A&E

Most published research is never checked. Re-running a study by hand is slow, skilled work that is rarely funded, so it happens late and to a handful of famous papers. When it is done, many results fail: the Open Science Collaboration re-ran 100 psychology studies, and 97% of the originals were significant against 36% of the replications (Science, 2015). Software solved the same problem with continuous integration: nothing merges until the tests pass. celsus does that for papers. The goal is a public library of checked papers, with the script, the numbers and the verdict open to anyone who relies on them. celsus is an AI workflow that replicates scientific research. Give it an arXiv paper and it reads the LaTeX, writes down every number the paper reports, rebuilds the analysis from public data in a sandbox, and tells you in plain words whether the published numbers hold up. It runs on Meta's Muse models. Muse builds the replication as an agent inside a Modal sandbox, with tools to run commands and write files. Only a clean run of its final script, in a fresh sandbox, counts. Code, not a model, judges every number, with tolerances from published replication practice. Coded guards stop typed-in results, samples trimmed or padded to a count, and fetching the authors' code, and every script is reviewed twice, independently. We tested it on 32 finance and economics papers from arXiv that nobody had looked at, drawn from 6,552 arXiv hits by a rule written down before any run. Each paper ran twice on Muse Spark Open. The 63 finished runs came out: holds up 1, partly holds up 26, does not hold up 3, could not check 28, could not check at the $2 budget stop 5. "Could not check" is the checker's own failure, given with its reason, and never counts against the paper. To test the checks, a sabotage battery plants faults into clean runs. Of 116 viable plants, the coded checks alone catch 94, and with two reviews added, 108. The reviews raised no false alarm in 12 reviews of untouched scripts. Before any change merges, a regression gate re-scores every stored run (215) and re-checks every stored plant (130). Try it with no API key: uv sync, then uv run celsus demo. The video shows a live run on Muse Spark from the typed command to the verdict: arXiv 2604.27287 in 3:20 for $0.24, with 23 of 25 checked results matching. All 67 held-out runs are in receipts/, and docs/USAGE.md goes from install to reading a report and checking it yourself. The idea came from a March run, before this event, that scored 374 arXiv papers for replicability: 20 could be run again in full, and 3 were replicated by hand. None of that code is in this repository. Not financial advice. It checks what academics published. AI can make mistakes: check the code and data in the report.