Research ethics, integrity and collaboration

Science works only because researchers can, by default, trust each other's reports. When a paper says that a measurement was made in a certain way and gave a certain result, readers assume it is true, build on it, and spend their own time and money doing so. Research integrity is the set of practices that justify that trust: honesty in recording and reporting, fairness in giving credit, care in handling data, and openness about methods and limitations. For an honors student these are not abstract principles. They govern everyday decisions about which data to keep, how to describe help from others, how to treat code and text from elsewhere, and who is named on a poster. This lesson sets out those practices, the policies behind them in Canada and the United States, and the ways honest researchers sometimes drift into questionable territory without meaning to.

Fabrication, falsification and plagiarism

The most serious breaches are usually grouped under three headings. Fabrication is making up data or results. Falsification is manipulating materials, equipment, processes, data or images so that the research record is not accurately represented, which includes changing or omitting data to get a desired result. Plagiarism is presenting another person's ideas, words, figures, data or code as your own. In Canada the Tri-Agency Framework on the Responsible Conduct of Research, which governs institutions funded by the three federal granting councils, lists these among its breaches, together with others such as destroying research records, redundant publication, inaccurate authorship and failing to manage conflicts of interest. In the United States the federal definition of research misconduct, used by agencies such as the National Science Foundation, is limited to fabrication, falsification and plagiarism, while universities set wider codes of their own. Both systems require institutions to investigate allegations, and findings can end careers.

Physics has had notable cases. In 2002 an investigation at Bell Laboratories found that Jan Hendrik Schön had fabricated data in many papers on organic semiconductors and superconductivity, some published in leading journals; the same noise curves appeared in figures that were supposed to show different devices, a duplication that colleagues outside his group spotted. Dozens of papers were retracted. The case is a reminder that peer review is not designed to detect fabrication, and that the strongest protection is the openness of data and methods that lets others check. It is also a reminder of how fabrication is usually caught: by people noticing that the data are too good, too smooth or too similar.

Questionable research practices

Outright fabrication is rare. Much more common, and much more dangerous for students because it can happen without bad intent, is a group of questionable research practices. They include dropping data points that disagree with a model without an independent reason, trying many analyses and reporting only the one that works, stopping data collection when the result looks good, adjusting cuts or fit ranges until a result reaches a significance threshold, describing an analysis decided after seeing the data as if it had been planned, and quoting statistical uncertainties while omitting systematic ones. Each of these can make a result look stronger than it is, and together they can produce results that do not replicate.

The defenses are the habits earlier lessons have built. A laboratory notebook written as you work records every run, including the bad ones, so there is no question afterward about what was observed. Hypotheses and comparison criteria written down in advance, as in lesson 5, separate planned tests from exploration. Rules for excluding data must be set before seeing their effect on the result, and every exclusion must be reported. Maya excluded two drops from her raw data, both from the first day, in which the magnet stuck briefly at a burr in the slot cut for the upper gate; her notebook recorded the sticking at the time it happened, she filed the burr smooth before collecting the data used in the thesis, and the thesis reports both the exclusion and the reason. That is a legitimate exclusion. Dropping the pipe whose ratio was furthest from the others would not have been.

Exploration is not wrong. Much of research consists of trying things and seeing what happens. The problem arises only when exploratory results are presented as if they were tests of a hypothesis fixed in advance. The honest practice is to say which is which. Maya's comparison of the three models was planned in her proposal; her estimate of how much off-axis motion could explain the remaining residual was exploratory, and the thesis says so. Examiners and readers can then weigh each kind of result appropriately.

Images, figures and data presentation

Figures can misrepresent data as effectively as numbers. Acceptable processing applies uniformly to a whole image or data set and is described in the methods: adjusting contrast across a whole image, subtracting a measured background, smoothing with a stated filter. Unacceptable processing changes the meaning: removing a feature from one part of an image, splicing images from different runs without saying so, or choosing axis ranges and scales that hide a discrepancy. A good rule is that a figure should allow a reader to see what the data actually show, including their uncertainties and their disagreements with models. Maya's comparison figure in lesson 8 plots every pipe for every model, with error bars, on axes chosen so that the point-dipole failure and the small finite-magnet residual are both visible.

Credit, authorship and acknowledgment

Authorship of a paper, poster or conference talk carries both credit and responsibility. Widely used criteria, such as those of the International Committee of Medical Journal Editors and many physics societies' guidelines, say that each author should have made a substantial contribution to the conception, design, data or analysis, should have helped draft or critically revise the work, should approve the final version, and should be accountable for it. People who helped in other ways, such as a technician who machined parts, a librarian who found sources, or a friend who proofread, belong in the acknowledgments. Large physics collaborations have their own rules; papers from the LIGO and Virgo collaborations and from CERN experiments list every member who meets the collaboration's criteria, often in alphabetical order, with thousands of names.

