Claim–Evidence — How it verifies

For CITED sentences in your text, it assesses whether the source you cited actually supports that sentence: supported, partial, unclear, unsupported, contradicted, or no evidence. The evidence text is taken from the full text of the PDF you uploaded, or failing that from the source's abstract; which one was used (“full text” or “abstract”) is shown to you, because an abstract may not represent the full text. The assessment looks ONLY at that evidence text. Any quote shown in the result is searched verbatim in the source; if it cannot be found, it is dropped — so a sentence that is not in the source can never appear in quotation marks. If we hold neither the full text nor the abstract, the result is “no evidence”; nothing is guessed. Sentences without citations are not assessed, because there is no source to compare against. There is a budget cap: at most the 8 most-cited sources and at most 4 sentences per source are assessed; in larger works the remaining cited sentences are not covered in that pass. This is a guide, not a reviewer's verdict.

Evidence location — How it verifies

Finds where in the source the passage supporting a claim actually appears. The source's full text is split into chunks and embedded; the chunk closest to the claim is selected by a deterministic similarity computation. The location is not asked of a language model — a model can invent a paragraph number, a vector match cannot. Works only for sources whose full text has been uploaded and indexed. If the full text is missing, it honestly reports “no chunks” rather than locating something in an abstract.

Research provenance record — How it verifies

Shows the timestamped events behind a work on a single axis: sources added, versions, comments, generation, evidence checks, exports. Only events carrying a real timestamp enter the timeline; an event whose timestamp cannot be resolved is never produced. Because part of the measurement data has no timestamp at all, two layers are used: events and stages. The timeline is structurally INCOMPLETE and says so on screen: automatic versions are pruned and some steps are not recorded. It is therefore never claimed to be complete. Most importantly: “not recorded” and “did not happen” are different things and are shown differently.

Generation trace — How it verifies

Measures how much of the text we generated still survives verbatim in the final text. When we generated a section, we store a fingerprint of the generated text (5-gram hashes) and later count how much of it is retained in the final text. The plain text itself is not stored. Whether a human wrote a given sentence CANNOT be measured and is never claimed. No authorship percentage is produced. The only thing measured is text retention; this is not AI detection — looking at prose and declaring “AI wrote this” is not measurement.

Institutional guide compliance — How it verifies

Checks the TEXT-level rules of the university template you selected: citation-style mismatch, heading numbering, empty sections. Entirely rule-based; no AI is used. A guide is already an explicit rule, and having a model interpret it would only add risk. Templates are derived from institutions' published guide documents and stored with a link to the source. Margins, font and line spacing are NOT checked here — they are already applied on export. Rules the template data does not carry (page limits, front-matter order, appendix rules) cannot be checked, and each of them is listed explicitly on screen.

Prior-art scan — How it verifies

Compares the contribution you state in your work against published patents, element by element. Technical elements are extracted from your contribution and each element is compared with the 8 most relevant prior-art documents found. Where an element is said to be “explicitly disclosed” in a document, a VERBATIM quote from that document's text is required; the quote is then searched automatically in the source and the cell is flagged if it cannot be found. This is NOT a patentability opinion, and the technique scanned is limited. Most importantly: finding nothing does NOT prove your contribution is novel — it only shows this scan found no match. The final assessment is a patent attorney's work.

How does Gödel verify a claim?

What distinguishes Gödel is not generating text but making text defensible. Every tool below follows the same rule: when the information is not available, nothing is invented — it says “could not verify”. We also state what each tool does NOT do, because a method statement without its limits is incomplete.

Claim–Evidence

What it does: For CITED sentences in your text, it assesses whether the source you cited actually supports that sentence: supported, partial, unclear, unsupported, contradicted, or no evidence.
How it verifies: The evidence text is taken from the full text of the PDF you uploaded, or failing that from the source's abstract; which one was used (“full text” or “abstract”) is shown to you, because an abstract may not represent the full text. The assessment looks ONLY at that evidence text. Any quote shown in the result is searched verbatim in the source; if it cannot be found, it is dropped — so a sentence that is not in the source can never appear in quotation marks.
What it does not do: If we hold neither the full text nor the abstract, the result is “no evidence”; nothing is guessed. Sentences without citations are not assessed, because there is no source to compare against. There is a budget cap: at most the 8 most-cited sources and at most 4 sentences per source are assessed; in larger works the remaining cited sentences are not covered in that pass. This is a guide, not a reviewer's verdict.
Example: “Method X reduces the error rate by 40% [3]” is flagged as unsupported if source 3 only reports a 12% reduction — and the relevant sentence from the source is shown.

