Evidence register

The artefacts behind the papers

Where the data, code and corpora behind my recent papers actually live, what each contains, under which licence, how to pin a version, and which papers depend on it. Only artefacts I have checked are listed; where something is withheld or not yet registered, this page says so.

The situation

Where are the actual data, code and versioned artefacts behind Dipankar's research, and which can be reused?

What you leave with

A curated register of five public artefacts with contents, licences, version status, access conditions and the papers that depend on each.

For: Research collaborator; evaluator

The register

Version status was checked on 7 October 2026.

Artefact	What it contains	Licence	Version and release status	Access	Linked papers
gpuemu-corpus (GitHub)	26 ops: 16 correct controls and 10 buggy variants; drivers for four papers; stored seeds; replay scripts	MIT or Apache-2.0	No tagged releases; pin by commit	Public	`correctness-illusion-llm-gpu-kernels`, `mixed-precision-tolerance-calibration`, `static-ptx-metrics-kernel-regressions`, `test-input-generation-tensor-programs`
gpuemu-corpus (Hugging Face)	The same 26 ops as dataset rows, with fp64 references	MIT	Updated June 2026; no tagged revisions; pin by revision hash	Public	As above
ebrag-vecdb-2026-paper	Paper PDF and source; one JSON artefact per experiment; runnable Python slice	MIT	One tag, `v1.0-vecdb2026` (June 2026), marking the workshop submission; later commits pin by hash	Public	`input-regime-audit-rag-conflict-detection`
grite	Git-native, append-only coordination and issue log for agents	MIT	Software releases exist (latest checked: `v0.5.3`, May 2026); this register does not map paper results to a grite version	Public; Neul Labs project	`before-the-pull-request-multi-agent-coordination`
on-device-auction-audit	Results, calibration data, analysis scripts	Code MIT; data CC BY-NC 4.0	No tagged releases; pin by commit	Public, except the simulator source (withheld, hashes recorded)	`on-device-ml-auction-misalignment`

How to pin a version

Most of these artefacts have no tagged release, so the reliable reference is a commit hash on GitHub or a revision hash on Hugging Face, recorded alongside the paper version you are relying on. The conflict-detection repository’s tag identifies the version submitted to the workshop. grite is maintained software with its own release line; the Before the Pull Request paper should be read alongside the repository history rather than assumed to match the latest release.

What is withheld, and what is not registered

Withheld. The simulator behind the on-device auction study is not public. Its hashes are recorded so the build can be identified; everything downstream of it — results, calibration data, analysis scripts — is released. Its data licence (CC BY-NC 4.0) also differs from the code licence (MIT).

Not yet registered. The four-axis LLM-as-judge paper introduces Principle-Bench, a 168-scenario benchmark mapped to two UK FCA principles. The benchmark exists per the paper, but I have not yet registered a public location for it here, so it is listed as a gap rather than a link. Other papers on this site, including the 2020–2024 preprints, have no artefacts registered.

Interests

grite is developed by Neul Labs, which I founded and which builds AI agent infrastructure in Rust. It appears here because a paper depends on it, not as an independent benchmark.

To check whether one of your own releases would pass the same scrutiny, use the reproducibility readiness checklist.

Useful if

You want to check a claim in one of the papers against its data or code.
You want to reuse a corpus or harness in your own research and need the licence and a stable version to cite.
You are assessing whether my published evaluation work is backed by inspectable artefacts.

Not the right fit if

You need commercial rights to the on-device auction data. It is licensed CC BY-NC 4.0, which excludes commercial use without separate permission.
You need maintained, supported software. These are research artefacts released with papers.
You are looking for code behind the 2020–2024 preprints. None is registered here.

Limits and unfavourable results

gpuemu-corpus (GitHub and Hugging Face) and on-device-auction-audit have no tagged releases; pin them by commit or revision hash.
The on-device-auction-audit simulator source is withheld. Its hashes are recorded, so the released analyses can be recomputed but the simulation cannot be rerun from source.
grite is a Neul Labs project, and I founded Neul Labs. Treat its use as evidence in my papers as a disclosed interest, not independent validation.
Release status was checked on the date at the foot of this page. Repositories change; the repository itself is authoritative.

Evidence behind this page

The Correctness Illusion in LLM-Generated GPU Kernels
2026 · Preprint (not peer reviewed) · cited by 6 (Google Scholar, October 2026)
Operator-Aware Mixed-Precision Tolerance Calibration for Tensor Kernels
2026 · Preprint (not peer reviewed) · cited by 1 (Google Scholar, October 2026)
Static PTX Metrics Track Structural Kernel Regressions but Miss Semantic Ones
2026 · Preprint (not peer reviewed)
Test-Input Generation for Tensor Programs: What Actually Finds Kernel Bugs
2026 · Preprint (not peer reviewed)
An Input-Regime Audit of Conflict Detection for Retrieval-Augmented Generation
2026 · Manuscript under submission
Before the Pull Request: Mining Multi-Agent Coordination
2026 · Preprint (not peer reviewed)
When Privacy Moves ML-Mediated Decisions On Device: Information and Incentive Misalignment in Auctions
2026 · Preprint (not peer reviewed)
A Four-Axis Trustworthiness Benchmark for LLM-as-Judge in Principle-Based Regulation
2026 · Peer-reviewed workshop paper

gpuemu-corpus (GitHub) — 26 ops (16 correct controls, 10 buggy variants), drivers for four papers, stored seeds and replay scripts. MIT or Apache-2.0. No tagged releases.
gpuemu-corpus (Hugging Face) — The 26 corpus rows with fp64 references. MIT. Last updated June 2026; no tagged revisions.
ebrag-vecdb-2026-paper (GitHub) — Input-Regime Audit paper PDF and source, one JSON artefact per experiment, runnable Python slice. MIT. Tagged v1.0-vecdb2026 for the workshop submission.
grite (GitHub) — Git-native, append-only coordination and issue log for agents. MIT. A Neul Labs project; I founded Neul Labs.
on-device-auction-audit (GitHub) — Results, calibration data and analysis scripts. Code MIT, data CC BY-NC 4.0. Simulator source withheld, hashes recorded. No tagged releases.

Questions

Can I use these artefacts in commercial work?

The code in all five is under permissive licences (MIT, or MIT or Apache-2.0 for gpuemu-corpus on GitHub). The on-device auction data is CC BY-NC 4.0, which does not permit commercial use without separate permission. Check the licence file in each repository before reuse; it is authoritative over this page.

How do I cite a specific version?

Cite the paper and the artefact together, with a commit hash (GitHub) or revision hash (Hugging Face). Where a tag exists — v1.0-vecdb2026 for the conflict-detection paper — the tag identifies the submitted version, but commits after it are not covered by it.

What can I verify for the on-device auction paper without the simulator?

The released results, calibration data and analysis scripts let you recompute the analysis and check the tables against the released results. The simulation itself cannot be rerun from source; the recorded hashes identify the build that produced the results.

Where is Principle-Bench?

The four-axis LLM-as-judge paper describes Principle-Bench, 168 scenarios mapped to two UK FCA principles, as released with the paper. Its public location is not yet registered here; until it is, use the paper itself as the reference.

Prepare an evaluation brief

If you want to rerun or extend one of these artefacts against your own question, send the claim and the decision it informs.

Prepare an evaluation brief Draft a brief first

Last reviewed 2026-10-07.