | +0 |
| word_mean_delta | — | — | +0 |
| duplicate_rate_delta | — | — | +0.0213 |
| vocab_size_delta | — | — | -644 |
| top-value jaccard | — | 0.05 | |
The query column is fully populated in curated (null_rate 0.0 across 279196 rows, 489 unique values led by 'firefaerie81.bsky.social' at 2.4%) but entirely null in firehose (null_rate 1.0 over 125645 rows, flagged all_null). This yields a null_rate_delta of 1.0 and a top_value_jaccard of 0.0, meaning the two slices share no overlap on this field. Curated also shows a healthy entropy of 7.68, while firehose has no distribution to compare.
report.columns["query"].delta
| stat | curated | firehose | Δ |
|---|---|---|---|
| n | 279,196 | 125,645 | — |
| null rate | 0.0% | 100.0% | +100.0% |
| unique | 489 | — | — |
| top-value jaccard | — | 0.00 | |
The columns diverge completely: curated has a 0.0 null_rate across 279196 rows with url_rate 1.0, while firehose is entirely missing (null_rate 1.0 over 125645 rows, flagged all_empty). With null_rate_delta of 1.0 and top_value_jaccard at 0.0, there is no overlap to compare further. Curated also shows one_word_rate 1.0 and a duplicate_rate of 0.0665 (18560 duplicates), but firehose provides no values to contrast.
report.columns["image_fullsize_url"].delta
| stat | curated | firehose | Δ |
|---|---|---|---|
| n | 279,196 | 125,645 | — |
| null rate | 0.0% | 100.0% | +100.0% |
| unique | 260,636 | — | — |
| top-value jaccard | — | 0.00 | |
report.columns["post_cid"].delta
| stat | curated | firehose | Δ |
|---|---|---|---|
| n | 279,196 | 125,645 | — |
| null rate | 0.0% | 0.0% | +0.0% |
| unique | 232,017 | 103,528 | — |
| n_unique_delta | — | — | -1.28e+05 |
| len_mean_delta | — | — | +0 |
| len_median_delta | — | — | +0 |
| len_p95_delta | — | — | +0 |
| word_mean_delta | — | — | +0 |
| duplicate_rate_delta | — | — | +0.00705 |
| vocab_size_delta | — | — | -389 |
| top-value jaccard | — | 0.00 | |
report.columns["image_index"].delta
| stat | curated | firehose | Δ |
|---|---|---|---|
| n | 279,196 | 125,645 | — |
| null rate | 0.0% | 0.0% | +0.0% |
| unique | 4 | 4 | — |
| n_unique_delta | — | — | +0 |
| mean_delta | — | — | +0.00735 |
| median_delta | — | — | +0 |
| std_delta | — | — | +0.00287 |
| q1_delta | — | — | +0 |
| q3_delta | — | — | +0 |
| min_delta | — | — | +0 |
| max_delta | — | — | +0 |
| outlier_rate_delta | — | — | +0.00826 |
| skew_delta | — | — | -0.0347 |
report.columns["post_uri"].delta
| stat | curated | firehose | Δ |
|---|---|---|---|
| n | 279,196 | 125,645 | — |
| null rate | 0.0% | 0.0% | +0.0% |
| unique | 232,017 | 103,367 | — |
| n_unique_delta | — | — | -1.29e+05 |
| len_mean_delta | — | — | -0.000684 |
| len_median_delta | — | — | +0 |
| len_p95_delta | — | — | +0 |
| word_mean_delta | — | — | +0 |
| duplicate_rate_delta | — | — | +0.00833 |
| vocab_size_delta | — | — | -402 |
| top-value jaccard | — | 0.00 | |
firehose alt_text is far dirtier and more skewed than curated: duplicate_rate jumps from 0.086 to 0.200 and the median length collapses from 127 to 84 even as the mean climbs to 281.9 and p95 nearly doubles to 1284, indicating a long tail of very long entries alongside many short repeats. Readability also craters (flesch_mean 67.3 vs 22.7) while url_rate (0.070 vs 0.010) and one_word_rate (0.063 vs 0.002) spike. The slices barely overlap on language mix (language_jaccard 0.35) or frequent values (top_value_jaccard 0.03), with firehose introducing ar, ceb, ko, he and others absent from curated.