Many journals now ask for a contribution statement using the CRediT taxonomy, a list of fourteen roles such as conceptualization, methodology, software, investigation, formal analysis, writing of the original draft, review and editing, supervision and funding acquisition. Using the roles explicitly is a good way to avoid disputes. When Maya and Professor Asante submitted an abstract to a conference, they agreed in writing that she would be first author and presenter, he would be the second author, and the department technician who built the release mechanism would be acknowledged. A thesis itself has one author, the student, but it should acknowledge every form of help, and it must make clear which parts of the work are the student's own. If a supervisor wrote a piece of code, or a previous student built the apparatus, say so.

contributions = {   # CRediT roles for a conference abstract
    "Maya Desjardins": ["conceptualization", "methodology", "investigation", "software",
                        "formal analysis", "visualization", "writing: original draft"],
    "Kwame Asante":    ["conceptualization", "supervision", "resources", "writing: review and editing"],
    "Department technician": ["resources"],
}
def qualifies(roles):
    substantive = {"conceptualization", "methodology", "investigation", "software", "formal analysis"}
    writing = {"writing: original draft", "writing: review and editing"}
    return bool(substantive & set(roles)) and bool(writing & set(roles))
for person, roles in contributions.items():
    status = "author" if qualifies(roles) else "acknowledgement"
    print(f"{person:22s} -> {status:15s} ({len(roles)} roles)")

Maya Desjardins        -> author          (7 roles)
Kwame Asante           -> author          (4 roles)
Department technician  -> acknowledgment (1 roles)

Using others' text, code and tools

Plagiarism is not only about copying sentences. Copying an equation's derivation, a figure, a data set or a block of code without acknowledgment is plagiarism too. Paraphrasing a paragraph closely, changing a few words but keeping the structure, is still plagiarism even with a citation, because the expression is not yours. The safe practices are those of lesson 4: take notes in your own words, cite every source, quote exactly when you must use someone's words, and redraw figures from data rather than copying them, citing the source of the data. For code, follow the software's license, keep its copyright notice, and cite it; open-source licenses permit reuse but almost always require attribution.

Artificial intelligence tools raise new questions, and policies are still evolving. Most universities and journals now allow some uses, such as checking grammar, suggesting code or explaining a concept, while requiring that the author take full responsibility for everything submitted, that significant uses be disclosed, and that AI systems not be listed as authors. Generated text must not be presented as your own analysis, generated references must be checked because they may be invented, and generated code must be verified like any other code, by the tests of lesson 7. Read your department's policy at the start of the project, ask your supervisor when in doubt, and keep a note of how you used any such tool, in the same spirit as a laboratory notebook.

Managing research data

Research data deserve the same care as the apparatus. A data management plan answers a few practical questions: what data will be produced, in what formats, how they will be documented, where they will be stored and backed up, who can access them, how long they will be kept and whether they will be shared. In Canada the Tri-Agency Research Data Management Policy, issued in March 2021, required institutions to publish data management strategies, which they did by 2023, and it is phasing in requirements for data management plans in certain funding competitions and, later, for depositing data that support publications. In the United States the 2022 memorandum from the White House Office of Science and Technology Policy, often called the Nelson memo, directed federal agencies to make the publications and supporting data from federally funded research freely available without embargo by the end of 2025; agencies issued policies on that schedule, with the National Institutes of Health applying its zero-embargo rule from July 2025, although reports in 2026 suggested the memo itself might be revised. The details will change; the habit of planning how data will be kept and shared will not.

For an honors project, a simple plan is enough. Keep raw data read-only in a clearly organized folder with a README describing each file. Back up to at least two places, one of them off the laptop, such as the university's storage service. Record checksums of the raw files so that any accidental change can be detected. Describe the data in the thesis well enough that someone could use them, and, if your supervisor agrees, deposit the data and code in a repository such as the Federated Research Data Repository or Borealis in Canada, Zenodo, or an institutional repository, which gives them a permanent identifier you can cite. Maya deposited her raw timing files, calibration readings and analysis code with her thesis.

import hashlib
raw = {   # file name -> contents (a few lines of each raw file)
    "gates_CuL_2027-01-18.csv": "drop,t_upper_s,t_lower_s\n1,1.8421,16.8812\n2,1.8406,16.8317\n",
    "hall_calibration_run4.csv": "z_mm,B_mT\n40,4.211\n50,2.236\n",
    "micrometer_CuL.csv": "position,OD_mm,wall_mm\n1,15.874,1.019\n2,15.878,1.023\n",
}
manifest = {name: hashlib.sha256(text.encode()).hexdigest()[:20] for name, text in raw.items()}
for name, digest in manifest.items():
    print(f"{digest}  {name}")
# later: someone edits one value by hand
raw["gates_CuL_2027-01-18.csv"] = raw["gates_CuL_2027-01-18.csv"].replace("16.8317", "16.8417")
changed = [n for n, t in raw.items() if hashlib.sha256(t.encode()).hexdigest()[:20] != manifest[n]]
print("files changed since the manifest was made:", changed)

63c46a5ca9f80969bbd6  gates_CuL_2027-01-18.csv
c54a5e843b18c6d51c05  hall_calibration_run4.csv
a5a5d48659104a489bb7  micrometer_CuL.csv
files changed since the manifest was made: ['gates_CuL_2027-01-18.csv']

Safety, care and respect

Integrity includes responsibility for people and equipment. Experimental projects must follow the laboratory's safety rules: Maya's neodymium magnets can pinch skin painfully when two snap together and can shatter, so she handled them with gloves and eye protection and kept them away from electronics and from anyone with an implanted medical device. Cooling a pipe in a freezer was safe; the department's rules would have required training before she could use liquid nitrogen. Computational projects have their own responsibilities, such as fair use of shared computing resources and secure handling of any data that involve people. Projects that involve human participants, even a simple survey, need research ethics approval before data collection begins.

Finally, integrity includes how you treat colleagues. Research groups work best when people share credit generously, ask for and give help freely, disagree about ideas rather than about persons, and report problems, including their own mistakes, promptly. Most universities have offices that can advise confidentially if you encounter harassment, unfair treatment or suspected misconduct. Knowing that such help exists, and using it if needed, is part of being a member of a research community.