About e2er
e2er is an open infrastructure that makes research objects reproducible, composable, and attributable, extending scientific inquiry to the age of AI agents.
Creating research
Templates describe research processes as a sequence of steps. Each template organizes specialists, i.e., AI agents with defined roles, and each specialist draws on skills, i.e., written descriptions of methods. Researchers can adjust any template, specialist, or skill and publish their version for fellow researchers.
Researchers decide at which steps the process pauses for their review.
A study produced with e2er comes with a dossier that helps others verify and reproduce the work and attributes everyone whose templates, specialists, and skills it used.
Verifying research
Before a study is published, e2er checks it against its dossier. The dossier lists every file of the study with its hash value. The check confirms that none of these files has changed since the specialists produced them, and that every table in the study can be rebuilt from the estimation results the specialists saved.
The check is built into e2er, so anyone who has the study's files can repeat it with the hash values in the dossier. e2er.org repeats the check when the study's files are in a public repository.
Examples:
The Bitcoin study (its dossier) has its code and study files in the public repository github.com/bhanneke/E2ER-project (folder examples/showcase), and e2er.org checks it against them. Of its data, the repository holds the data dictionary and a summary.
The FOMC study (its dossier) keeps its files private, so its dossier shows only the check its author ran.
The dossier lists every file of the study with its hash value. The check computes each hash value again and compares the two. The 17 changes made to the abstract and the paper after the export are recorded as amendments.
73 of 73 files match
[PASS] integrity 73 files hash-verified against provenance.json; 346 table cell(s) traced in the derivation graph; 17 change(s) after export recorded as amendments (paper/abstract.tex, paper/paper.pdf, paper/paper.tex); the abstract rewritten after export by e2er's abstract step (paper/abstract.tex, paper/paper.pdf, paper/paper.tex, report.html)Every value in the tables of the study is looked up in the estimation results the specialists saved. For each coefficient, the check computes its p-value again from its test statistic.
290 table values found in the estimation results
48 of 48 coefficients consistent
In the text, 244 of 245 numbers are values from the results; 1 differs from the quantity named next to it. The text does not decide the check.
[PASS] numbers 290 table cell(s) are source values at the precision shown; 48 coefficient(s) checked, 0 inconsistent; text (not gating): 244 of 245 number(s) are source values, 1 differ from the quantity named nearby, 0 untracedThe check rebuilds each table from the saved estimation results. The rebuilt table is compared with the table in the study.
13 tables rebuilt identically
2 more tables in the paper were not built from the saved estimation results, so the check cannot rebuild them
[PASS] tables 13 table(s) reproduce byte-identically from the sidecars; 2 not renderer-produced (rawcontrast.tex, sumstats.tex)Before the estimation, the specialists declared how the effect is identified. The check compares the estimation with that declaration.
[PASS] spec estimation implements the declared identificationThe check looks up every citation key of the study in its list of references. Each key found there is verified against its source.
33 citation keys found in the list of references
33 citation keys verified
0 citation keys could not be verified
[PASS] citations 33 cite key(s) resolve in refs.bib; snapshot: 33 verified, 0 unverifiable at run timeNewer runs record which script wrote each results file. With that record, the check confirms that every results file is what the study's script wrote. This study was run before e2er kept the record, so the check is skipped.
[SKIP] results from code not checked: the run predates output recordse2er.org repeats the check when the study's files are in a public repository.
Its code and study files are in a public repository.
✓Checked by e2er.org: 73 of 73 files match
its dossier (version 4)The files are private.
Its dossier shows only the check its author ran.
its dossiere2er verify on the Bitcoin study Demonstration, version 4, one step at a time. The result lines are the real output of e2er 0.18.0.Reproducing research
The replication template reproduces a published study from its replication package on Zenodo. Specialists read the package and plan which of the study's numbers to compare. The researcher reviews that plan before anything is executed.
The authors' code is then executed again in a container on the researcher's computer, without network access. Every number the study reports is compared with the number from the new execution and labelled reproduced, reproduced with a minor difference, or not reproduced. The researcher reviews the report, and the dossier records every comparison.
Example:
The replication of Peixoto and Leite, "Political Rupture and Selective Damage to Local Service Delivery: Evidence from Brazil" (its dossier): 12 of 17 numbers reproduced, 2 with a minor difference, 3 not reproduced. All five differences are in numbers from the Callaway and Sant'Anna estimator: two estimates, one standard error and two minimum detectable effects. The replication is not an assessment of the authors' work; differences can come from e2er or from software versions.
The replication package is downloaded from Zenodo. It is checked against Zenodo's checksum and unpacked read-only.
97 files unpacked, read-only
The replication planner reads the package. It plans which of the authors' scripts to execute and which numbers to compare.
The numbers are taken from the result files in the package. The linked preprint has a different title, so the numbers printed in the paper were not compared.
Level 1 only. Keep the 16 packages.…
- sent back 2026-09-28
- sent back 2026-09-28
- sent back 2026-09-29
- approved
The researcher sent the plan back three times with written instructions, then approved it. Every decision is recorded in the dossier.
The authors' code is executed again in a container on the researcher's computer. The container has no network access.
- image
- rocker/r-ver:4.6.1
- network
- none
- replication package
- read-only
- R packages
- as of 2026-08-30, the day the package was published; did 2.5.1, DRDID 1.3.0
- processor
- aarch64; the authors used x86_64
6 of 6 scripts finished without errors
The reproduction comparer compares each planned number with the number from the new execution and labels it.
12 reproduced
2 reproduced with a minor difference
3 not reproduced
All five differences are in numbers from the Callaway and Sant'Anna estimator: two estimates, one standard error and two minimum detectable effects. They are marked with a dashed outline.
The check reads every number of the report again from the files of the new execution. Three of the 17 numbers:
The researcher reads the report and approves it. Every decision is recorded in the dossier.
Dossier
source: Zenodo 10.5281/zenodo.22179151Get started · Install · The Bitcoin study Demonstration · its dossier · The FOMC study Demonstration · its dossier · The replication Demonstration · its dossier · Documentation · How e2er works in practice · Source on GitHub