LivoDraft
◇
How it drafts
Short answer: yes, but only if the tool asks for the research first. Most AI writing tools take a topic and hand back a document, which means the parts of a thesis that are yours — your objectives, your method, your measurements — get written by reading other people’s papers. LivoDraft takes the study first and refuses to draft without it: a Thesis / Dissertation requires 16 fields, so a submission carrying only a topic is turned away fifteen fields short.
A prompt carries one thing: the topic. A thesis needs the topic and the research design behind it. That gap is not a matter of prompt quality — no wording recovers a measurement that was never taken. So the honest division of labour is the one this tool enforces: the student does the research, the tool drafts and formats it, and the assistance is declared.
| Field | A prompt | LivoDraft |
|---|---|---|
| Topic | Yes | Required |
| Research objectives | Invented | Required, used verbatim |
| Methodology | Invented | Required |
| Study design | Invented | Required |
| Data source | Invented | Required |
| Research gap | Invented | Required |
| Problem statement | Invented | Required |
| Your findings | Invented | Required, locked |
Those eight sit alongside the identity and format fields a title page cannot be printed without — name, department, university, supervisor, roll number, degree, year, citation style. Sixteen in total for a Thesis / Dissertation. A Research Paper and a Journal Article require thirteen each. The list is held in one place and read by both the form and the server, so the page can never ask for less than the generator enforces.
This is the part most tools get wrong in a way that is invisible until a viva. If a student’s measurements are handed to every chapter, the Introduction and the Literature Review start reporting them — and in those chapters a number sits next to a citation, so the student’s own result ends up credited to somebody else’s paper. It reads fluently and it is a fabrication.
So the findings are routed, not broadcast. They reach Results, Discussion and Conclusion, the chapters where the work is the student’s to report. They are withheld from the background chapters, which are written from the literature and cite it. The build log records which way each chapter was routed.
A proposal or a synopsis is written before there is data, and the honest voice for that is the future one: what the method is expected to show. The failure mode is a tool that fills the gap with plausible numbers instead — “an accuracy of 94.3%” for an experiment nobody ran. A deterministic pass removes invented figures rather than softening them into vaguer ones, because a vaguer fabrication is still a fabrication. Literature values carrying a citation and a year are left alone; they belong to the paper they came from.
Prompting a model to respect a value is not a guarantee, and a long document gives it many chances to drift. After the draft is written, every figure in it is checked against the data that was supplied — the dose, the metric, the study site, the method. A value that has drifted is corrected and re-verified. The alternative is worse than a wrong number: a document that quietly argues with the student’s own table, in front of a panel that will read both.
Fabricated citations are the best-documented failure of general AI writing tools — a reference that is correctly formatted, plausibly titled, and does not exist. References here come from Crossref, and before the bibliography is finalised every DOI is resolved against the official DOI handle service; one that is dead or unregistered is dropped. Strict verification is on, so a DOI that cannot be reached after repeated attempts is also dropped — unverifiable is not verified.
Two guards keep that honest in both directions. If the DOI service itself is down, a whole bibliography would look unverifiable at once; above a set proportion of failures strictness stands down for that run and logs loudly, because losing most of a bibliography does not mean most of it was invented. And records with no DOI are always kept: an IS, ASTM or IRC standard is a real reference that never had one.
The bibliography is then reconciled against the body, so every reference listed is actually cited and every citation in the text maps to a reference that exists. A list padded with uncited entries is the oldest tell of a generated document.
Yes, if the tool takes the data as input rather than inventing it. LivoDraft requires your objectives, methodology, study design, data source, research gap and findings before it will draft, and it checks every figure in the finished document against what you supplied.
A general chat model will, because a topic is all it has to work from. Here, invented figures are removed by a deterministic pass, and where no data has been supplied the results are written as what the method is expected to show rather than as measurements that were never taken.
Often not — fabricated but well-formatted citations are a known failure of general models. References here are fetched from Crossref and every DOI is resolved live before the bibliography is finalised; dead or unverifiable ones are dropped, and the list is reconciled so nothing is listed that the text does not cite.
Because those fields are the thesis. A Thesis / Dissertation requires sixteen; eight of them are the research design and findings, and the rest are what a title page and a citation style cannot be produced without. A tool that asks for fewer is inventing the difference.
Yes — for a research proposal or a synopsis, where expected outcomes are the honest framing. For a final thesis or a case study it will not waive the evidence requirement, because that is the point at which a document would be reporting an experiment nobody ran.