report.columns["alt_text"].delta
| stat | curated | firehose | Δ |
|---|---|---|---|
| n | 279,196 | 125,645 | — |
| null rate | 0.0% | 0.0% | +0.0% |
| unique | 255,318 | 100,455 | — |
| n_unique_delta | — | — | -1.55e+05 |
| len_mean_delta | — | — | +78.9 |
| len_median_delta | — | — | -43 |
| len_p95_delta | — | — | +582 |
| word_mean_delta | — | — | +2.39 |
| duplicate_rate_delta | — | — | +0.115 |
| vocab_size_delta | — | — | +2.12e+04 |
| top-value jaccard | — | 0.03 | |
| language jaccard | — | 0.35 | |
report.columns["collected_at"].delta
| stat | curated | firehose | Δ |
|---|---|---|---|
| n | 279,196 | 125,645 | — |
| null rate | 0.0% | 0.0% | +0.0% |
| unique | 279,196 | 125,645 | — |
| n_unique_delta | — | — | -1.54e+05 |
| len_mean_delta | — | — | +0 |
| len_median_delta | — | — | +0 |
| len_p95_delta | — | — | +0 |
| word_mean_delta | — | — | +0 |
| duplicate_rate_delta | — | — | +0 |
| vocab_size_delta | — | — | +0 |
report.columns["cursor"].delta
| stat | curated | firehose | Δ |
|---|---|---|---|
| n | 279,196 | 125,645 | — |
| null rate | 15.9% | 0.0% | -15.9% |
| unique | 2,391 | 103,528 | — |
| n_unique_delta | — | — | +1.01e+05 |
| len_mean_delta | — | — | -7.64 |
| len_median_delta | — | — | -8 |
| len_p95_delta | — | — | -8 |
| word_mean_delta | — | — | +0 |
| duplicate_rate_delta | — | — | -0.814 |
| vocab_size_delta | — | — | +1.7e+04 |
| top-value jaccard | — | 0.00 | |
report.columns["langs_json"].delta
| stat | curated | firehose | Δ |
|---|---|---|---|
| n | 279,196 | 125,645 | — |
| null rate | 0.0% | 0.0% | +0.0% |
| unique | 56 | 189 | — |
| n_unique_delta | — | — | +133 |
| entropy_delta | — | — | +0.833 |
| top-value jaccard | — | 0.29 | |
report.columns["image_count_in_post"].delta
| stat | curated | firehose | Δ |
|---|---|---|---|
| n | 279,196 | 125,645 | — |
| null rate | 0.0% | 0.0% | +0.0% |
| unique | 4 | 4 | — |
| n_unique_delta | — | — | +0 |
| mean_delta | — | — | +0.0161 |
| median_delta | — | — | +0 |
| std_delta | — | — | +0.000328 |
| q1_delta | — | — | +0 |
| q3_delta | — | — | +0 |
| min_delta | — | — | +0 |
| max_delta | — | — | +0 |
| outlier_rate_delta | — | — | +0.00408 |
| skew_delta | — | — | -0.0177 |
report.columns["created_at"].delta
| stat | curated | firehose | Δ |
|---|---|---|---|
| n | 279,196 | 125,645 | — |
| null rate | 0.0% | 0.0% | +0.0% |
| unique | 232,017 | 102,375 | — |
| n_unique_delta | — | — | -1.3e+05 |
| len_mean_delta | — | — | +1.17 |
| len_median_delta | — | — | +0 |
| len_p95_delta | — | — | +5 |
| word_mean_delta | — | — | +0 |
| duplicate_rate_delta | — | — | +0.0162 |
| vocab_size_delta | — | — | -436 |
| top-value jaccard | — | 0.00 | |
report.columns["author_did"].delta
| stat | curated | firehose | Δ |
|---|---|---|---|
| n | 279,196 | 125,645 | — |
| null rate | 0.0% | 0.0% | +0.0% |
| unique | 489 | 34,785 | — |
| n_unique_delta | — | — | +3.43e+04 |
| entropy_delta | — | — | +3.77 |
| top-value jaccard | — | 0.03 | |
firehose records carry a url_rate of 0.4649 versus curated's 0.1494, tripling link density and triggering its url_heavy alert. Language mix also diverges sharply (language_jaccard 0.25): firehose adds de, fi, fr, ja, ru, and tr on top of en/pt, earning the multilingual alert, while curated is effectively English-only (396 en vs 2 pt). firehose payloads are longer too (len_mean 1290.7 vs 1136.37, len_p95 2774 vs 2354) with a far more negative readability_flesch_mean of -422.28 vs -155.70, and its one_word_rate of 0.0242 dwarfs curated's 0.0005.
