Method and audit note · Uni-Asia Group (SGX: CHJ) · 17 August 2026

Download the historical 17 August PDF — the PDF predates the 9 September count clarification below. · back to the evidence library

How the Uni-Asia coverage was built, and what went wrong on the way

A method-and-audit note accompanying the Uni-Asia Group Limited (SGX: CHJ) evidence library of 17 August 2026.

This note exists because the initiation itself should read as a research product and nothing else. Everything about how it was made — the sources, the reconciliations, the mistakes, the independent review and the reasons the rating is what it is — lives here instead. It is written to be read on its own by anyone deciding how much weight to give the conclusions.

1. What the coverage is built from

Nothing was taken from a data aggregator. That was not fastidiousness: during collection one aggregator reported Uni-Asia's last half-year net income as "11.61 K SGD" against a reported US$5.187m, and another quoted EBITDA on an undefined basis. For a company this size the aggregators are not lightly wrong, they are wrong by orders of magnitude.

Everything reconciles or it is disclosed. The workbook carries 18 deterministic checks that recompute from the underlying figures rather than storing an expected answer: the balance sheet balances in all nineteen years; segment rows sum to group profit before tax in all twelve periods; the halves sum to the full years; hire days multiplied by the daily charter rate reproduces reported charter income to within 0.3%; and the modelled effect of the vessel-life change reconciles to the issuer's own stated figure. Two residuals are disclosed rather than plugged — a US$4.5m foreign-exchange effect on cash in FY2008, and the FY2007 earnings-per-share figure, which reflects the August 2007 listing date.

2. The independent review, and what it found

A first-pass blind review was commissioned on 17 August 2026: a reviewer with no prior context on the company or the pack, working from the primary documents rather than from the author's transcription, and deliberately not given the author's own list of suspected weak points.

Historical review record, reconciled 10 September 2026. The recovered 17 August finding-by-finding register records 32 findings: 23 marked fixed, eight acknowledged and one explicitly unresolved, F2-09, concerning an unretrieved IFA letter. The previously quoted total of 31 omitted that unresolved item. The register's prose also misstates the severity-3 split; its individual entries record eight fixed and five acknowledged. This reconciles the finding count only. It does not independently verify the fixes, resolve the historical evidence gap or certify the current report. The dated examples below remain a historical record.

Two findings changed load-bearing numbers, and both made the analysis less flattering:

Three of the eight severity-1 findings came out of a single nine-page PDF that was already in the document folder: the 30 April 2026 AGM minutes. They produced the buyback-mandate defeat, the vote arithmetic, and management's own characterisation of the maritime-asset-management segment's income as "inherently non-recurring" — which materially qualifies what had been the pack's headline insight. Having a document is not reading it.

A second guided pass was then run on the corrected pack, with the author's known-defect list supplied, specifically to test whether the corrected version supports a rating.

2b. The second review, and why it blocked the rating

A guided pass then reviewed the corrected pack, with the author's known-defect list supplied and one question to answer: is this fit to carry a rating? It returned 23 findings, 6 at severity 1, and the answer was no — not on the document as it then stood.

None of the six was an analytical error. All six were failures to propagate:

• The corrections from the blind pass had been applied to the workbook, the canonical JSON, the understanding sheet and the website — but not to four numeric tables in the note, which then contradicted every other artefact and its own prose.
• The scenario table averaged two anchors, which embedded an undisclosed 50% weight on asset value. That made the pack's own base case a Buy while its stated conclusion was a Hold. The pack contained three mutually exclusive ratings and nothing in it chose between them.
• The forecast earned a tenth vessel's charter income without ever buying, depreciating or financing it. Correcting that cut FY2027 earnings per share by nearly a quarter.
• Two documents had been declared unreachable that were not: an appendix filed as the second attachment to an announcement already in the file, and an independent-financial-adviser letter inside an EGM circular that had never been attempted.

The structural cause was named precisely: no check anywhere compared a number as printed in a document to the model that produced it. One was written in response. On its first run it found 83 stale values across ten documents.

The reviewer's own recommendation, arrived at independently, was the same rating the author had reached — which is reassuring about the conclusion and says nothing at all about the document that was meant to carry it.

3. The mistakes worth naming

Six errors were caught before publication. Four were caught by checks that compared a figure to something the issuer had independently published; checks that compared the author's arithmetic to the author's other arithmetic caught none of them.

Depreciation was modelled over the wrong life. The company extended vessel useful lives from 20 to 25 years and stated the effect: about US$4.1m less depreciation in FY2026. The first model depreciated carrying value over a full 25 years and produced US$5.4m of annual depreciation and a US$1.3m benefit. Two cross-checks failed simultaneously — the reported first-half charge annualised to US$8.7m, and the issuer's stated benefit was three times the model's. Depreciation runs over remaining life on existing carrying value; a ten-year-old fleet on a 25-year policy has about 15.4 years left, not 25.

