One Company, Every Lens, Six Years Deep: How the Health Check Reads a Thai Stock Before You Do — and Where the AI Summary Is Not Allowed to Go

This is a sentence an auditor signed, about a company most Thai investors have bought something from this week:

"The valuation of inventories was focused on because the estimation of net realisable value of inventories involves management’s judgment relating to estimate quantity and valuation for deteriorated, damaged, obsolete and slow-moving and shrinkage inventories."

— CP ALL (CPALL), auditor's report, key audit matter, labelled FY2025 · Form 56-1 on the SEC site · p. 3

The same sentence, word for word, appears under the labels FY2024 and FY2023 as well. That repetition is not a scandal and I am not presenting it as one. A key audit matter appears in perfectly clean reports — it is where the auditor spent judgment, not where the auditor found a problem. The point is different: you would only know it repeats if you had read three years of auditor's reports side by side. Almost nobody does that for one holding, let alone twenty.

That is the job of the Health Check. This article explains exactly what it does, what the AI summary written on top of it is and is not allowed to say, and how I would use the two together before an earnings release or an investment committee.

The question a search engine answers badly

Most of what Boom Leverage Terminal does is cross-sectional: type a theme, and see which companies across the market discussed it. That is a retrieval problem, and semantic search is the right instrument for it.

The Health Check answers the opposite question: "what did this company disclose about itself?" And for that question, retrieval is the wrong instrument twice over.

• Top-k truncates. A search returns the passages closest to your query. For one company over six years, the closest 25 passages are not "everything relevant"; they are whatever ranked highest for the words you happened to type.
• The answer depends on your phrasing. If you did not think to ask about working capital, a search will not volunteer it.

So the Health Check is not a search. Every finding in the corpus is already labelled by domain and theme at the moment it is extracted, and the card is built by a metadata pass over those labels for one ticker — no embedding, no relevance floor. One request, and it returns every extracted, citation-verified finding for that company, sorted into fixed clusters.

Say it precisely: the card is complete over what the system extracted and verified, not over everything in the document. A finding the extractor never produced cannot appear. The on-screen coverage line says exactly that — "every item the system extracted and proved against the original file" — and never "everything in the filing". If you remember one sentence from this article, make it this one.

The eight clusters, in the order the card shows them

The order is fixed in code and mirrors the engine's own definition. It does not reshuffle per company, which is what makes two companies' cards comparable at a glance.

The headline strip counts only three of them — red flags, performance and forward view. Eight numbers at the top of a page is the information overload this feature exists to remove; the rest are sections you open.

Two of those descriptions carry the phrase "not a warning sign", and that is deliberate. Every listed company has rows under operating quality. A warning colour there would paint the entire market and teach you to ignore the colour — which would cost you the one cluster where the colour does mean something.

Worked example: CPALL, measured 11 October 2026

I ran the card for CPALL against the index as served on 11 October 2026, over the six-year window FY2021–FY2026 (the depth a six-year pack sees; the free tier sees roughly the latest five quarters). Here is what each cluster held.

502 findings matched in total, and one was withheld because its quote could not be found in the source file. These are counts of extracted findings, not a score. FY2026 is a partial year, which is why its counts are lower — that is calendar, not a trend.

Before you read a number into this table: a count is how many passages were extracted, not how serious anything is. 241 performance findings means management explains its results at length, quarter after quarter. It says nothing about whether those results were good. This product deliberately assigns no score to any of it.

Now the rows behind the counts. Every quote below is byte-for-byte what the card serves, with the SEC filing and the page the citation check found it on.

Auditor observations — one, from 2021

"I draw attention to note 4 to the financial statements, during the year ended 31 December 2021, the Group completed the business combination by entire business transfer of Lotus’s Group, the independent valuers are still in the process of appraising the fair value of the identifiable assets acquired and liabilities assumed from the business combination."

— CPALL, auditor's report, emphasis of matter, FY2021 · SEC filing · p. 4

Read it for what it says: a purchase-price allocation still in progress after an acquisition. It is an emphasis of matter — the auditor pointing at a note — and the opinion was not modified, which is why the red-flag cluster is empty.

Disclosed risks — the newest rows

"Elevated energy prices resulting from uncertainty surrounding conflicts in the Middle East increased business operating costs and the cost of living."

— CPALL, MD&A, FY2026 Q2 · SEC filing · p. 1

"Consumers remained cautious in their spending and placed greater emphasis on value for money."

— CPALL, MD&A, FY2026 Q2 · SEC filing · p. 1

Operating quality — the working-capital lens, quarter by quarter

This is where the cluster earns its place. The same sentence structure recurs across quarters, and lined up it becomes a series the company wrote itself:

Each quote begins with "Consequently, the Company and its subsidiaries" in the source; I trimmed only that opening to fit the table. I am not telling you what a negative cash cycle means for this business — the point is that the company's own number, in its own words, is now a time series you can check in one click per row.