report.columns["raw_record_json"].delta
| stat | curated | firehose | Δ |
|---|---|---|---|
| n | 279,196 | 125,645 | — |
| null rate | 0.0% | 0.0% | +0.0% |
| unique | 232,017 | 103,528 | — |
| n_unique_delta | — | — | -1.28e+05 |
| len_mean_delta | — | — | +154 |
| len_median_delta | — | — | +81 |
| len_p95_delta | — | — | +420 |
| word_mean_delta | — | — | -2.45 |
| duplicate_rate_delta | — | — | +0.00705 |
| vocab_size_delta | — | — | +1.19e+04 |
| top-value jaccard | — | 0.00 | |
| language jaccard | — | 0.25 | |
report.columns["text"].delta
| stat | curated | firehose | Δ |
|---|---|---|---|
| n | 279,196 | 125,645 | — |
| null rate | 0.0% | 0.0% | +0.0% |
| unique | 210,092 | 92,997 | — |
| n_unique_delta | — | — | -1.17e+05 |
| len_mean_delta | — | — | -7.78 |
| len_median_delta | — | — | +17 |
| len_p95_delta | — | — | -4 |
| word_mean_delta | — | — | -5.95 |
| duplicate_rate_delta | — | — | +0.0123 |
| vocab_size_delta | — | — | +7.65e+03 |
| top-value jaccard | — | 0.05 | |
| language jaccard | — | 0.43 | |
The image_thumb_url column is fully populated in curated (null_rate 0.0) but entirely missing in firehose (null_rate 1.0, all_empty alert), a null_rate_delta of 1.0. Curated also shows url_rate 1.0 across 279196 rows with 260636 unique values, while firehose contributes no values at all, yielding a top_value_jaccard of 0.0. The two slices share no overlap on this field.
report.columns["image_thumb_url"].delta
| stat | curated | firehose | Δ |
|---|---|---|---|
| n | 279,196 | 125,645 | — |
| null rate | 0.0% | 100.0% | +100.0% |
| unique | 260,636 | — | — |
| top-value jaccard | — | 0.00 | |
report.columns["image_alt_length"].delta
| stat | curated | firehose | Δ |
|---|---|---|---|
| n | 279,196 | 125,645 | — |
| null rate | 0.0% | 0.0% | +0.0% |
| unique | 1,951 | 2,036 | — |
| n_unique_delta | — | — | +85 |
| mean_delta | — | — | +78.9 |
| median_delta | — | — | -43 |
| std_delta | — | — | +246 |
| q1_delta | — | — | -36 |
| q3_delta | — | — | +58 |
| min_delta | — | — | +0 |
| max_delta | — | — | +1.5e+04 |
| outlier_rate_delta | — | — | +0.0695 |
| skew_delta | — | — | +9.61 |
The author_handle column is fully populated in curated (null_rate 0.0 across 279196 rows) but entirely missing in firehose, which triggers an all_null alert at null_rate 1.0 over 125645 rows. As a result, curated's 489 unique handles and top value 'firefaerie81.bsky.social' (top_rate 0.02446, entropy 7.6765) have no counterpart in firehose, yielding a top_value_jaccard of 0.0. This is a structural divergence: firehose is not carrying author_handle at all.
report.columns["author_handle"].delta
| stat | curated | firehose | Δ |
|---|---|---|---|
| n | 279,196 | 125,645 | — |
| null rate | 0.0% | 100.0% | +100.0% |
| unique | 489 | — | — |
| top-value jaccard | — | 0.00 | |
report.columns["indexed_at"].delta
| stat | curated | firehose | Δ |
|---|---|---|---|
| n | 279,196 | 125,645 | — |
| null rate | 0.0% | 0.0% | +0.0% |
| unique | 232,015 | 103,528 | — |
| n_unique_delta | — | — | -1.28e+05 |
| len_mean_delta | — | — | +8 |
| len_median_delta | — | — | +8 |
| len_p95_delta | — | — | +8 |
| word_mean_delta | — | — | +0 |
| duplicate_rate_delta | — | — | +0.00704 |
| vocab_size_delta | — | — | -389 |
| top-value jaccard | — | 0.00 | |
click to copy
@misc{saturn-bluesky-alt-text-curated-vs-firehose-2026,
author = {Steuber, Luke},
title = {Saturn compare: curated vs firehose},
year ={2026},
howpublished = {\url{https://dr.eamer.dev/saturn/view/bluesky-alt-text-curated-vs-firehose}},
note = {Profiled with saturn-dissect v0.2.0, prompt saturn-insight-v2, model anthropic:claude-opus-4-7},
}
Steuber, L. (2026). Saturn compare: curated vs firehose. Profiled with saturn-dissect v0.2.0 (saturn-insight-v2, anthropic:claude-opus-4-7). Retrieved from https://dr.eamer.dev/saturn/view/bluesky-alt-text-curated-vs-firehose