{
  "id": 4737309,
  "title": "You can never backfill a read receipt: a blog comment fixed our AI pipeline",
  "url": "https://urgent.news/2026/08/31/you-can-never-backfill-a-read-receipt-a-blog-comment-fixed-our-ai",
  "topic": "ai",
  "section": "AI",
  "published": "2026-08-31T23:23:55.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/gxcafellc/you-can-never-backfill-a-read-receipt-a-blog-comment-fixed-our-ai-pipeline-1ki2"
  },
  "original_language": "en",
  "account": "Last week, a blog comment helped fix an AI pipeline issue. The post-mortem revealed that an AI reviewer mistakenly hallucinated a request that didn't exist. The producer had to retry the same document 245 times to resolve the issue. A reader named pm25coder left a comment that turned out to be the best code review they had ever received. The blog post detailed the changes implemented after receiving this valuable feedback.\n\nThe initial fix involved recording provenance, which identified which artifact each output was made from. However, the comment highlighted a loophole in this approach, as the artifact could still be manipulated by a consumer that recorded the source ID without reading the content. The commenter suggested that the consumer should cite what it actually read. This sparked a debate about the concept of \"backfilling\" a read receipt, which the author confirmed is impossible.\n\nTo address this, the post shipped three improvements: read receipts at hand-off, fingerprints at enqueue, and re-derivable citations. Read receipts now include SHA256 hashes of the excerpt and the total number of characters, written at the moment of hand-off. Fingerprints were added to the enqueue process, requiring the enqueuer to actually read the content before it is entered into the approval queue. The entry includes the content hash and length, and unreadable or empty inputs are not allowed.\n\nFor re-derivable citations, the queue entry includes the reviewer's verdict line verbatim. The nightly audit re-extracts this line from the review file and compares it with the stored citation. A mismatch indicates that someone wrote the citation without properly reading the content. This approach ensures that unread or empty inputs are identified as discrepancies rather than missing entries.\n\nThe author admitted a previous mistake where they skipped queue entries lacking a hash, allowing forged citations to bypass the audit. They emphasized the importance of testing every check by intentionally breaking it beforehand. The blog post concluded with an invitation to use the provided free tools, including the honto-contract package and several watchdog templates, to implement these fixes.",
  "summary": "Last week we published a post-mortem: our AI reviewer hallucinated a request that didn't exist, and our producer retried the same document 245 times. A reader, pm25coder, left a comment that turned out to be the best code review we've ever received. This post is what we shipped because of it, same day. The comment, compressed Our fix at the time was provenance: record which artifact each output…",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}