Audit of the first cloud weekly reading (the reading of 2026-10-19)

Written 2026-10-19 by the session that manages this tracker, before any of the day's other changes landed. It changes no record: besides this file, the only change in the same commit is two sentences added to the standing instruction (proposals 1 and 2 below). The format follows the audits of the first scheduled readings on the sibling sites (The Quiet Galaxy's docs/readings/audit-2026-10-17.md, Outrunning Ageing's and The Upgrade Tree's docs/readings/audit-2026-10-18.md).

What was audited. The commit 3a99138 on main ("2026-10-19: weekly reading, 18 read, 5 rows appended, 2 flagged"), made at 13:04 universal time (06:04 Pacific) by the cloud routine "AGI Zeitgeist weekly reading", its first run; the Mac task it replaced was switched off on 2026-10-18. The commit holds the reading record data/readings/2026-10-19.json, five rows in data/aggregates.csv and 37 rows in data/search_log.csv. It was checked against the design (docs/weekly-reading.md), the standing instruction the routine follows (docs/weekly-reading-prompt.md), the validator, and the sources themselves, each opened again for this audit.

A limit on this audit. The run's own log was not read; this session has no access to it. What the run opened is inferred from its search log, which names every page it fetched, and from its notes, which quote those pages word for word.

The seven verdicts

The five new rows, re-checked

The sample and dates of the AI Impacts survey also match (1,580 respondents kept, fielded December 9 to December 24, 2024). The respondent counts on the rows add up the question groups the report names (471 is printed for one group; the other is named in the row's note).

Noticed, no ruling needed. The AI Impacts report restates the 2023 survey's figure for full automation of labour as 2115, after small changes to its code and data cleaning; the record's 2023 row says 2116, which is what the 2023 paper itself printed. The 2023 row stays as its own source gave it; the 2024 row's note already records the restatement.

The quiet and noted items

Nine quiet and four noted. The quiet ones each give a reason that fits the item (no new model in METR's file; the automation index still at two points; no new Google statement; no open letter, report or government action dated in the window). The noted ones carry their sources. Two weak spots, neither of which put anything into the record:

• The AI releases item cites an aggregator page for the dates of two releases. An aggregator is not a source this record accepts for a row; it is harmless here because nothing was added, but the item should name the companies' own announcements (proposal 2).
• The government-action item did not open a primary page for the Sanders and Casar bill; it searched for the senator's own site but fetched nothing, and says so. Congress's own bill register is the page to open (proposal 1).

The two flagged items, as rulings for me

1. A second Anthropic figure for how much of its own work AI does. Anthropic's page "When AI builds itself" (opened for this audit) says that "as of May 2026, more than 80% of the code we merge into Anthropic's codebase was authored by Claude", and that in a March 2026 poll of 130 research staff the median person estimated about four times the output with the company's newest model. The scored measure for the "AI doing the improving" part stays the company's automation index (26 percent of research work led by its model, August 2026), as I ruled on 2026-10-17.

• The question: should the 80 percent code share be recorded beside that part, shown but not scored?
• Recommended: yes, as a row of the same kind as Google's "75 percent of new code" figure, which the record already shows beside that part and does not score. A company's share of code written by AI is a proxy for its share of research work at best, and the definition already says so. The four-times poll would not be recorded: it measures output, not a share.
• The alternative: record nothing, on the ground that three figures from one company already sit beside the gauge.

2. Does a bill that has only been introduced count as a government action? Senator Bernie Sanders and Representative Greg Casar introduced a bill in September 2026 that would ban superintelligence and create a Department of AI. The record's government actions are orders, laws and declarations that took effect.

• The question: does an introduced bill go into the events list?
• Recommended: no. A bill enters the list when it passes a chamber of a legislature or becomes law, dated by the legislature's own record. Introduced bills are statements of intent by their sponsors and would flood the list. This bill stays out for now, and the reading keeps watching it.
• The alternative: record introduced bills as public statements by their sponsors, dated from the legislature's own record.

Proposed changes to the procedure

1. Open the primary page for anything flagged. A flagged item names the record that would settle it and opens it (for a bill, the legislature's own register, such as congress.gov), so the ruling put to me rests on the record, not a news report. Applied in the same commit as this audit (a sentence in step 3 of docs/weekly-reading-prompt.md); a procedure detail, not a method change.
2. An aggregator is not a source. For an AI release or any dated event, the date comes from the maker's or the institution's own announcement; an aggregator may point the way but is never the source cited. Applied in the same commit (a sentence in step 3).
3. Nothing else. The run already does the two rules The Upgrade Tree's audit asked for: it opened the source of every row it added, and it aimed one search at each pollster's own site.

Items for me

The two rulings above. Nothing else from this run needs him.

Rendered from docs/readings/audit-2026-10-19.md at version a7bd17a.