Evidence location

What it does: Finds where in the source the passage supporting a claim actually appears.
How it verifies: The source's full text is split into chunks and embedded; the chunk closest to the claim is selected by a deterministic similarity computation. The location is not asked of a language model — a model can invent a paragraph number, a vector match cannot.
What it does not do: Works only for sources whose full text has been uploaded and indexed. If the full text is missing, it honestly reports “no chunks” rather than locating something in an abstract.
Example: If the claim matches a paragraph in section 4 of the source, that paragraph and its similarity score are shown so you can open and read it yourself.

Research provenance record

What it does: Shows the timestamped events behind a work on a single axis: sources added, versions, comments, generation, evidence checks, exports.
How it verifies: Only events carrying a real timestamp enter the timeline; an event whose timestamp cannot be resolved is never produced. Because part of the measurement data has no timestamp at all, two layers are used: events and stages.
What it does not do: The timeline is structurally INCOMPLETE and says so on screen: automatic versions are pruned and some steps are not recorded. It is therefore never claimed to be complete. Most importantly: “not recorded” and “did not happen” are different things and are shown differently.
Example: When we hold no record for a step, the screen says “not recorded” — it does not say “this step was skipped”, because we do not know that.

Generation trace

What it does: Measures how much of the text we generated still survives verbatim in the final text.
How it verifies: When we generated a section, we store a fingerprint of the generated text (5-gram hashes) and later count how much of it is retained in the final text. The plain text itself is not stored.
What it does not do: Whether a human wrote a given sentence CANNOT be measured and is never claimed. No authorship percentage is produced. The only thing measured is text retention; this is not AI detection — looking at prose and declaring “AI wrote this” is not measurement.
Example: If you generated a section and then rewrote it, retention drops. That is not proof you wrote it; it is a measure of how little of our generated text remains.

Institutional guide compliance

What it does: Checks the TEXT-level rules of the university template you selected: citation-style mismatch, heading numbering, empty sections.
How it verifies: Entirely rule-based; no AI is used. A guide is already an explicit rule, and having a model interpret it would only add risk. Templates are derived from institutions' published guide documents and stored with a link to the source.
What it does not do: Margins, font and line spacing are NOT checked here — they are already applied on export. Rules the template data does not carry (page limits, front-matter order, appendix rules) cannot be checked, and each of them is listed explicitly on screen.
Example: If the guide requires APA 7 and your work is set to IEEE, this is shown as a high-priority finding — one of the most common reasons a thesis is returned.

Prior-art scan

What it does: Compares the contribution you state in your work against published patents, element by element.
How it verifies: Technical elements are extracted from your contribution and each element is compared with the 8 most relevant prior-art documents found. Where an element is said to be “explicitly disclosed” in a document, a VERBATIM quote from that document's text is required; the quote is then searched automatically in the source and the cell is flagged if it cannot be found.
What it does not do: This is NOT a patentability opinion, and the technique scanned is limited. Most importantly: finding nothing does NOT prove your contribution is novel — it only shows this scan found no match. The final assessment is a patent attorney's work.
Example: If one component of your method is explicitly described in a 2019 patent, that cell is marked “disclosed” and the sentence from the patent text is placed beside it.

One shared rule: what cannot be measured is not claimed

None of these tools issues a verdict, and none of them treats a language model's confidence as verification. If a registry could not be reached, the result is “could not verify”, not “invalid”. If a measurement could not be made, no number is produced. The final judgement always rests with the author.

Try it on your own work.

Open Gödel

In exactly what ways do AI-suggested sources go wrong? We list twelve error classes on a separate page, stating for each what we detect and what we do not.

Error taxonomy →