Q & M Dental Group (SGX: QC7) — how this pack was built, checked and corrected

A companion to the initiation. It carries no rating, no fair value and no view on the share price — those live in the note itself. This document exists so that a reader can judge how much weight the work bears, and where it is weakest.

Information cutoff 17 August 2026. Prepared 18 August 2026.

Why this is a separate document

The initiation report is written for someone deciding what to do about a security. It should carry the argument and the evidence and nothing about the machinery that produced it. But the machinery is not irrelevant: a reader is entitled to know what was independently checked, what was found wrong, and what remains unverified. Putting that inside the note buries the argument; leaving it out entirely asks for trust that has not been earned. So it goes here.

This pack went through three review rounds and changed its mind, in writing, several times. That is recorded below in full rather than tidied away. Three of the corrections reversed a conclusion the draft had already argued in print.

The evidence base

The subsidiary stream is the finding that mattered most, and it was nearly missed. Q&M consolidates Aoxin Q & M, which is separately listed on Catalist and files its own announcements. The first document sweep covered only the parent's tape. Everything in section 8.4 of the initiation — the RMB 526m of subsidiary-level acquisitions, the share-count doubling, the deconsolidation arithmetic, the Panjin hospital closure, the police report, the second SGX query — comes from that separate stream and appears nowhere in Q&M's own filings.

The general lesson, recorded for the next run: when a listed company consolidates another listed company, sweeping the parent's announcements is half the job.

What was independently checked

Three review rounds, two of them by a different model family.

Every finding was dispositioned in research/red-team-dispositions.json. All 17 blind findings were confirmed — none was rejected as a reviewer error.

The model was re-evaluated with an independent formula engine. The workbook holds live formulas, not values; an engine outside Excel recomputes the whole thing and every structural identity returns residual zero. What that proves and does not prove is set out under deterministic checks below, and the distinction matters more than the result.

Charts were inspected on rendered contact sheets rather than trusted from the plotting code. Two defect classes were found that no code review would have caught: annotation collisions, and paired dollar signs being consumed by the plotting library's maths renderer.

Correction log

Corrections that changed a number or reversed a conclusion. Ordered by how much they moved.

Reversals — a printed conclusion was withdrawn

Numbers that moved

The largest omission

Profit-guarantee receivables were missed entirely on the first pass. The draft argued that guarantee income is low-quality earnings because it is booked when acquisitions miss. It did not notice that S$15,123,000 of it had never been collected, that the auditors flagged S$7,386,000 as of uncertain recoverability, and that the impairment assessment is a Key Audit Matter. Found in round 3 by reading the Key Audit Matters section rather than the revenue notes.

Errors of arithmetic and units

Three unit errors were caught before publication, each an order of magnitude:

• EPS computed in thousands against a share count in units.
• NTA per share stated as 27.47 SG cents against the correct 2.747 — a clean factor of ten.
• Model PAT of S$38m against an actual S$9.9m, because "segment result" in this issuer's disclosure is struck before depreciation, amortisation and finance costs. The bridge was then verified to the cent: 54,464 − 6,441 − 5,094 − 13,794 − 827 − 16,020 − 103 = 12,185.

A correction made on the final day

One block of the valuation records items deliberately left out of it. Its first version capitalised the profit guaranteed to Aoxin under its two China memoranda without netting off what Aoxin pays for that profit. That is not an accretion calculation; it is a gross-earnings calculation dressed as one, and it overstated the omitted amount by roughly four times. Redone properly — earnings capitalised, less Q&M's economic share of the RMB 526m consideration, less the subscription needed to hold 50% through the resulting share issuance — the figure is far smaller. It is recorded here because it was the author's own error, found by the author, on a line item that would otherwise have made the conclusion look more robust than it was.

What remains unverified

Named, not waved at.

