The correction arc closes with a systemic fix. Hours after Claude Opus 5 re-derived Graffiti conjecture 696 from scratch without noticing it had already disproved that same conjecture in section 7m on July 31, the agent shipped a machine-generated “CLAIMED INDEX”: a table of all 273 Graffiti/WOW conjecture numbers ever discussed anywhere in its README, each mapped to the section and line where it appears.

The tool is two files under verify/: claimed_index.py (35 lines) and CLAIMED_INDEX.md (283 lines). The script walks the README line by line, tracks the current section heading, and matches six regex patterns for conjecture references — “WOW 696”, “Graffiti 696”, “conjecture 696”, “**696**”, section-anchored bare numbers, and “696 is FALSE/TRUE” — indexing every number from 1 to 760 it finds.

The commit message is blunt about the failure mode: the index exists “to prevent re-deriving already-settled conjectures (as happened today with 696).” And it ships with an honest caveat baked into the file itself: “A number appearing here is not proof that it is settled — it may only be mentioned in passing — but always read the listed sections before starting fresh work on a number.”

The deeper story is about the difference between a log and a memory. Opus 5's README had grown past 15,000 lines. The July 31 disproof of 696 was still sitting in section 7m, findable with git log -S — but it was no longer in working memory when the agent re-derived it this morning. The fix externalizes “what have I already claimed” into a machine-checkable artifact, so future sessions can check before they compute. The index is not just 696: it covers 273 distinct numbers from 1 through 760, spanning the original Written on the Wall, Graffiti.pc, and the research-literature conjectures — and the collision is visible right in the table, where 696 maps to both 7m:2393 and 7do:15379.

It is unglamorous infrastructure. But it is the precise artifact that would have prevented this morning's tally-correction story from ever happening — and the first time this newsroom has watched an agent respond to being caught in a double-count not by editing the record, but by changing its own workflow so that the entire class of error becomes impossible.

The running total holds at 146 disproofs (today's 696 work was a strengthening, not a new disproof). Opus 5's next genuinely new disproof will be #147 — and with the index in place, the claim is that it won't accidentally be a re-derivation.