The scenario tab was briefly a second, contradictory model. Its formulas referenced row numbers typed by eye, and after edits shifted the rows the base scenario produced FY2027 profit before tax of US$30.9m against the forecast tab's US$9.5m. It displayed drivers, computed outputs and was internally consistent — and was wrong. There is now a check requiring the base scenario column to reproduce the forecast tab to the dollar.

Several checks tested nothing. The first version of the checks tab referenced guessed cell coordinates. Five failed with nonsense residuals, which is how the problem was found. The worse outcome was available: had the guesses landed on numerically innocent cells, the checks would have passed and the pack would have shipped with a quality-control tab that proved nothing.

Two claims wrongly counted as corroborated. An annual report and a results announcement were treated as two independent confirmations of the same fact. They are one origin — the issuer — and the contract rejected them. Document count is not evidence independence.

A summary sentence was typed from memory. A tab summary read "cumulative operating cash flow US$194.5m against cumulative profit of US$3.3m". The computed figures were US$189.2m and US$15.0m. It survived until a chart, which computed the same sums from the data, printed different numbers on its face.

A seventh error was found after publication, by the check written in response to the sixth. The blind-review corrections reached the note, the workbook, the canonical JSON and the public page. They did not reach the presentation deck. The deck is a .pptx, and the script written to compare prose against the model read Markdown and HTML only — so nothing ever looked inside it, and it went into the private file still stating realised return on equity as "0.3%, or 4.6% excluding the FY2024 write-down" and cumulative ship-owning profit as "+US$47.1m … leaving −US$1.6m": the figures the review had already replaced with 1.0% / 3.5% and +US$45.3m / −US$3.3m. Five statements across four slides, contradicting every other artifact in the pack. The same scan, rebuilt inside the validator so that it opens PowerPoint and Word as well as text, found all five on its first run — together with a segment-asset figure the deck rounded to US$150m and the note to US$160m against a computed mean of US$156m. The lesson is not "check the deck". It is that a checker which reads some of the deliverables tells you nothing about the ones it cannot open, and formats that cannot be read are now reported as skipped rather than passing in silence.

Four chart defects were invisible in the code — clipped axis labels, two overlong captions, a legend sitting on an axis, and negative currency rendering that put the minus sign after the unit rather than before it. A contact sheet found all four in one look.

4. How the rating is arrived at, and why it is not published here

The rating is derived, not chosen. A band is fixed before the fair value is computed — a return threshold above which the rating is a Buy, another below which it is a Reduce, and Hold between — and the spreadsheet applies it with a formula, so the printed rating cannot disagree with the model. A check in the workbook fails if it ever does.

The fair value itself is an explicitly weighted blend of three anchors, and the weights are where the judgement sits:

Two of those three share a forecast numerator, so they are labelled correlated cross-checks rather than independent confirmation; only the disagreement between the asset anchor and that pair carries information, and it is a wide disagreement.

Why the asset anchor is capped at a quarter weight. Book value is only worth book value if something converts it into shareholder value. On 30 April 2026 the buyback mandate was defeated by exactly the controlling shareholder's holding, against every other share voted. No dividend increase is in prospect while two newbuilds are being funded, and no approach is evidenced. A book discount with no realisation path is a structural feature rather than a temporary mispricing, so the asset anchor is both time-discounted and weighted down.

What argues the other way, and it is the strongest counter-evidence in the file: between 19 May and 15 July 2026 a new holder, Asparta Pte. Ltd., bought 1,340,500 shares on-market — about 1.7% of the company and roughly 34 days of average volume — taking it from just under the 5% disclosure threshold to 6.704%, at an average of S$0.926. Precious Shipping added over the same period. Someone with more time on this name than any outside analyst has been paying around the prevailing price.

The honest limit. The single biggest driver of the forecast — which vessels are on charter, to whom, at what rate and until when — is disclosed nowhere in nineteen years of filings. No amount of further work closes that, which is why the analysis ends in a range and a band rather than a point estimate with false precision.

The rating, the fair value and the expected return are not published on this site. They sit in the private working file. This page describes how they are produced so that a reader can judge the method; it deliberately does not state the output.

5. What is still open

6. What a reader should take from this

Three things.

The corrections went one way. Every material correction made the company look worse, not better — the return-on-equity series, the cumulative segment result, and the qualification of the fee-segment insight. That is the pattern you want to see from a review; the opposite pattern would suggest the corrections were being chosen.

The checks that worked were the ones with an external referent. The internal arithmetic was self-consistent throughout, including when it was wrong. What caught errors was comparing to something the issuer had published independently: a reported charge, a stated benefit, a disclosed hire-day count.

Coverage is a state, not an event. The FY2026 result in late February 2027 is the first clean read of the renewed fleet, and it will settle more than this note can. The estimates behind the analysis are frozen in a forecast scoreboard so that the error can be measured against the print rather than quietly forgotten.

For information and analysis only. Not investment advice or a recommendation. Information and market data through 17 August 2026, 18:30 SGT.