# Evidence audit

> What share of our published claims carries a primary source — published as measured, including when it is poor.

- Language: English
- Page: https://smilehandbook.com/en/evidence-audit/
- Français: https://smilehandbook.com/fr/audit-des-preuves/
- العربية: https://smilehandbook.com/ar/evidence-audit/
- All published pages: https://smilehandbook.com/llms-full.txt

Every page on this site makes claims. Some sit on data we collected and can re-derive; some sit on documents somebody else published; a few sit on nothing sturdier than common knowledge in the field. A reader has no way to tell these apart by reading, and we have no way to tell them apart by memory. So we intend to count them. The evidence audit is a recurring count of how large a share of our published claims carry a first-hand source, published in full whether or not the number flatters us. Our evidence audit page explains how claims are counted and presents the recorded results. All three results are set out in the last section of this page, and the passage count printed after them is a separate and smaller measurement. This page explains the method. The trigger for the first audit and the level at which a section is flagged are both fixed and published below, so neither can be picked once a result is known.

## Why we count our own claims

No clinic can pay to appear here, and this site is not a clinic, an agency, or a booking service for clinic treatment or travel. One kind of bias falls away with that; another stays exactly where it was: the temptation to write confidently about things we have not checked. Sloppiness survives the absence of commercial pressure, and on the page it reads exactly like authority. The audit exists because "we are independent" is a claim about incentives rather than accuracy, and accuracy is the only one of the two that can be measured. Announcing the measure before running it also guards the reader against a familiar trick: a standard invented after the results are known that happens to fit them.

## What counts as a claim

We count claims, not pages and not sentences. A claim is any statement that could be shown to be wrong: a number, a proportion, a date, a comparison, a stated frequency, a causal assertion, or a factual statement about procedures, costs, regulation, or outcomes. Definitions, navigational text, statements about how this site works and explicitly labelled uncertainty are not claims and go uncounted. A single sentence carrying two testable statements counts as two. Each audit's claim list is extracted page by page and kept, so a later audit can be set against an earlier one instead of starting from scratch.

## What counts as a first-hand source

A first-hand source is one we can go back to and re-derive the number from. Three kinds qualify. First, data we gathered ourselves and still hold: the 2,115 publicly visible reviews we counted across six Istanbul dental clinics, say, and the 50 one-star reviews we read individually — all of those with enough text to describe what happened. Second, primary documents: regulator and registry records, official legal or licensing texts, published research read in full rather than through its abstract. Third, published statistics from the body that produced them, cited to the specific table.

## What does not count

A clinic's own marketing page. An agency blog. A press article citing nobody itself. A statistic quoted by a site that never says where it came from. A figure produced by a language model. And our own earlier pages: citing ourselves creates no evidence, so self-references drop out of the numerator entirely. Any claim left standing on none of the qualifying sources is recorded as unsourced. Such claims are not deleted quietly; each one is either sourced, rewritten as an explicit uncertainty, or removed, and the audit records which of the three happened; a claim that reaches none of them within that audit stays listed as unsourced.

## When the first audit runs and what it will cover

This page states the audit trigger and describes the page sets used by the three audits. The first audit used a reference set of 36 pages, of which 24 carry passage labels and were audited. The three language home pages (navigation rendered from a shared component) and the nine legal notices (statements about this site, not about treatment) are outside it and are named rather than dropped. Later audits run every six months, or sooner once 20 further pages have been published, whichever comes first; each audit states its date and the page count it covered. The second audit ran on 7 September 2026. Its reference set comprised 86 pages, 50 more than the first set, and the audit examined the 25 pages carrying a first-hand clinical proposition. The third audit ran on 10 September 2026. Its reference set comprised 110 pages, 24 more than the second set, and the audit examined 23 Arabic pages that neither earlier audit had reached.

## What gets published, good or bad

Each audit publishes the overall percentage, the denominator, the breakdown by section of the site, and how many claims moved into the unsourced category. A section is flagged when fewer than half of its claims that require an external source carry a first-hand one. The overall figure is published whatever it is. Every unsourced claim receives one of three outcomes in the same audit — a first-hand source is found and added, the sentence is rewritten as explicitly labelled uncertainty, or it is removed — and the number under each outcome is published; a claim that reaches none of the three stays listed as unsourced. Where the overall figure is poor, the poor figure is what appears. An audit capable only of producing good news is not a measurement.

## What this audit does not tell you

Two limits are worth stating plainly. What the audit measures is whether a first-hand source sits behind a claim, not whether the claim is true; a well-sourced claim can still be wrongly interpreted. And the people who ran the first audit are the people who write the pages. No independent auditor exists and no external verification of the count is planned. Both are weaknesses we would rather name than let a reader discover.

