Revision and methodology note · Civmec Limited · 18 August 2026

How this coverage was built, and what was wrong with it

A record of the research process behind the Civmec coverage: the errors found in the first two versions, how each was found, and what was corrected. Published because the alternative — quietly correcting figures that were already public — is worse. It contains no rating, no valuation and no investment opinion. Updated 18 August 2026 with a second, guided review.

18 August 2026 · accompanies the initiation dated 17 August 2026 (build cvl-2026-08-18-d)

This note records how the Civmec initiation was built, what was wrong with the first two versions, and how each error was found and corrected. It exists separately from the initiation because a research note should carry analysis, not an account of its own drafting. It is published because the alternative — quietly correcting figures that were already public — is worse.

1. Version history

2. Why the first version carried no rating, and why that was wrong

Build -a withheld a rating behind three stated gates: an unquantified FY2025 recognition release, an undisclosed purchase-price allocation for the Luerssen acquisition, and the absence of an independent review.

Two of those three were not true. The 1H FY2026 interim financial statements, published on 12 February 2026 — six months before the cutoff — contained both the segment note that resolved the margin question and the full purchase-price allocation. The third was a process gap, not an analytical one.

The remaining objection was that FY2026 results were imminent. That is a real consideration but it is not a reason to withhold a view; it is a reason to state the horizon and the sensitivity, which the initiation now does. Withholding a rating because a result is coming would postpone every rating indefinitely, since a result is always coming.

The correction is that an analyst owes a conclusion. Qualifications belong in the note, not in place of it. Build -c therefore carries a rating, with bear and bull cases, the derivation of the target multiple, and the specific price levels at which the rating would change. That material is not published on this site.

3. The independent blind review

The pack, the workbook and the underlying filings were given to an independent reviewer with the author's suspected-defect list withheld, and with an instruction to work from the primary documents rather than from the author's claim ledger — because the transcription was the thing under test.

It returned 37 findings: 6 severity-1, 13 severity-2, 18 severity-3. Every severity-1 finding was re-derived from the primary filing before being accepted. All six held; none was rejected.

On the reviewer's independence. It was a cold-context agent instance of the same model family as the author. That makes it genuinely blind — it had no access to the author's reasoning — but not vendor-independent, and it may share systematic blind spots. It is recorded as a supplementary pass. A future vendor-independent review remains desirable.

The six severity-1 findings

Consequences

Finding 4 inverted a headline conclusion. At 0.72x, the price implied an order book near A$1.7bn, above the record — supporting a reading that the market was paying for further growth. At the correct 0.826x it implies approximately A$1,486m, which is roughly the book already in hand. The published conclusion was the opposite of what the evidence supported.

Finding 3 meant the earnings-quality section had missed approximately A$5.5m a year of non-cash contract-liability amortisation running to FY2029 — about 13% of FY2025 profit to owners.

Postscript, 27 August 2026 — finding 3 was right about the disclosure and wrong about the outcome. The A$5.5m a year was computed from the amortisation policy the 1H FY2026 statements stated: straight-line over four years from acquisition. The FY2026 accounts, published 27 August 2026, recognise no release at all in FY2026 and leave the carrying amount at A$22.195m, with the release pattern now “to be assessed”. So the review correctly caught a disclosure the pack had missed, and the number derived from it never materialised. This paragraph is left as written because it records what the review found on 18 August; the correction is on the company page.

3b. The guided review, and the two things it broke

The blind pass was run with the author’s suspected-defect list withheld. A second pass was then run the other way round: the reviewer received the pack, the filings and the author’s own list of where the work felt weakest, and was asked to attack those areas specifically. That is a weaker independence claim than the blind pass and is recorded as such — but it is a stronger test of the places the author already doubted.

It returned 16 findings: 2 severity-1, 8 severity-2, 6 severity-3. Both severity-1 findings were re-derived from the primary filing before acceptance. Both held.

The first: the model repeated an error the historicals had already fixed

The blind pass had found that Civmec books lease interest inside cost of sales, so the profit-and-loss finance-cost line is not the group total, and that FY2025 EBITDA on the correct basis is A$91.7m — exactly the issuer’s own published figure — rather than the A$87.9m the pack had carried. That was corrected in the historical tabs.

It was not carried into the forecast. The forecast EBITDA formulas still added back only the profit-and-loss line. The arithmetic that settles it: building profit before tax from gross profit, other income, administrative expenses and the profit-and-loss finance-cost line reproduces the reported FY2025 figure of A$60,575k exactly; doing the same with total finance costs gives A$56,778k, short by A$3,797k — precisely the lease interest sitting in cost of sales. So the profit-before-tax build is right on the P&L line and the EBITDA add-back needs the total, and the model had used one basis for both.

Correcting it raised the forecast earnings figure and, with it, the valuation held in the private pack. Neither is published here. The rating did not change.

The instructive part is not the arithmetic. It is that a defect found, verified, written up and fixed in one part of the pack was left standing in another part that depended on it, because the fix was applied where the finding pointed rather than everywhere the finding implied. A correction is not complete until every consumer of the corrected input has been re-derived.

The second: a governance claim that the filing contradicts

The note stated that all three independent non-executive directors were replaced on a single day “with no overlap period”. The FY2025 directors’ report says the retirements were “to ensure compliance with tenure requirements under the SGX Listing Rules”, and that “prior to their departure, there was a transitional period during which newly appointed independent directors” had joined the Board.

