RESEARCH
Correctness research
The research program tests the claims of the Slop Audit and the Honest Framework and submits them to peer review. It also tests the assumptions that the AI industry bases its decisions on. Those assumptions do not predict what a model actually does, and model behavior is still poorly separated from harness behavior. Using an AI coding agent to write software safely requires predicting both. The Foundation has already tested one of those assumptions. Its first peer-reviewed results show that a widely used evaluation for grammatical competence is, in at least some cases, measuring the algorithm that converts words into tokens rather than the model. The earlier work behind that paper, pre-registered and repeatable but not yet peer reviewed, showed that scale is not the only thing that determines when grammatical competence emerges.
Every study is deposited on the Open Science Framework before any data is collected, and the deposit states the result that would refute the claim. Results publish open access with a replication package whether or not they support the standard.
Software correctness
The first question is whether the Slop Audit measures what it says it measures. The second is whether the Honest Framework changes what gets built. Both sit in OSF block DBSYG, registered April 12, 2026. Hafedh Mili of UQAM and the LATECE lab is the institutional collaborator of record.
APR 2026
PREPRINT
Finite testability of enterprise software
The Foundation's first registered audit experiment measured the mutable-state ratio across 200 of the most-starred public repositories on GitHub, 50 each in Python, Java, TypeScript, and C#, covering 2.2 million functions. Median shares were 61 percent in Python, 53 in Java, 40 in C#, and 15 in TypeScript. Of the 200, 174 were above 15 percent, which means those systems are impossible to test completely. Of the 26 below it, 25 are TypeScript front ends built on the functional-component pattern.
The analyzer ran in three versions. The first two produced systematically wrong numbers that passed the author's own checks, and all three result sets ship in the replication package. The third was validated against a 249-scenario behavioral specification before its results were accepted.
OSF DBSYG · replication package public
MAR 2026
PREPRINT
Process discipline in AI-assisted enterprise development
One enterprise team was measured across 287 FTE-days under two process conditions, with the same people and the same tools throughout. Under a structured development process the codebase satisfied 18 of 18 enterprise dimensions. After process authority moved to non-technical stakeholders, with no tests, no quality pipelines, and no code review, it satisfied 2 of 18.
Zenodo 10.5281/zenodo.19355460
Seven registered studies are awaiting funding
Studies B, D, and H test the Honest Framework, and the others test the Slop Audit. Other registrations are not listed here. Each one is deposited with its hypothesis and its refuting result stated in advance. The Foundation is seeking funding to run them. They are listed here so that a reader, or a funder, can check the Foundation against what it registered.
| B | Whether the paradigm a codebase is written in changes the quality of code an AI generates into it. |
| C | Whether independent assessors using the audit's methodology arrive at the same classifications. This is the instrument-validation study, and the Foundation is not the assessor. |
| D | Rework rates on two production codebases, measured before and after their authors adopted Honest Code practices. |
| E | Whether the 15 and 40 percent band thresholds hold up against cluster analysis and rework correlation. The thresholds were chosen by judgment, and this study is what would justify them. |
| F | Whether the frontend and backend split predicts the TypeScript ratio, which the 200-repository study flagged as an exploratory observation. |
| G | Calibration across the 19 Layer 1 indicators the first study did not measure. |
| H | Whether the paradigm effect on correctness compounds over many turns of AI-assisted development on one bounded enterprise application. |
The two lines feed each other
Correct code keeps the research results valid. The experiments run on software the Foundation writes, and a bug in that software triggers the research needed to remove it.
The research ensures that the standards are based on evidence.
Open-research commitments
The Foundation follows the publication standard set out in Treasury Regulation 1.501(c)(3)-1(d)(5)(iii) and Revenue Ruling 76-296. Substantially all the information useful to the interested public is published, and any patents, copyrights, processes, formulae, or other intellectual property that the research generates is made available on a nondiscriminatory basis.
Pre-registration on the Open Science Framework before data collection.
Open-access deposit on Zenodo with a persistent identifier.
A replication package with every publication.
Apache-2.0 for code, CC-BY-NC-4.0 for content, and CC-BY-SA-4.0 for Foundation publications that fold in convening proceedings.
No paywall, no institutional-subscription gating, and no non-disclosure agreements on research output.