A real output
One real paper, and everything we measured about it
Judging a research tool from its marketing is impossible, so here is an actual run — the paper in full, the eight checks the system runs on its own work, and what those found. Nothing on this page is illustrative. The faults are published beside the strengths at the same size, because a tool that cannot describe its own limits is not one you should trust with a thesis, and one that reports only its limits has told you just as little.
What the audit scored it
- None
- invented citations
- 30/30
- figures verified at source
- 21/21
- references actually cited
Weakest dimension: grounding, 6.4. The overall is capped near the worst of grounding and coherence on purpose — otherwise a perfect structure and a long reference list average away unsupported claims, and a paper with five serious grounding faults scores “solid”.
Read it yourself
Everything below is evidence about the paper. The paper itself — 3,383 words, 8 sections, all 21 references — is published in full, on screen and as the PDF the system typeset.
The run
“Knowledge and practices of operating room nurses regarding surgical specimen handling and perioperative safety: A review of published evidence and implications for Korle-Bu Teaching Hospital”
Written in
10 min
3,383 words, 8 sections
Records searched
1,059
80 carried into the review
References
21 of 21
every reference is cited in the text
Off-topic excluded
81
records sharing no term with the question
Local studies found
6
the question named a specific hospital
Figures verified
30/30
quote found verbatim in the cited study
Invented citations
0
checked deterministically, not by a model
How the score breaks down
Computed from countable signals only — never asked of a model, because a score a model invented would be the most trusted-looking number here with nothing behind it. A dimension with nothing to measure returns no score and is dropped from the average rather than counted as a pass.
| grounding | 6.4 | 6 claims flagged; 30/30 datapoints verified against source text |
|---|---|---|
| coherence | 8.9 | 1 issue: objective unanswered |
| citations | 8.8 | 21 references, 21 peer-reviewed; 22 statistics with no citation |
| literature | not scored | no comparison with prior work |
| completeness | 10.0 | 8 sections; all expected parts present |
Deliberately not scored
- — Novelty — requires judgement against the field, not a countable signal
- — Writing quality — style is not measurable from the pipeline's outputs
- — Methodological soundness of the underlying study
- — Repetition of figures — counted in the integrity report, not scored here
What the audit found
What held up5
No invented citations
Every citation in the text resolves to a real source in the reference list. The audit checks this deterministically, not by asking a model.
Every figure traced to its source
30 of 30 extracted figures proved their quote appears verbatim in the study cited. Any that could not were discarded rather than printed.
No padded bibliography
All 21 references are cited in the text. Sources the prose never used were dropped and the rest renumbered, because References means works cited — the search collected far more than the paper needed.
81 off-topic records thrown out
Records sharing no term with the research question were excluded before ranking, rather than being allowed to fill the evidence window because they were famous.
6 studies from the named setting
The question named a specific hospital, so work from that setting was ranked above otherwise comparable work from elsewhere. Ranking by citation count alone finds the famous foreign review and misses the local field study that measured the actual question.
What did not4
22 figures we could not place
Of 72 numbers the text states, 22 carry no citation in their own sentence. Some are cited in the sentence before — the check is deliberately strict — but this is the paper's real weakness and we would rather say so.
9 figures restated too often
Counted and reported, deliberately not scored: the underlying problem is already reflected in the citations and grounding scores, and scoring it a third time would punish one weakness three ways.
6 claims flagged for review
A second pass re-reads every claim against the evidence. It is imperfect in both directions — on an earlier paper it wrongly accused five sources of not existing — so its findings are shown, never hidden, and never allowed to be the only check.
The paper does not answer everything it set out to
The self-consistency pass compares the paper against itself — objectives against conclusions, sections against each other — and found 1 issue: objective unanswered. A paper can pass every claim sentence-by-sentence and still not close the question it opened.
What it is not
This paper reviews published evidence. It is not a study: no data was collected from anyone, and the system said so in the document itself rather than letting a title imply otherwise. If your topic needs primary data, Fortell will tell you that and produce the instrument and sampling plan for you to run — it will not invent results.
Research runs cost real money, so they are metered — but the whole ML pipeline is free and unmetered, and you can see how your own file is read before spending anything.