So the event had a stated, checkable and entirely ordinary cause — a nine-year independence tenure cap — and the overlap the note denied did exist. The claim was not merely unsupported: the document that was the obvious place to check it said the opposite, on a page already read for other purposes. This correction is reflected on the evidence library page.

And one that was accepted only after being re-derived

The reviewer proposed anchoring the target multiple on Civmec’s own trailing range. That number was not adopted on the reviewer’s word. It was reproduced from the pack’s own data — the 52-week price range over FY2025 reported earnings per share — and only then used, and it is now carried as live formula rows in the workbook rather than as prose. Reviewer output is evidence to be checked, not a source to be cited.

What the guided pass says about the blind pass

Six of the sixteen findings were things a blind reviewer would have had no particular reason to look for, because they were the places where the note sounded most confident: the capacity argument, the target multiple, and a claim that a half-year margin had been achieved without a one-off. Confidence in the prose is not a signal of strength in the evidence, and the two passes fail in different directions. Running both is not redundancy.

3c. The multiple that rested on nothing

Build -c justified its target multiple as “a mid-band multiple for Australian contracting peers”. There was no peer multiple set anywhere in the pack. None had been built, none was cited, and the band was never quantified — it was a plausible-sounding sentence standing in for evidence.

This is the same failure the blind pass had already flagged once, under a different name: a number that entered the work as a modelling convenience and was then restated in prose as though it were an observation. The blind pass caught it in the order-book conversion ratio, where 0.72x was a model driver that appeared in the text as a historical fact when the observed average was 0.826x. The target multiple was the same species of error, and it survived the pass that caught its sibling.

The multiple is now anchored on evidence the pack actually holds — the stock’s own traded range — with the direction of the judgement stated against the stock, and with sensitivities given so a reader can disagree at the point where disagreement actually matters. The multiple itself, like every other valuation input, is not published on this site.

4. Root causes

One document, two files. Civmec published the 1H FY2026 interim statements twice: a 24-page scanned copy with no text layer, and a 43-page copy with one. The first build extracted the scanned version, found it unreadable, and moved on. The 43-page version contained notes 12 and 26 — between them the source of three of the six findings. When a filing exists in more than one version, open all of them.

A model driver leaked into prose. The 0.72x figure was an input on the model's driver block. It appeared in the note as though it were a computed historical fact. Nothing recomputed it, because prose is not a consumer that a build pipeline checks. The controls verified the workbook against the filings and the model against its drivers; nothing verified the note against the workbook.

A rule was read and not applied. The methodology this pack follows states explicitly that finance costs for an EBITDA build must include any portion booked in cost of sales. That instruction was read before the model was built and violated anyway. The cheapest available check was also skipped: the issuer publishes its own EBITDA, and one comparison would have exposed the A$3.8m gap immediately.

Sequence. The build order was: build, validate, publish, review, correct, republish. It should have been: build, review, correct, validate, publish. Publishing before the review converted internal defects into public corrections.

A fix was applied where the finding pointed, not where it reached. The lease-interest correction was made in the historical tabs and not propagated to the forecast that consumed them. The pack had no step that asked, for each accepted finding, what else reads this input. It now does.

A clean validator run was treated as a ceiling. Eight severity-1 defects across two reviews is a poor showing for a pack that had already passed a deterministic validator with zero failures. The validator checks structure, provenance and internal consistency; it cannot check whether a sentence about facility size is true, whether a target multiple rests on anything, or whether a fix reached everything downstream. Those need a reader. The validator is a floor.

5. What held

The review confirmed the transcription. Seven years of the income statement, six of the balance sheet and cash flow, five years of segment data and the entire Singapore-dollar era tie line-for-line to the filings, including two subtle reclassifications correctly sourced. Every margin, both return-on-equity bases, net asset value per share, net debt and each multiple recompute correctly, with no order-of-magnitude errors. The customer-concentration figure and the management quotation are exact.

Every defect was in interpretation, definition or completeness — none was in transcription. That is a useful signal about where review effort belongs: not on re-keying figures, but on whether a definition is the issuer's, whether a document is the complete one, and whether the prose still matches the spreadsheet.

6. Verification of the pack itself

• Deterministic validator: 0 failures, 0 warnings, 33 passes, 6 recorded manual sign-offs.
• Workbook identities: 69 checks, all passing, evaluated by an independent formula engine at zero error cells.
• Negative testing: the identity checks were verified by deliberately corrupting two claims — a segment revenue by A$1,000k and a total-liabilities balance by A$500k. Both were caught with the exact residuals, and the pack returned to a pass when the corruption was reversed. A check that has never failed is not evidence of anything.

7. Gaps that remain open

• No vendor-independent review. The blind pass was a sibling-model agent.
• Civmec Defence Industries' contribution is undisclosed, and the issuer states it is prevented from disclosing it by contracted confidentiality obligations. Defence is therefore carried in the valuation at nil.
• The order book has no published conversion profile and has been stated on at least three different bases.
• FY2026 results are imminent and will restate the earnings base. The initiation states its horizon and the sensitivity accordingly.

8. Correction policy

Figures that were published and later found wrong are corrected in place, with the superseded value quoted alongside the replacement so a reader can see what changed. The public evidence page carries a dated corrections section naming each changed figure. Nothing is silently rewritten.

This note is a record of research process. It contains no investment opinion, no rating, no valuation and no expected return. Those are held in the initiation, which is distributed separately and is not published on this site.