Performance — management's own explanation

"In 1Q26, the Company reported net profit at Baht 9,118 million, increasing by 20.2 percent YoY."

— CPALL, MD&A, FY2026 Q1 · SEC filing · p. 3

"In 2Q26, the Company reported total revenues at Baht 265,760 million, increasing by 3.6 percent year-over-year (YoY)."

— CPALL, MD&A, FY2026 Q2 · SEC filing · p. 2

Forward view — what they said they would do, then whether they did

"In 2025, the Company has expanded new stores in total of 700 stores in Thailand, including corporate, store business partner (SBP) and sub-area stores, which reached the target at 700 new stores per year."

— CPALL, MD&A, FY2025 · SEC filing · p. 4

And from the other-findings cluster, four years earlier, the same target:

"The Company has planned to open 700 new stores per year, which during 2021 CPALL has successfully opened 702 new stores."

— CPALL, MD&A, FY2021 · SEC filing · p. 3

A stated target, and the company's own report against it, five years apart. That is the kind of read-across I describe at length in tracing one company's narrative across periods — the Health Check just puts both ends on the same card.

Where the auditor focused — and the repetition

Back to the hook. The inventory key audit matter appears with identical wording under FY2023, FY2024 and FY2025 (FY2023, p. 3 · FY2024, p. 3 · FY2025, p. 3). So does a second one:

"The Group has significant goodwill and other intangible assets with indefinite useful life which was mainly arising from a business acquisition. The management normally assesses the impairment from the estimated recoverable amount at the end of the period. The recoverable amount was based on its value in use, determined by discounting future cash flows to be generated from the continuing use of assets and related assumptions."

— CPALL, auditor's report, key audit matter, FY2024 · SEC filing · p. 3

Stable KAMs year after year are common and usually mean the same judgment areas persist. A KAM that appears, disappears, or changes wording is the thing worth a second look — and you can only see that with years side by side. Which is what the year chips are for.

Year chips: rewinding one cluster's timeline

The card's first fold shows each cluster in importance order — auditor opinion first, then risk signals, then the latest year — with the label "important first". For a company with years of rows, that means you never walk back through time.

Under each cluster sits a row of year chips. Each chip shows how many findings that year holds; a year with zero cannot be pressed. Press one, and that cluster's rows for that fiscal year open — up to 25, which is the engine's own per-cluster ceiling.

The rules the code enforces:

• 1 credit per (company, cluster, year). Re-pressing the same chip, or coming back to it after re-running the scan, is free within the same credit window.
• The company's own Health Check must already have been bought in that window. A year view cannot be a cheaper back door into the 3-credit card.
• Refunded if the engine is down or the year sits outside your pack's depth window.

The model-validation view — read the timeline, not the year: one year of disclosed risks tells you what management worried about once. Six chips across the same cluster tell you which worries are permanent furniture, which ones arrived, and which ones quietly stopped being mentioned. In model validation we never sign off on a single snapshot; the drift is the finding.

The AI summary over a Health Check — what it may and may not write

Since 11 October 2026, a signed-in reader can switch on AI summary over three things already bought on /scan: a search result set, a Health Check card, and one cluster-year from the chips. It is a layer over evidence you already have, not a new search.

For a Health Check, the summary is bound by rules written into the prompt and checked again by the validator afterwards.

Shape.

• Headings follow the card's cluster order, one heading per cluster, then 1–3 bullets per cluster.
• Every bullet names the period as labelled on the source (for example "FY2026 Q1") and ends with a numbered citation [n] pointing to a source you can open.
• A cluster with no sources gets no heading. The card already shows which clusters are empty; the model is forbidden to fill or guess one.

Voice.

• Red flags, auditor observations and where the auditor focused come from the auditor's report, and the summary must relay the auditor's own wording, not interpret beyond it.
• Operating quality is a comparative reading of the MD&A, not a warning signal and not a fraud check. The model may not use the words fraud, abnormal or warning sign unless the source itself uses them.
• ▲ and ▼ may appear only where the source itself states the direction.

A single cluster-year (from a chip) gets a tighter format: no headings, 2–6 bullets ordered by the weight the sources give them, and when several sources say the same thing they are merged into one bullet that cites all of them, for example [2][5].

Years. This rule exists because of a real incident. In the production test on 11 October 2026, a CPALL Health Check summary came back with "ปลายไตรมาส 1/2566" and "1Q66" for sources labelled FY2026 Q1 — the model had "translated" a Western 2026 into the Thai year 2566. The correct Buddhist-era year would have been 2569. The prompt now forbids converting between calendars or abbreviating years to two digits, and the validator cuts any line carrying a Thai year that none of its cited sources can account for.

Why I let a model near this card at all: the summary is not trusted; it is checked. Every factual line must carry a citation to a numbered source, every number in that line must appear in the source it cites, and a line that fails is cut before you see it. How that validator works — and why a mostly-cut answer is not charged — is its own article: how every line of an AI answer gets checked.

