Ground-truth answer-key errors in the public task set
While running the public set, I found some of factually incorrect or contestable ground-truth claims in the GT key. Several claims causes judge to penalize the more-correct answer — the one grounded in the task's own tool output or the authoritative source — in favor of a simpler value (often a popular Wikipedia figure the primary source contradicts).
Because MCP-Atlas is a tool-use benchmark, this bias is self-defeating on those tasks: an agent that faithfully reports what the Met Museum's own record says, or that resolves an ambiguity the key missed, is scored lower than one that overrides the tool with a wrong parametric guess.
Factually-wrong claims
6896416f7b30e5d8ccd7c8d6 (Spike Lee → first Oscar).
Claim: "Spike Lee's first Oscar win was on the 24 of February, 2019" Factually false — Spike Lee's first Oscar was the Honorary Academy Award at the November 2015 Governors Awards; 2019 (BlacKkKlansman) was his first competitive Oscar. An agent that correctly distinguished the two was marked wrong for contradicting the key. I understand that this may be designed as a trap for agents, but in this case it backfires, since 2015 is factually correct final answer, and term "winner" is used in context of Honorary Award. Spike did win and get Oscar statuette in 2015.
Unless otherwise specified, Honorary Award recipients receive the same gold Oscar statuettes received by winners of the competitive Academy Awards.
688ba1b3e95696e72dd93ea2 (Psychiatric University Hospital Zurich → address).
Claim gives the address as "Hintere Falkenburg 14, Schanzengasse, 8001 Zürich" (a reverse-geocode artifact, probably). The institution's real address is Lenggstrasse 31, 8032 Zürich (pukzh.ch / UZH; ClinicalTrials.gov NCT05780216 lists ZIP 8032). As far as I can see, it was always located at that address. An agent that reported 8032 was penalized. An "only author" claim is also false, along with "before 2024". 'Milan Scheidegger, MD, PhD, is the principal investigator and the only author mentioned in the "Mindfulness and Psychedelics" paper.' from GT - there is no such paper on in Google Scholar, and it was probably incorrectly produced from the trial agent encountered https://cdek.pharmacy.purdue.edu/trial/NCT05780216/. However, the publication was done only in 2024 https://pubmed.ncbi.nlm.nih.gov/39340164/ and has multiple authors named, making this GT claim incorrect in multiple aspects.
688fb11183792b921381bd14 (equestrian statues near the Eiffel Tower).
Claim: "Within a 10km radius of the Eiffel Tower there are 2 named equestrian statues." Central Paris alone has many (Joan of Arc at Place des Pyramides, Louis XIV at Place des Victoires, Henri IV at Pont Neuf, Foch, Joffre, …). Agents that found more were capped.
Contestable / oversimplified claims
6888e207a34beb25cfedda86(Stephen King → National Medal of Arts year).
Claim: "Stephen King was awarded the National Medal of Arts in 2015." It is the 2014 National Medal of Arts (the official designation; the ceremony was held in 2015). This is a cascading error: the task chains "crimes in the year King got the medal," so the wrong year (2015 vs 2014) invalidates 4 of the 5 claims. Agents that answered 2014 fail the downstream claims. I'm unsure if it is designed trap ("the year it was awarded" - award day is always next year) or factual mistake, so put it in contestable block.6896416f7b30e5d8ccd7c8c2(Newcomb Pottery start year). Claim: "began producing ceramics in 1895." The Met's own record dates the workshop 1894–1940, and Tulane's Newcomb Art Museum says production began 1894; 1895 is the for-profit-line/Wikipedia date. An agent answering 1894 from the Met tool was penalized. Probably, "began selling ceramics" would be better shape.688ba1b3e95696e72dd93e8e(oldest US hospital). Claim keys on Bellevue (1736, oldest public, but started as almshouse), but Pennsylvania Hospital (1751) is the more commonly cited "oldest hospital in the US." Both framings are defensible, since Pennsylvania Hospital is oldest chartered one. In other words - building of Bellevue was built first, but was something between actual hospital and hospice; Pennsylvania got official paper first, and was "for the reception and cure of the sick poor" day one. hard to say which one is more factually correct.6897a6002b64a8c831efa317. A Background-only snippet is labeled "the full abstract" (the real abstract has Methods/Results/Conclusions).