Job search
Use a Zenodo DOI for a Research Software Work Sample: Choose the Right Version
Two actual local software versions show why a fixed citation, input identity and reproduction packet must agree. No DOI was assigned or reserved.
Use a specific-version Zenodo DOI when your research-software work sample claims a particular implementation or result. A concept DOI describes the software across versions and leads to its latest record. Neither identifier supplies missing dependencies, authorizes sharing private code or proves that someone can reproduce your computation.
We built and executed two small local software versions to show why that distinction matters. They process the same five synthetic measurements but apply different QC policies, producing different group means. Their complete version map and application-review packet are saved locally. Neither version has an assigned or reserved DOI; no Zenodo upload, publication or account action was performed.
LandOffer publishes this guide. Compare the responsibilities in recent research-software and data roles with the work you can personally explain. The example is AI-authored teaching software, not a candidate's observed research contribution, a published study or a hiring outcome.

Decide whether you mean one result or the evolving project
Zenodo's current versioning FAQ distinguishes the DOI for a particular version from the concept DOI covering all versions. It normally recommends the specific version for citations; the concept DOI resolves to the latest version's landing page. Use the former when an application discusses a fixed result, and the latter only when you deliberately mean the evolving project. Zenodo DOI versioning FAQ
Suppose a reviewer reads “this program excludes measurements rejected by QC.” A link that later opens a different implementation can obscure which policy produced your reported result. Check whether the landing record, downloaded files and explanation identify the same artifact. A project name or citation badge does not identify which implementation produced the result.
Do not construct a new DOI by appending .v2 to an existing one. The FAQ explains that a version suffix is not a DOI-system feature that selects another release. Retrieve the identifier actually associated with the intended record. A software label such as 1.1.0 and a repository identifier have separate jobs; record both rather than assuming one encodes the other.
Before choosing any public route, establish permission to share the code, included data and documentation. Employer ownership, collaborator agreements and confidential research inputs require their own review. A DOI does not settle those questions. An unpublished, accurately described local sample is preferable to an unauthorized public deposit.
Separate the record from the executable experiment
Zenodo describes a record as metadata, files and a persistent identifier. Its preservation promise is at the bit level; a preserved proprietary format may eventually become unreadable. Public metadata also does not necessarily mean unrestricted file access. Check the record’s stated file-access conditions rather than treating a readable citation as proof that a recruiter can obtain every attachment. About Zenodo records
For a work sample, make the executable contract explicit beside that record identity: accepted input, parameter values, dependency versions, command, expected output and limits. Preserving source bytes and reproducing an interpretation are different achievements. A script can run successfully while implementing a policy different from the one your application describes.
Record the actual environment used, including external packages or their absence. Our runs used existing CPython 3.9.6 with standard-library modules only. That older runtime is reported as execution evidence, not recommended as a new deployment target. We did not test another operating system, a clean installation or a modern interpreter and do not claim cross-platform reproducibility.
Keep third-party authorship and your own decisions visible. Here the assistant wrote the software, input, documentation and verification code. The fictional application narrative below does not establish that a human candidate performed those actions. Replace its claims with your actual work before using a similar structure in an application.
Completed asset one: map versions, inputs and DOI status
Our original toy summarizes integer measurements by group. The supplied CSV has five rows: group A contains 10 and 14 marked accept, plus 100 marked reject; group B contains 8 and 12 marked accept. These are invented teaching values, with no patient data, scientific source dataset or real experiment behind them.
Version 1.0.0 includes every row. Version 1.1.0 includes only rows whose QC field is exactly accept. Both print each group's count and exact rational mean as JSON. The second policy is appropriate for this example's declared accept-only claim; it is not a universal instruction to discard inconvenient research observations.
| Completed map field | Local version 1.0.0 | Local version 1.1.0 |
|---|---|---|
| Input policy | Include every row | Include only accept rows |
| Group A result | Count 3; exact mean 124/3 | Count 2; exact mean 12 |
| Group B result | Count 2; exact mean 10 | Count 2; exact mean 10 |
| Implementation identity | Full script SHA-256 in the manifest | Different full script SHA-256 |
| Dependencies | Existing CPython 3.9.6; standard library | Same runtime and input |
| Version DOI | Unassigned; not reserved | Unassigned; not reserved |
| Concept DOI | Unassigned; not reserved | Unassigned; not reserved |
This completed map is delivered with both implementations, the original CSV and version-manifest.json. The manifest records each script hash, README hash, input hash, command, process exit code and output identity. The version numbers are author-assigned local labels, not evidence of published Zenodo releases.
The old A result is (10 + 14 + 100) / 3 = 124/3; the new result is (10 + 14) / 2 = 12. Keeping the exact fraction avoids concealing the difference behind rounding. Group B stays at (8 + 12) / 2 = 10, so checking only that group would miss the policy change.