To be explicit about what I am not showing you: I have not pasted an AI-generated CPALL summary into this article. The quotes above are the source rows themselves. If I hand-composed the shape for illustration, it would look like this — every bullet built only from rows quoted earlier:

Illustration only — hand-composed from the quoted rows, not captured AI output:

Auditor observations

• FY2021: the auditor drew attention to the Lotus's business combination, with independent valuers still appraising fair value [1]

Operating quality

• FY2026 Q2: cash cycle of approximately negative 23 days as of 30 June 2026 [2]

Where the auditor focused

• FY2023–FY2025: inventory valuation recurs as a key audit matter, with the same wording each year [3][4][5]

What it costs

From the credit table in the code, as of 11 October 2026:

The same summary requested again within 24 hours is replayed for 0 credits; pressing "try a fresh summary" runs the model again and charges again. The AI summary needs you to be signed in, and model usage is also capped per week by plan (the free tier's weekly allowance is 30,000 output tokens). A free account gets 100 credits a day, so a Health Check plus a summary — 5 credits — is well within reach. Full pricing is in the pricing article.

A workflow you can copy: one holding, before results or before committee

1. Run the Health Check on the ticker. Look at the red-flag count first. Zero is the expected answer for most companies; anything else gets read in full before anything else.
2. Open auditor observations and where the auditor focused. Note which KAMs are stable year to year. Press the year chips on "where the auditor focused" for the last three years and look for a KAM that appeared, disappeared, or changed wording.
3. Open operating quality and pick the one or two lenses that matter for this business model (working capital for a retailer, customer concentration for a contract manufacturer, FX for an exporter). Use the year chips to turn those rows into a timeline.
4. Read forward view against later performance. Put what management said it would do next to what it later reported — the 700-store pair above is the template.
5. Switch on the AI summary only now, as an index over what you have already read. Use its [n] citations to jump back to rows, not as a substitute for them. Export it as Markdown for your notes; the export carries the citations and the quotes.
6. Take the open questions somewhere else. Anything the card raises but does not answer — "what did management say about this on the call?" — is a question for the chat over 56-1, MD&A and Opp Day, framed as a theme, as described in the theme-not-year research workflow.

Four traps I see people walk into

1. An empty cluster is not a clean bill of health. Zero red flags means no modified opinion or going-concern paragraph was extracted and verified inside your depth window. It does not mean none exists outside the window, and it does not certify anything about the business. "Not found by this system" and "not there" are different sentences.

2. Filings lag. The newest Health Check row is only as fresh as the last filing. A quarterly MD&A arrives weeks after quarter-end; an annual auditor's report arrives months after year-end. Between filings, the card cannot know what you know from the news.

3. Auditor wording is the evidence — paraphrase is not. "Emphasis of matter" and "qualified opinion" are different instruments with different consequences, and a summary that blurs them is wrong even if it sounds reasonable. That is why the card files them in different clusters and why the AI is told to relay the auditor's own words. When it matters, open the page.

4. One year is a snapshot; the chips are the film. A single year of disclosed risks or KAMs tells you almost nothing about change. The repetition in the hook only became visible across three years. Budget the extra credits for the chips on the one or two clusters that matter.

For how the auditor clusters relate to the whole market, see the 68 companies whose audit opinion was not clean; for the operating-quality lenses read together across companies, see when all six lenses light at once. For the cross-sectional cousin of this feature — a summary of a whole result set grouped by who gains and who loses — see the AI summary on /scan.

Overview

The Health Check answers the holder's question — what has this one company disclosed about itself — with a fixed eight-cluster card built from labels, not from how well you phrased a search. The year chips turn any cluster into a timeline. The AI summary sits on top under rules that keep it inside the evidence: fixed order, a citation on every line, no invented sections, the auditor's words, no alarm vocabulary the source did not use, and no charge if it fails the check.

Try it on a stock you own: sign in at terminal.boomleverage.com, type the ticker into the company lookup, and run the Health Check. The free tier's 100 daily credits cover the card, a few year chips and a summary.

For a desk, an investment committee or a risk team that wants this across a whole coverage list — multiple seats, deeper history, bulk export with the source passages attached — the scoping conversation is at boomleverage.com/th/enterprise or contact@boomleverage.com.

Disclaimer: This article is written for education and to explain how to read disclosures that listed companies publish themselves. It is not investment advice, it recommends no security, and it guarantees no outcome. · CP ALL is used as a worked example because it is large and well covered; nothing here is a view on the company, its management or its shares · Every quotation is reproduced verbatim from Form 56-1 filings and MD&A published on the SEC website, linked at the page where the citation check located it · Counts are counts of extracted, citation-verified findings as served on 11 October 2026 for the FY2021–FY2026 window; they are not scores and they change as filings enter the index · Classification into clusters and lenses is the output of automated systems, not the opinion of a licensed professional or an auditor · AI summaries may be wrong — check every line against its cited source before relying on it · Investing carries risk; past results do not guarantee future ones