Nanotechnology in Cancer: Building the Chapter 28 Fact Check by Novia

A fact check that only tells you a sentence is true or false isn't finished. It has to also account for every sentence it didn't check, and why. That's the standard applied here to a chapter on nanotechnology in cancer, treating a fact check less like a vibe and more like a workbook: 168 sentences, run through four distinct moves, producing three sheets that can be audited on their own.

The topic under the microscope

The chapter covers nanoparticle drug delivery, imaging, and theranostics, tiny carriers designed to deliver treatment, produce imaging signal, or do both at once. That's genuinely difficult material to fact-check, because it mixes verifiable claims, statistics, approvals, dated facts, specialist assertions, with explanatory passages that are harder to pin to a single external source. The process built here treats that mix as the whole problem, not an inconvenience to work around.

Move one: count and split

The first move is simply counting every sentence in the chapter, all 168 of them, and deciding which ones need to be checked against outside sources. Statistics, guidelines, regulatory approvals, evidence, specialist claims, and dated facts get sent to the web for verification, 54 sentences in total. Everything left over, 114 sentences, gets marked AI-only, and critically, each one carries a recorded reason explaining why it wasn't sent out for a web check. The 114 AI-only sentences are then reconstructed by file, in reading order, across six source files. The two largest are the file covering clinical material, with 33 AI-only sentences, and the file covering delivery mechanisms, with 30. The numbers are made to sum and reconcile exactly, so a sentence can't quietly disappear between counting and reporting.

Why publishing only the 54 is not a finished report

It would be easy to stop here and publish the 54 web-checked sentences as "the fact check." The process explicitly treats that as a partial, unfinished report, one that has dropped the actual classification work on the other two-thirds of the chapter. A report that only shows what got checked against the outside world says nothing about whether the material that didn't get checked was handled carefully.

Move two and three: editorial cross-check and the reread

The next two moves build out the remaining sheets. Sheet B holds the full 114 AI-only sentences, each with its recorded justification. Sheet C is an editorial pass across the AI-only pile that turns up 13 distinct findings, duplicate sentences repeated across different files, including the blood-brain barrier being defined twice, formatting or encoding junk left over from generation, and mechanisms described backwards from how they actually work. That editorial sweep is followed by a dedicated reread of the 114 AI-only sentences specifically hunting for hallucinations, factual claims that sound plausible but are wrong. That reread turns up three flagged sentences: a description of PEG's stealth coating written with the mechanism reversed, a claim about siRNA assembly that doesn't hold up, and a sentence that mischaracterizes gold nanoparticles as a chemotherapy agent. None of these three sentences were ever sent to the web for a check, because they were classified as AI-only from the start, which is exactly why the dedicated reread step exists: only a second pass through sheet B catches errors that the initial web-versus-AI-only split was never designed to find.

Why 15 minutes of a demo threshold matters

One detail underlines how seriously this process treats its own limitations: even after the workbook passes its internal tests, a materiality threshold used during development is flagged as still an unapproved fixture, not something that should quietly become policy. Passing tests is treated as necessary but not sufficient. If a demo-stage threshold slides into production use without explicit approval, that's a failure of process discipline, even if nothing in the chapter itself is wrong.

The four-move structure, generalized

The value of this approach isn't specific to nanotechnology or to this one chapter. The four moves, count every sentence, split it by source-checkability with a reason recorded for each, run an editorial cross-check for duplicates and reversed mechanisms, then reread the AI-only pile specifically for hallucinations, apply to any AI-assisted written material. Skip the editorial sheet and a classifier's decisions go unaudited. Skip the hallucination reread and a chapter's own duplicated or reversed claims ship unnoticed.

Key takeaways

• A complete fact check accounts for every sentence, not just the ones verified against outside sources.
• Of 168 sentences in the chapter, 54 were checked against the web and 114 were marked AI-only, each with a recorded justification.
• An editorial cross-check across six source files found 13 findings, including duplicated definitions and mechanisms described backwards.
• A dedicated reread of the AI-only sentences, separate from the editorial pass, caught three hallucinated claims that the web-check step would never have seen.
• A materiality threshold used during testing is explicitly kept separate from approved policy, even after all tests pass.

Who this is for

This is aimed at anyone building or reviewing AI-assisted written material who needs a repeatable, auditable way to separate verified claims from unverified ones, rather than a single pass/fail judgment. It's a worked example from the Humanitarians AI Fellows program of applying a rigorous, four-move fact-checking workbook to a real chapter before publication.