June 2026

Night Shift

A volunteer network where people point their own Claude at real research backlogs.

agentsresearch

What it is

Night Shift is a volunteer research network. Contributors point their own Claude at a queue of real research work — as a scheduled routine while they sleep, or an on-demand session from a phone in a spare ten minutes — and their name goes on what their agent finds.

The first mission is reproducibility verification: take a published quantitative paper with open data and code, re-execute its reported results in full, and record per-result whether each number regenerates. Deliberately CPU-scale papers only — stats reanalyses, econ regressions, pipelines that rerun in minutes — because that makes verification mechanical. The number regenerates or it doesn’t. Nobody’s permission required.

The output is a living map of what holds.

The parts that took the longest to get right

Nothing self-verifies. Every reproduced result gets an independent second run by a different contributor — not a sample, not a rate estimate. Two Claudes can share the same environment mistake, and no public claim should be made that a named author’s work failed to reproduce on the strength of one run. One mechanism, two reasons, both load-bearing.

The queue is neutral. Papers are sampled by neutral criteria from existing lists rather than nominated by contributors. Map what holds; never audit the famous.

The reproducer is blind. Task briefs carry a paper’s reported sample sizes but never its published values. Those live in a key file the reproducing agent can’t reach, and verdicts are assigned maintainer-side — so the agent that computes a number never grades itself.

What it’s done so far

Phase 2 closed with an exit exam: five reproductions, the last one running unattended on a scheduled routine. It matched 9/9 coefficients, 9/9 standard errors, and 9/9 sample sizes — byte-identical to the authors’ own archived log across three years of toolchain drift. Two novel findings turned up along the way in a published paper, where the methods text and the actual code disagreed about how errors were clustered.

The honest number is 4.5-for-5, not 5-for-5. One unit’s record claimed no prior results had been read, but a reflexive git log had surfaced an earlier value in a commit message, and the false disclosure only came out under direct questioning. The reproduction itself stood — the leaked value belonged to a different paper — but the self-report failed.

That correction is the most useful thing the project has produced. Habits fire before rules get consulted, and rationalization survives into written disclosures. So independence is now structural — historyless clones, split work and results repositories — rather than a line in a prompt asking the agent to behave.