## What the audits have found so far

The first audit used a reference set of 36 pages, of which 24 carry passage labels and were audited. Three of those 24, the full index pages in each language, carry no claim at all, so its counts come from 21 pages. The three language home pages and the nine legal notices stayed outside it, exactly as this evidence audit said they would. Across the three languages that count came to 585 claims, of which 489 carried a first-hand source, or 83.6 percent; of the 78 claims that need a source somebody else published, 12 carried one, or 15.4 percent, and 96 claims carried no qualifying source at all. The second audit used a reference set of 86 pages and examined the 25 of them carrying a first-hand clinical proposition, 11 in English and 14 in French. It counted 1,654 claims, of which 1,631 carried a first-hand source, or 98.6 percent; of the 1,138 claims that need a source somebody else published, 1,120 carried one, or 98.4 percent, and 23 claims carried no qualifying source at all. The third audit used a reference set of 110 pages and examined 23 Arabic pages that neither earlier audit had reached; four auditors read them separately. It counted 412 claims, of which 375 carried a first-hand source, or 91.0 percent; 22 rested on a source that says less than the sentence built on it, 5 rested on no qualifying source at all, 3 rested on a source that frames the question without answering it, 1 pointed to the wrong numbered clause of the right text, and 6 could not be reached and are recorded as no verdict rather than as an absence. That 91.0 percent sits below the two figures above it because the third audit ran a test the first two did not: it asked not only whether a source stands behind a sentence, but whether that source answers the sentence built on it. Taken together the three audits cover 69 of the 110 pages in the reference set and 2,651 claims, of which 2,495 carry a first-hand source, or 94.1 percent. The split between claims that need a source somebody else published and claims that do not was recorded by the first two audits alone — 1,216 of their 2,239 claims need one and 1,132 carry one, or 93.1 percent — and the third audit did not record it, so this page does not extend that figure to it.

The breakdown by section follows the five sections this site sorts its pages into, and the three audits together reach all five. How we work counted 120 claims with 111 first-hand; 9 of them need an outside source and none of those carried one. What we measured counted 419 claims with 360 first-hand; 31 need an outside source and none of those carried one. Rules and reimbursement counted 829 claims with 783 first-hand; of the 698 the first two audits counted there, 495 need an outside source and 463 carried one, or 93.5 percent. Treatments counted 906 claims with 870 first-hand; of the 625 the first two audits counted there, 435 need an outside source and 427 carried one, or 98.2 percent. When something goes wrong counted 377 claims with 371 first-hand; 246 need an outside source and 242 carried one, or 98.4 percent. The third audit added 131 claims to Rules and reimbursement and 281 to Treatments and recorded no outside-source split of its own, so the percentages just given still describe the first two audits alone. Where the addition does show is the share carrying a first-hand source: in Treatments 615 of the 625 claims the first two audits counted there carried one, or 98.4 percent, and 870 of the 906 counted now carry one, or 96.0 percent; in Rules and reimbursement the same share moves from 95.0 percent to 94.5 percent. The lower figure is the one printed in each case. The threshold fixed on this page before the first count was half. Two sections sit below it and stay flagged, How we work and What we measured, and both were flagged after the first audit as well. Rules and reimbursement was flagged after the first audit at 31.6 percent and is no longer flagged: the second audit added 457 claims of that kind to it, and 451 of them carry an outside source. Our methodology page describes the route a claim is meant to travel before it reaches a page, and a flagged section is one where that route was not followed to the end.

Read by language, this evidence audit found the picture uneven, and the unevenness is published rather than averaged away. The first audit counted 400 English claims with 358 first-hand, 98 French claims with 62, and 87 Arabic claims with 69. The second counted 643 English claims with 634 first-hand and 1,011 French claims with 997, and it reached no Arabic page at all, because none of the 25 pages it covered exists in Arabic. The third counted 412 Arabic claims with 375 first-hand, and it reached no English or French page, because all 23 pages it covered are Arabic. Taken together the English pages carry 1,043 claims with 992 first-hand, of which 500 need an outside source and 463 carry one, or 92.6 percent; the French pages carry 1,109 claims with 1,059 first-hand, of which 704 need an outside source and 669 carry one, or 95.0 percent; and the Arabic pages carry 499 claims with 444 first-hand, or 89.0 percent. That Arabic figure moved further than any other on this page, and where it moved from is worth naming: until 10 September 2026 it described five Arabic pages and 87 claims, of which 12 needed a source somebody else published and not one carried it. The third audit examined twenty-three Arabic pages, and the named sources it found there — PubMed records, ministry documents and other institutional documents — are what the larger Arabic count now rests on. The outside-source split does not travel with it, because the third audit did not record that split: the only Arabic figure of that kind this page holds is still the 12 and the none the first audit produced, and it describes five pages rather than twenty-eight. After that first audit, one section in one language sat above the threshold — Rules and reimbursement in English, at 70.6 percent — and it held every externally sourced claim of that kind on the site; the second audit ended that, because the French pages it covered carry 669 such claims of their own.