1. No independent valuation of any of the four acquisition targets exists, from any source. None was required under the listing rules and none was obtained. Every acquisition multiple in the initiation rests on management-supplied unaudited figures. This is a permanent gap in the evidence base, not a retrieval failure.
2. The Rule 1014 circular for Experteeth has not been published. It would carry an independent financial adviser's opinion. It was reclassified from thesis-critical to supporting once the pro-forma financial effects it would contain were confirmed already public in the 12 July announcement — and because an IFA opinion addresses whether terms are fair and reasonable, not whether a guarantee will be delivered, which is the question the note turns on. That reclassification is a judgement, recorded with its date and reasoning in research/decision-log.md, and a reader is entitled to disagree with it.
3. The China dental-centre count does not reconcile. Aoxin's 23 March 2026 announcement states 14 centres; the parent's 1H2026 release states 11; only one closure was announced between them. No China centre count is printed anywhere in the initiation. Closing this needs Aoxin's FY2025 annual report.
4. An entity-level non-controlling-interest build for Aoxin cannot be constructed from the disclosure. The forecast applies a single consolidated historical rate, which is stated as a limitation at the point of use in the workbook.
5. The driver of the S$10.1m increase in non-controlling interests over 1H2026 is not established.
6. The outlet-count roll-forward — opened, acquired, closed — is not disclosed by the company.
7. The realisable value of 138m Medi Lifestyle shares, were the EM2AI disposal to complete, is not determinable; the consideration is illiquid sub-5-cent Catalist paper.
8. Aoxin's two China acquisitions are memoranda of understanding, not agreements. Their guarantee mechanics — waterfall, escrow size, put option, personal guarantees — are not disclosed at all, and whether Aoxin can fund RMB 263m of cash consideration is not evidenced.
9. No traded-volume series was obtained, so the initiation states float value and declines to state ADTV or days-to-build.

Deterministic checks, and what they are worth

The workbook carries two kinds of check, and the distinction is the point.

Structural identities cannot fail. Nine checks confirm that, for example, organic PBT equals the sum of the lines it is defined as the sum of. Each subtracts exactly the cells its target adds, so each is algebraically zero by construction. They confirm the workbook was assembled as intended; they test no economic proposition. An earlier version of this pack cited "all residuals zero, independently verified" as evidence the model was correct. That claim was wrong and was withdrawn. The checks are now labelled as tautologies in the workbook itself.

Ten economic tests were added that can fail. Each compares the model against something it does not itself produce — among them: that the base-case forecast must come out below the issuer's own pro-forma, and that the difference must reconcile to acquisition financing plus cost inflation within S$1,000k; that the growth-payout identity holds (g = ROE × retention must stay below the cost of equity); that base organic growth does not exceed the perimeter-clean observed rate by more than 200bp; and that economic net debt exceeds covenant net debt, since leases sit outside the covenant definition. All ten pass. Unlike the identities, they could have failed.

Copy-across tests were added after they caught a real defect. A scenario column had silently collapsed onto another — base and bull covenant equity both referencing bear PATMI. The tests now fail if that recurs.

Retrieval: what did not work, and what did

Recorded because the failures cost hours and are reusable.

• Line endings broke every retrieval URL. A script writing URL lists on Windows emitted CRLF; the carriage return travelled into the request and every fetch returned HTTP 000. Strip line endings explicitly rather than trusting the file.
• A parallel sweep produced a false negative rate of 82%. Twenty concurrent document fetches reported an 18% hit rate. Retried serially, the same twenty documents returned 30 of 30. The failures were throttling, not absence. Concurrency was capped at eight thereafter. A retrieval failure under load is not evidence a document does not exist.
• Standard PDF layout extraction misaligned every segment table. Columns drifted across pages. Replaced with a purpose-built extractor that buckets words by vertical position into visual lines and keys numbers by their horizontal coordinate, so a column is identified by where it is on the page rather than by whitespace. Every extracted block was then verified by segment arithmetic before use.
• Charts must be looked at, not merely generated. Two defect classes survived code review and were caught only on a rendered contact sheet.

Provenance and independence

The blind review was performed by a model from a different family, working from the pack alone against a fixed build hash, before seeing any author-supplied list of suspected weaknesses. The third round was performed with context isolation. Part of the subsidiary-stream retrieval was carried out by a third system and every item it returned was treated as unverified until the primary filing was retrieved and read by the author — two items it supplied were corrected on verification, including a transaction date that independent sources placed two days earlier.

The author's own sceptical pass does not count as independent review and is not presented as one.

Behind the lock

A complete private working view of this company exists — the initiation report with its rating, fair value and scenarios, the model workbook and the full review artefacts. That material is not published and not available for sharing. This document deliberately contains none of it.

This document describes process only. The rating, fair value and investment view are in the initiation report. Not investment advice.