We ran each version in two separate processes. All four exited with code zero, and the two saved outputs for each version matched byte for byte. A successful exit alone therefore cannot choose the correct citation: both programs ran, but only one supports the stated accept-only result. The saved source and synthetic input let a reviewer trace the changed A result to the QC rule.
Keep metadata corrections distinct from changed software
Zenodo's Manage versions documentation distinguishes a metadata edit from changed record files. A new version has separate metadata, files and identifier while remaining linked to the other versions. It also directs minor-error cases to separate file-modification guidance. Do not reduce every possible correction to a blanket promise of permanent immutability or silently replace a materially changed implementation. Manage Zenodo versions
For our example, switching the QC rule changes both source and computed A result. That is a substantive version distinction to preserve. Correcting a descriptive typo is a different task; it does not justify claiming that an unchanged computation has been rerun. Keep a dated change note explaining which layer changed and which evidence remains applicable.
If you already have an authorized public record, read its exact title, creators, software version, file list and related-version information. Follow its actual specific-version DOI, then check the accessible files against your manifest. This guide did not perform that resolver check for a real software record, and the local sample has no DOI to resolve.
GitHub's citation documentation explains how a repository-root CITATION.cff communicates citation metadata and can supply a repository citation prompt. That is useful presentation support, not registration of a Zenodo record or execution verification. Keep version fields and any genuine identifier consistent with the cited files; never copy the documentation's example DOI into your own sample. GitHub citation files

Completed asset two: a fixed-version application review packet
Inez Calder is a fictional applicant preparing a research-software example. Her goal is to show an accept-only summary, so the completed packet selects local version 1.1.0. It contains the selected source identity, shared input identity, literal expected results, reproduction instructions and an explicit unpublished, DOI-unassigned citation note. It makes no claim that she deposited the software or personally wrote this article's code.
Her complete teaching statement reads:
QC-aware group summary, local teaching version 1.1.0. On the supplied five-row synthetic CSV, the declared accept-only policy returns A: count 2, exact mean 12; B: count 2, exact mean 10. The manifest identifies the implementation, input and tested runtime. The sample is unpublished, with no assigned or reserved DOI. The code and checks were AI-authored for this example; personal contribution must be described separately from the saved execution evidence.
From the delivered package directory, a reviewer can inspect the code and repeat the selected small computation with:
python3 1.1.0/summarize.py input.csv
python3 verify_packet.py
The first command prints JSON without overwriting the original receipts. The second reads the preserved files and checks their hashes, expected results and packet selection. It does not import either implementation or rerun the writer. The supplied README states the working directory, filenames and limits so the command does not depend on an author's private home-directory path.
The actual local read-only verifier passed against all four saved runs; its saved JSON receipt records the checks. It also rejected an in-memory packet changed to version 1.0.0 and another changed to an incorrect input hash. Those negative controls show that this particular review checks more than the presence of a result file. They do not exhaust malicious changes, validate arbitrary CSV input or prove the scientific appropriateness of the QC policy.
The verifier uses literal expected values rather than importing the summary function as its own oracle. That separates the application-selection check from the implementation under review. The small arithmetic is also inspectable by hand: the rejected row changes only group A, while group B provides an unchanged comparison. We retained both outputs instead of describing the older result as a crash or claiming that it failed its own all-rows contract. A later change to accepted QC labels, numeric parsing or grouping would require new evidence; matching today’s hashes would not validate tomorrow’s different contract.
The two assets answer different questions. The version map establishes how implementations and results differ. The application packet chooses one of them and verifies that its statement refers to the same source and input. A repository's latest branch could not substitute for that second decision without rechecking the claimed result.
Give the reviewer a bounded, usable route
Choose a format the application actually accepts. If it permits a public link, identify the fixed version and explain what the reviewer should inspect first. If it requests an attachment or forbids external materials, follow those instructions instead. The application instructions determine whether a link is useful; a DOI does not establish that an employer will open the sample.
For a genuine published record, check access without relying on your owner session. Confirm that the cited files are available under the intended conditions and keep any remaining dependency or data-access requirements visible. Do not present “the page loaded” as a completed reproduction test. If files are restricted, explain the permitted review route without disclosing credentials or private research material.
Our completed decision is to retain version 1.1.0 for the stated local demonstration, preserve version 1.0.0 as the comparison and leave both DOI fields unassigned. No upload is needed to establish that the current evidence is incomplete for a claim of published archival availability. Publication would be a separate action with its own permission and record checks.
Use LandOffer's job-search tools to find roles where this kind of software evidence is relevant. Select one relevant opening, follow its work-sample instructions and identify the exact version you checked, your contribution and the remaining limits.


