What TraceFolio is — and what it isn't

The Certified Execution Record proves that a specific evidence set produced a specific report under recorded parameters — model, prompt version, pipeline version, schema version — and that neither the inputs nor the output have been altered since. That is a strong, checkable claim about process.

It is not a claim about truth. If a source document contains a false statement, a faithful, correctly cited report will faithfully report that false statement. Verification tells a reader that nothing was fabricated or swapped after the fact; it does not tell them the underlying documents are accurate.

• Extracts text from your documents and records where each passage came from (file, page, section and character range).
• Produces structured findings — key facts, entities, timeline events, contradictions and uncertainties — each tied to a specific source passage.
• Drops any finding whose quoted excerpt cannot be matched verbatim against the extracted source text.
• Distinguishes explicit statements in your documents from inferences drawn across them.
• Hashes your inputs, the analysis parameters and the delivered document, and certifies that binding so it can be checked independently later.

• Does not determine whether your documents are truthful, complete or authentic.
• Does not authenticate signatures, detect document forgery, or establish chain of custody for the originals.
• Does not provide legal advice, and is not expert evidence or a forensic examination.
• Does not guarantee that every relevant fact in your documents has been surfaced.
• Does not decide outcomes — a reader must still check each cited passage and reach their own conclusion.

• Scanned and photographed documents. Text is recovered with OCR, which can misread poor scans, handwriting and unusual layouts. Extraction confidence is recorded so low-confidence sources are visible.
• AI variability. Language models can miss a relevant passage or over-generalise. Citation validation removes unsupported claims, but it cannot invent a finding that the model failed to make.
• Contradictions. Detected contradictions are candidate conflicts between cited passages, not adjudications of which passage is correct.
• Dates. Ambiguous or partial dates are labelled as such rather than guessed into a precise position on the timeline.
• Evidence set boundaries. The report can only reason about what you uploaded. Missing context is invisible to it.

Treat it as a structured, checkable index of your own material: read the finding, follow the citation to the page, and confirm it. The value is that every claim points somewhere specific, and that the whole package can be shown to a third party without asking them to trust us.