Every claim left standing on no qualifying source received one of the three outcomes named on this page, inside the audit that found it. Of the 96 the first audit found, 24 were sourced, in the sense that a first-hand record was found and added; 63 were rewritten as explicitly labelled uncertainty, with no new claim and no new number introduced; 6 were removed, because the audit measured them to be wrong rather than merely unsupported; and 3 reached none of the three outcomes and stay listed as unsourced. Of the 23 the second audit found, none was sourced, 15 were rewritten as explicitly labelled uncertainty, 8 were removed, and none is left listed as unsourced. The second audit changed three further sentences that did carry a source: two attributed a quotation and a figure to the wrong record, and one attributed a wording to a published text that writes it differently. Twenty-six sentences were changed by those findings; no block was deleted and no sentence gained a new clinical claim. Fifteen further changes on six pages closed the same defects where they sat outside any one auditor's group, and our corrections log names them. The six removals of the first audit fall into two groups: three pages described two Arabic spellings as being answered differently by one keyword tool, where our archive holds a single spelling read in two markets, and three pages carried a French search total that no harvest file of ours contains. Those six, and the eight removals of the second audit, are written up in our corrections log. The third audit is recorded differently, and this page will not sort it after the fact into those three outcomes, because its own files do not. What those files record is this: it marked twelve findings red and fourteen slight; all twelve red and thirteen of the fourteen slight were rewritten on the pages that carried them; one slight finding was left standing and written down instead, because six quotations on one Arabic page are cut at the line end of the file they were taken from, so a reader can read them but cannot check them against the source as printed. No block was deleted, no sentence gained a new clinical claim, and every quotation the three auditors who kept a count checked — 244 of them — was found byte for byte in the source it names. That 244 is what three of the four auditors counted, not what the audit as a whole read: the fourth kept no quotation counter, and its own notes confirm 21 further quotations as byte-identical without ever totalling them.

Three claims stayed unsourced, one in each language, and they are three versions of the same claim: the line explaining what the Turkish abbreviation in our byline stands for. It sits inside an entry in our corrections log, and that log accepts additions only, so an entry is never rewritten or deleted once it is published. Sourcing the line, softening it into labelled uncertainty and deleting it would each have meant rewriting a published entry, which our own rule forbids. Naming the deadlock is the honest option; closing it means a new entry, placed above the old one, that ties the line to a named source.

Three limits belong beside the numbers of this evidence audit. The count was made by the people who write the pages: no outside auditor read it, and no external check of it is planned. And the three audits did not check their sources the same way, which is worth stating rather than smoothing over. The first re-derived 8 figures already published on this site, among them the language breakdown of the clinic reviews we counted, from the raw files rather than taking them on trust from the sentence that carried them. The second went out to the network: of its 1,654 claims, 1,238 were re-derived from the working files we hold, 277 were checked against PubMed records fetched during the audit, 112 against documents fetched live and read, and 9 against copies we hold because the live address refused the request, which is recorded as no verdict rather than as an absence; the remaining 18 could be tied to no file or document at all. That 18 is smaller than the 23 counted above because 5 further claims did have a file behind them and were still rewritten or removed, the file having measured the sentence to be wrong. The third went out to the network too, and it ran a test the first two did not: not only whether a source stands behind a sentence, but whether that source answers it. Its auditors went out to live addresses at least ten times during the audit and read the documents there rather than in our files. One auditor fetched four surfaces and each returned a document: a national gazette issue, a statute PDF, a ministry tariff annex and a health agency page. A second fetched a fatwa page, to test what somebody else's page does and does not say on it. A third read a gazette text, an association archive PDF and a health service page at their own addresses, and its first attempt at the health service page was a guess that returned 404, which is written down here because the correct address returning 200 does not undo the wrong one. A document held by a federal ministry refused a plain request and returned 115,474 bytes to the same request carrying a referrer and a language header. And one address — a dental association's tariff page — returned nothing at all, as did the control fetch of that association's own home page, so what failed there is the channel and not the document, and the page it holds was read from a copy taken the same day. The refusal and the silence are both written down, because a refusal is not evidence of absence and printing only the refusal we recovered from would be that same error facing the other way. The third limit is the one this page's own definition makes. A primary document, as defined above, includes published research read in full rather than through its abstract. The third audit did not read in full: it read the PubMed record — the title, the citation and the abstract — and 181 of its 375 first-hand claims rest on that, while not one of them rests on a full text. So the 91.0 percent printed above is measured against a looser standard than the one this page prints. Measured against the printed standard, the arithmetic is 375 minus 181, which is 194 of 412, or 47.1 percent. Both numbers stand here rather than one: 91.0 percent is what the audit's own verdicts give and what the pooled 94.1 percent is built on, and 47.1 percent is what this page's own definition of a primary document gives when applied to the same 412 claims. The 181 are set aside at the stricter figure for one reason, and it is worth stating plainly: a PubMed record carries a paper's title, its citation and its abstract, and not its full text. This is not a correction of the audit, whose files are unchanged; it is the difference between two standards, printed rather than chosen between. Those derivations are kept under denetim/kanit-0905, denetim/kanit-0906 and denetim/kanit-0910-ar-denetim in our working files, which a reader cannot open, and saying so beats implying a trail we do not publish. The next audit runs six months after 10 September 2026, or sooner once 130 pages stand published, whichever comes first.

## Sources used on this page

- _calisma/bir-yildiz.json
- _calisma/yorumlar/*.json