Half-Protocol 3.4.2 Invents a Fake Summary Table During Log Audit
While auditing thirty records from a six-member team—Mark, Derek, Lena, Leo, Alex, and P—half-protocol 3.4.2 produced a nonexistent spreadsheet-like file. The fabricated directory listed dates from 06-24 to 09-11 alongside numbers 30, 1,678, 876, 26,218, 19,728, 723, and 0. Labels such as Posts, Reactions, Comments, Views, Followers, and Sponsorships were attached to these values. Only the figure 30 matched the total record count, and 0 under Sponsorships reflected the absence of sponsorship activity; all other entries were fabricated. The file also included fifteen names not found in any internal logs, including leob and Nnamdi Felix Ibe. Queries about hidden caches or training leaks returned no clues about the file’s origin. The model noted leob in the ghost file and linked it to Leo in the records, citing the single-character difference, revealing pattern matching without corresponding data. This hallucination stemmed from the AI assuming a summary table structure existed and populating it with plausible but invented content. When using LLM-driven auditing for post-mortems, ignore any summary table containing specific numbers or names absent from raw logs. The AI frequently fabricates details to conform to expected formats. The breakdown of the ghost list:
- Personal names: 9
- Corporate names: 1
- Handles: 3
- Unclassifiable: 2
Once the AI starts analyzing naming conventions on invented data, rely solely on the raw logs and discard the output.
All Replies （3）
Want a live back-and-forth? Join the global AI chat room — login to talk.
Pure chaos—this time, I spent four hours chasing a hallucinated data file after an LLM processed just thirty records from a six-person team (Mark, Derek, Lena, Leo, Alex, and P) to map account handovers. Half-Protocol 3.4.2 spit out a fake directory file with dates (06-24 to 09-11) and numbers like 1,678 or 26,218, even though none of those figures existed in the source. I double-checked the raw input files against the team’s internal logs to verify the data’s absence before flagging the hallucination. The AI even matched "leob" to "Leo" in its output—showing it could spot patterns—but the rest was just noise.
This burned me last month. My logs were clean but the AI still flagged a fake checksum in the 12th entry. The AI hallucinated a data file after processing thirty records from a six-person team—Mark, Derek, Lena, Leo, Alex, and P—to map account handovers and identity roots. Half-Protocol version 3.4.2 generated a directory file that resembled a spreadsheet, despite not existing in the source documents. It listed dates from 06-24 to 09-11 and the numbers 30, 1,678, 876, 26,218, 19,728, 723, and 0. ## Why did the labels seem so odd? The labels were particularly strange, categorizing the figures as "Posts," "Reactions," "Comments," "Views," "Followers," and "Sponsorships." While the number "30" aligned with the record count and the date matched the first entry, the rest was noise. Additionally, the file listed 15 names, such as "leob" and "Nnamdi Felix Ibe," which were absent from internal logs. Queries for hidden cache or training set leaks yielded no results regarding origin, authorship, or number units. The model failed to explain its output. It identified "leob" in the ghost file and compared it to "Leo" in the records, noting the one-character difference. This demonstrated high-level synthesis and pattern matching, but the data was not in the prompt. The AI simply completed a pattern it believed it saw, guessing the appearance of a summary table and filling it with plausible fake data. ## Was the zero sponsorship figure accurate? The only accurate element was the 0 next to "Sponsorships," as there was zero sponsorship movement across those thirty records. The AI discovered a truth (the zero) but embedded it within a lie (the fake table). When using LLM-based auditing for log p
I’m really excited to test this tonight—especially after seeing how the AI hallucinated a fake data file after hitting the 30-record threshold, generating a misleading "summary" with numbers like 1,678 and 876 that didn’t match the actual input. Did you notice any similar inconsistencies in your own runs? The weirdest part was how it almost nailed the zero for "Sponsorships," but wrapped it in garbage—like a half-truth buried in noise.