The week in AI from the people who actually move the field.

The week in three numbers: 481 million, the transcripts Anthropic screened in a broader search that found no additional incidents of similar or worse severity · zero, the observed violations in 250 controlled runs once an agent's original constraints were preserved, against 62% when they were not · $664 billion, Oracle's contracted work not yet delivered, of which the filing says the company expects to recognise approximately 13% as revenue within twelve months.

This week's question: what would you have to be able to show, and could you?

One question for your team this week: if we had to explain one agent's actions to an auditor, which parts could we produce ourselves?

Still running from earlier issues: the NIST comment deadline on October 15 and the CFTC comment deadline on October 20.

This tracks selected public AI value claims and the evidence behind them. It is not a census of whether AI creates value, and a low rung means our confidence in one disclosure is low, not that nothing happened. The evidence ladder: 0 · Narrative (a story, no numbers) · 1 · Operational (activity counted) · 2 · Financially linked (a number tied to AI, mixed with other causes, or asserted outside reported results) · 3 · P&L-attributed (a reported profit or margin change the company credits to AI, in a filed document) · 4 · Sustained (Stage 3 held four quarters).

This week's ruling. Adobe's release states "AI-first ARR grew more than 150% year over year" and prints no dollar figure, in the text or the tables. The chief executive gave the scale on the call: "now exceeding $650 million." A growth rate tells you direction and pace, not size, and neither disclosure isolates a profit contribution, which is what Stage 3 requires. Computed this issue: $650 million against $27.50 billion total ending ARR is about 2.4% of the book.

Next tests. Oracle's Q2 FY2027 is the biggest, because the filing published a recognition schedule and that makes it checkable rather than rhetorical. Duolingo is the only row within sight of Stage 4: it needs its Q3 and Q4 to hold the attribution, so Q3 is the next step rather than the finish. Every row in this table was re-opened and re-checked at its source this cycle. Four changed as a result, and all four are in Corrections at the end of this issue.

Three stories this week are about the same operating problem in three different forms: a record nobody could query, a summary that dropped the limits it needed to carry forward, and a review whose meaning changed under volume. What they share is that a record only helps if it answers the question you actually have.

Anthropic had the transcripts the whole time. Both searches ran against records that already existed, and the fourth incident was found by neither of them. The work was in the finding, not in the keeping.

The constraint-loss paper names a second version of the problem, in the cleanest form available. When an agent's context is compressed to keep a long task running, "a progress summary often sufficiently preserves task coherence while omitting critical negative constraints." The summary is an accurate account of progress. It is simply not sufficient to govern the next action, because the limits are the part that dropped out. In the paper's compaction test, keeping the control constraints in the summary yielded zero loss of control against 87% when they were dropped. A separate paired test across 250 runs took 62% to zero by restoring the original boundary, while leaving the unsafe action perfectly executable.

The verification story is the same question asked of people. Microsoft's Edge team changed how it assesses quality because submissions outgrew the reviewers. That change might be better than what it replaced, or worse, or adequate for some checks and not others, and nothing published this week settles it. What it does establish is that at least one major platform's review now runs partly without a person reading each submission, and the team's own account says volume is why. So an approval needs to say what was checked, not that something was checked.

Anthropic's second search is the capability this issue says is missing. It ran across 481 million records and produced an answer. So the binding constraint may not be retrieval at all. It may be obligation. Anthropic searched because it had already published three incidents and signed an outside investigator. Most operators have no equivalent forcing event, and some would find, if one arrived, that their records were more answerable than they assumed. If that is right, the thing to buy is not a better archive. It is a standing reason to query the one you have.

That is the distinction worth carrying, because treating them all as documentation hides it.

For an operator the next step is to name the question the evidence has to answer before asking whether you have the evidence. What was the agent authorised to do? What did it actually change? What checks support accepting the result? Then test whether your team can retrieve and keep enough to answer those three.

That is the point of reconstructing one completed task before you widen an agent's authority: not to prove a supplier is withholding something, but to find out whether your own operation can explain the work it is already delegating.

Four Margin-Proof Tracker rows published in Issue 012 were wrong. Each is corrected below against its source.

How we label evidence: Primary source · Corroborated · Reported · Vendor claim · Analysis. Written and edited by Mario Suarez · Independent analysis · Where we could not open a source, this issue says so rather than implying coverage.

A weekly decision brief for business leaders, about what AI changes in a business rather than how it is built. Every week we ask one standing question of the news: when the dust settles, who actually keeps the money? We grade each claim by the strength of the record behind it, and link to primary documents or to clearly identified reporting.