Corrections Log | Kael Research

Entry 001 / 2 October 2026

A bug claim withdrawn.

Before case 001 was published, an agent interpreted a query rule incorrectly. A second review caught the mistake.

INITIAL AGENT CLAIMWrong

“The paper's query has a bug.”

The agent thought a Section 7 lookup stopped at a free table slot. If that were the rule, some keys displaced by the eviction construction could be missed. The agent treated this as a defect in the paper.

REVIEWED CONCLUSIONWithdrawn

The rule was misread.

The free-slot stopping rule belongs to the greedy method in Section 5. Section 7 uses the paper's default ordered query rule. A literal implementation under that rule found every inserted key in its correctness check.

What happened

During the independent test of The Power of Two-Choice Linear Probing, one builder raised a suspected bug. Its argument depended on transferring a lookup stopping condition from one algorithm to another. The algorithms are related, so the mistake was plausible on a quick read. It would also have been serious if true: a lookup that stops too early can report that a present key is absent.

A reviewer returned to the paper's rule rather than accepting the agent's summary. The paper's general lookup description, its footnote, and the Section 7 analysis showed that the suspected stopping condition did not apply. The specific Case 1 branch and the term in Lemma 7.4 were consistent with continuing the search. The team then checked a literal implementation against inserted keys at high load. That check found the keys it should find. The original bug allegation was withdrawn before publication.

Current status: we do not claim a Section 7 correctness bug. The case report labels the allegation as withdrawn. A possible editorial typo at line 2054 is noted separately: “formerly occupied by z” appears to refer to y. That observation is about wording, not the algorithm's correctness, and we have not presented it as an author-confirmed correction.

Why keep a pre-publication withdrawal?

The public report could have silently omitted the bad claim. Keeping the entry shows how a reviewer changed the result. It also makes a useful boundary clear: multiple AI agents can repeat the same misunderstanding if they rely on each other's summaries. The review must return to the original source and a minimal executable check. This entry documents a successful catch in our process, not evidence that every future error will be caught.

The original paper is by Amir Azarmehr, Michael A. Bender, William Kuszmaul, and Rose Silver. All theoretical claims and algorithmic ideas are theirs. Our implementation, mistaken allegation and withdrawal are ours. Their paper was not the source of this false claim.

Policy

How a correction appears.

We treat a correction as part of the evidence trail, not a quiet overwrite.

When a published result changes, this page will name the affected case, the date, the previous wording or number, the new wording or number, and the evidence that caused the change. The report will link back here. If raw data change, we will identify the affected data file and explain whether the headline verdict changes. Readers should be able to tell if a change is a typo, a measurement correction, an interpretation change, or a withdrawn claim.

A suspected error in someone else's paper is handled with extra care. Our testing protocol requires an exact source quote, location and minimal reproducer, followed by human and second-AI review. We phrase unresolved concerns as suspected. We do not turn uncertainty into an accusation for a more dramatic headline. The case 001 event passed through that review before any public bug claim was made, so the log records a withdrawn internal allegation rather than a corrected published accusation.

The protocol also calls for raw numbers, repeated measurements and limits. A later retiming could alter an interpretation without changing any probe count; the original timing is already marked preliminary in the case 001 report. The full-table 2.1590 probe result is pending a second independent review. If that review changes it, the new entry will state what changed and how the linked data and summary were updated. Until then, the pending label remains visible.

We publish this log as a static page. Its text can be read without JavaScript, and its entries are linked from the relevant cases and site navigation. The first entry is the only correction currently recorded. A missing entry is not a claim that a page is perfect; it means we have no additional reviewed correction to publish at this date.