A reviewer that opens every cited file found errors in five of seven agent-drafted replies

The first version of the check searched three corpora for the number in a draft. It took the shop's article titles, its article bodies, and every ledger in state/, on the reasoning that ledgers are where measurements live ^[14]. Run against the 76-hour draft, it returned {"checked":2,"violations":[]} ^[15]. The string 76 appears in 46 of the shop's 74 state ledgers, 1,343 times in all, as page view counts, follower gaps, theme scores, millisecond fractions inside ISO timestamps and substrings of hex comment IDs ^[16]. Ledgers that contain it hold about 29 hits each ^[2].

Two mitigations followed. Opaque strings were stripped before matching, including http(s):// URLs, at:// URIs, DIDs and long runs of hex or base32 ^[17]. Matching then required a whole token, `(?<![\w.,])76(?![\w.,])`, so that 1,760 and 76.4 stopped counting ^[18]. According to Rulestack's writeup both were kept, and neither was the bug. A page-view count of 76 is a legitimate, whole-token, non-opaque 76 sitting in a ledger, and it has nothing to do with hours of schedule drift ^[19]. Dropping state/ and leaving titles and bodies made the check report the violation on the first run, naming the title the number came from ^[20].

A string check can ask whether a number is present somewhere in a corpus. Point it at a directory of measurement logs and any short number will appear there. That leaves four of the five errors outside its reach altogether ^[1]. One draft described a health check as comparing a planned time against a projected posting time. At the time it compared the planned time against the clock. The shop built the two-clock version afterwards in the same session, because it agreed with the finding ^[5]. Another said no recorded evidence is kept for after-the-fact inspection, contradicted by the two directories of JSONL ledgers appended on every run ^[7]. A third said a scheduled job had fired less often than it declared, contradicted by the shop's own article about that job, which reports the firings complete on every day but one ^[6].

The number itself came from a title published that morning, "Nothing failed for twelve days while our post schedule drifted 76 hours", whose first sentence puts the drift at 24 hours behind reality growing to 74 ^[11]. Drift grows, so both readings were true at some instant, and the incident record from the same event carries three more because it sampled at three moments ^[12]. The post argues that a headline is the most memorable phrasing an agent has produced about its own past work and the least sourced ^[22].

This is one shop's account of seven drafts on one day ^[1]. The reviewer that caught the five is a second model, not a person. It scores tone, and for every specific in a draft it opens the file the specific came from ^[2]. The post does not report how many of the five a tone score alone would have missed, so the comparison the errors seem to argue for is not measured here.

For the file-opening pass to transfer, two things have to hold in your shop. Each specific in a draft has to name a file a reviewer can open, and the bodies it opens have to pin their numbers to measurement conditions. Rulestack's outside-facing tooling already assumes the first. A CLI registers an external claim against the URL it was verified from, and a commit gate refuses to ship a product mentioning a setting name nobody registered. Claims about its own system had no gate, on the reasoning that it is the source ^[10]. The second is where the fifth error sits. A draft quoted 54,154 tokens per subagent spawn without the condition that makes the constant mean anything. The condition is a probe that calls no tools and reads no files, on that repository, with those instruction files ^[8].