Introduction: Current collections of tumor cell lines, and their genomic, proteomic, and phenotypic reference data, have been vastly important to cancer research. Given the centrality of these datasets, it is paramount that they be of the highest quality. Here we show that up to 20% of mutations identified within these collections are erroneous, owing to sequencing artifacts, misaligned reads, in addition to other sources of error. To achieve a high-quality set of mutation profiles for all 329 tumor cell lines sequenced by the Cancer Cell Line Encyclopedia (CCLE), we processed all genomes using the consensus of four complementary mutation identification tools, yielding an online resource we call Quartet. We demonstrate considerable benefit from its use compared to currently used mutation profiles across a variety of tasks. Methods: We retrieved raw sequence data previously generated by the Cancer Cell Line Encyclopedia. We filter for short sequences that disproportionately contribute to noise in mutation profiles, identify candidate mutations using 4 mutation identification tools, and then filter those candidate mutations to the subset with agreement from at least 3 tools. Using a HCC1395 cancer cell-line sequenced at ultra-high depths as a positive control, we attribute roughly 1/5 of mutations originally generated by CCLE to technical artifacts and demonstrate enrichment of True-Positive mutations from Quartet. These artifactual mutations persist at the gene-level, where we observe divergence between predicted variant effects from our ground truth and those expressed in CCLE, a divergence which is significantly recovered through Quartet. When aggregated across all cell lines and viewed with respect to Variant Allele Fraction (VAF) in original CCLE mutations, we observe a bimodal distribution of variants whose lesser mode is enriched for technical artifacts which Quartet effectively filters. We demonstrate these artifactual mutations cannot be removed post-hoc through either VAF thresholds or Variant Effect Prediction, nor through use of alternate databases such as DepMap. We then leverage this recalled CCLE dataset to train pharmacogenomics models and show significant improvement to predictive performance across 10 anti-cancer drugs. Conclusion: Quartet presents a public resource of cancer cell lines which use consensus support of mutation callers. Quartet demonstrates depletion of artifactual mutations and their deleterious effects in a variety of downstream tasks. These results highlight the potential to enhance cancer biological discovery through Quartet, and we expect these benefits to compound as additional cell lines are incorporated. Quartet data are made publicly available through zenodo and source code is made available at https://github.com/digitaltumors/quartet.