Every Number in Your Thesis Should Be Ctrl+F Findable

The precise hallucination

Large models write numbers beautifully. Sample size N=337. Internal consistency α=0.821. R²=0.594. Significance p<0.01. The formatting is right, the decimal places are consistent, the statistical vocabulary is used correctly, and the whole passage looks as if it were copied out of a published empirical study.

Then your supervisor asks one question: "Which paper does the 0.594 come from?"

From there there are only two roads. Either you spend two hours turning pages through a PDF until you find the line, or you never find a source at all — and in that second case the number was not a typo. It was invented.

Numbers in a thesis are not decoration. Each one is a claim about the outside world, and a claim is either checkable or merely number-shaped.

Method one: the whitelist gate

The first gate in a search is not the keyword. It is the source.

Only elite universities, core journals and authoritative institutions in the country you are actually studying may enter the retrieval pipeline. Domains that are not on that list are discarded during the search stage, before a single abstract is even parsed.

This is not about looking rigorous. Mixed sources contaminate the evidence pool: a number propped up by one weak article drags the apparent confidence of your conclusion down to a level nobody can defend. The gate turns "trustworthy source" from the author's good intentions into the default behaviour of the system.

The gate also changes what a search returns. The same query against a mixed pool gives you a hundred plausible-sounding links. The same query against a gated pool gives you a dozen sources you can actually cite, each with a named institution behind it. Fewer results, more of them usable, and a much shorter argument with your supervisor about why the list is so short.

Method two: anchoring every number to its sentence

A number separated from its sentence loses its meaning.

WriterOne stores each number together with its anchor sentence — the exact sentence that carries it in the original text — then compares them word by word. That sentence has to be findable in the archived full text, and the number inside it has to match the number on the card. If the sentence cannot be found, or the number in the sentence does not match, the evidence is void — not downgraded to "needs review", but discarded outright.

The rule is harsh, and it is the foundation of the whole evidence chain.

Method three: missing means "not reported"

If the original paper does not report a statistic, the ledger says "not reported". Nothing is filled in.

Authors often find gaps unflattering, so they ask the model to smooth them over: Cronbach's α, RMSEA, CFI. Those invented values then travel into the chart, into the conclusion, and into the defence room, where no sentence in the original ever supported them.

The same rule applies to charts. When a card carries fewer than two verifiable data points, no chart is drawn. It degrades into an explicitly labelled quality audit chart that states which indicators are missing and why, instead of a picture that looks like data.

Engineering: fingerprints and a 22-column ledger

Verifiability needs somewhere to live. The archived original is stored as Markdown and carries a SHA-256 fingerprint: change one byte of the file and the fingerprint changes, so both you and your supervisor can confirm that the text you are checking is the text that was captured.

Behind every argument card sits a 22-column data ledger: metric definition, unit, time period, geography, source, sample size, sampling method, reliability, validity, confidence interval, model fit, significance, comparability of definitions, limitations and more. Every column either holds a verifiable value or is explicitly marked "not reported". A gap is not an empty cell — it is a finding.

Two copies of the same PDF can differ by a single character in the reference line, and a student who cannot show which copy was used cannot answer the question at all. The fingerprint turns "I promise this is the original" into a fact that either matches or does not.

Conclusion: verifiability beats completeness

A literature review with 30 cards, of which 12 carry numbers that can be traced word by word back to the source text, is stronger than one with 50 cards and 20 numbers nobody can place.

That second review is the one supervisors actually check. Verifiability matters more than completeness, because a number that cannot be checked does not state a finding. It states a risk.

Every verified number is a sentence you can defend out loud, and that is what a thesis defence actually is: a long conversation in which somebody keeps asking where each number came from.