Boom Leverage

This is the first edition, so there is no prior version to diff against. The table below is the dated record of what moved in the weeks before publication.

Every listed Thai company files a One Report (Form 56-1) once a year. The financial statements inside it are structured, standardised, and sold by every data vendor in the region. The prose is not. Management's discussion of what actually happened, the risk factors they chose to name, the matters the auditor decided were key — that material arrives as a several-hundred-page PDF, in Thai, once a year, from roughly nine hundred issuers at the same time.

This produces two different readers with the same problem. A foreign fund cannot read the document at all. A domestic desk can read any one of them and cannot read all of them, which is the same constraint expressed as arithmetic: the analyst covering thirty names has perhaps a day per name per year to spend on the part of the filing that is not a number.

So the qualitative half of the disclosure regime — the half where a deteriorating business explains itself before the numbers turn — is functionally unread at market scale. Not unavailable: unread. That distinction matters commercially, because it means the product is not selling access to documents that are already public and free. It is selling the ability to ask one question across every one of them at once, and to be handed the sentence that answers it, in the company's own words, with the page it sits on.

A result in this system is not a passage. It is a passage plus the coordinates that prove where it came from: the filing, the page, and the verbatim sentence as the company wrote it. That bundle is what the term bounding box refers to here — the boundary drawn around a claim so that the claim and its evidence cannot be separated downstream. A retrieval layer that returns text without it is returning an assertion; the same layer with it returns a citation. The index also retains delisted issuers — 860 of the 916 are still listed with the regulator, and the remainder are precisely the companies a failure study needs.[5]

The bundle is enforced by a gate that runs over the index and reports two numbers, not one: how many passages resolved to their page, and how many did not. In the Core Coverage window that is 98.83%, decomposed by axis rather than blended — auditor 99.48%, financials 98.77%, risk 98.85%.[3] A blended figure that cannot decompose is not auditable, and a pass rate published without its quarantine count is a pass rate with the failures deleted.[10]

There is a second measurement over the whole index, taken a day later, of 99.14%.[4] Both are correct. They count different populations on different dates and this document does not average them, reconcile them, or quote whichever is higher — a discipline that is easier to write about than to hold, and the single most common way a metric on a vendor's website becomes untrue without anyone lying.

For a reader whose desk does not read Thai, each cited line arrives twice: the company's Thai sentence verbatim, which is the authoritative text, and an English rendering beside it, labelled as ours rather than as the regulator's. The page number and the link to the document sit on the same row. That construction is deliberate and its limit is deliberate too — the regulator publishes English versions for some issuers and not others, so the product does not promise English filings it does not control.

The governing assumption is that in this product a wrong answer costs more than a missing one, and the two are not symmetric in any way that a benchmark score captures. A miss costs an analyst a search. A confident false positive costs them a paragraph in an investment committee memo, and it is discovered by the person they were trying to persuade.

That assumption is expressed in four mechanisms rather than in a policy document. Rows that fail verification against their source are withheld and never rendered, so the reader's failure mode is an empty result rather than an unsupported one.[10] Pipeline verdicts combine worst-of and never average, so a category that is 10% complete cannot be carried to green by a sibling that is 90% — the average of a truth and a falsehood is a falsehood with a smaller error bar. Nothing defaults to green: a stage whose status cannot be read fails closed. And scope floors are pinned as literals in code rather than read from configuration, so extending the corpus backwards cannot move a headline coverage figure that was measured on a narrower window.[9]

The same instinct governs the vocabulary. The axis a marketing department would call forensic is labelled operating quality everywhere a customer reads it, because in finance forensic denotes an investigation into fraud, and what the axis actually holds is six lenses over what management wrote about itself. The English term survives in the machine-read surfaces a compliance buyer searches, and nowhere a customer would mistake it for an accusation.

Three classes of leak are engineered against, and it is worth naming them precisely because only one of them is what the phrase normally means.

The first is cross-market leakage. The engine serves several storefronts from one codebase, so shared navigation, canonical tags and the command-palette search index are all capable of handing a reader of one market the content of another. This is checked by driving a real browser over every page of each market and asserting that no link, no canonical URL and no meta tag crosses the boundary — a source-code grep cannot do it, because the navigation is assembled at render time and a grep would report the defaults as the answer. The failure it exists for is not hypothetical; it shipped once, and was found by a person opening the page.

The second is fabrication leakage: model output that reads like the document but is not in it. The citation gate is the control, and its design decision is that failures are quarantined and counted in public rather than dropped silently.[10] A gate whose rejections are invisible cannot be audited by anyone outside the company, which makes it a claim rather than a control.

The third is staleness leakage — a figure that was true in one month presented as current in another. The build refuses a live figure that jumps beyond a set tolerance and requires a person to record why before the new number may be published, and figures that are deliberately frozen to a measurement date are labelled as frozen rather than refreshed to look current. This is the least glamorous of the three and has caused the most real-world error.

What this edition does not claim is set out in the unanswered questions above: a formal tenancy model, retention schedule and sub-processor list are not published here, and an institutional security review should expect to ask for them separately.

Two prices appear on one page, and their order is the mechanism rather than a layout choice. A bespoke forensic report is priced at $1,000 as a floor, sold per unit, and rendered above the subscription ladder.[7] A reader who meets that number first reads a monthly subscription underneath it as small; a reader who meets the ladder first reads $1,000 as expensive and stops. Anchoring only works in one direction, and the page says out loud why both numbers are on it — an unexplained expensive option beside a cheap one reads as a bait price to exactly the audience being addressed.

Subscriptions are collected in Thai baht. The dollar figures on the site are a reference line derived from a single fixed constant, rounded up, and never from a live exchange rate: a cached marketing page and a checkout page quoting different rates would put the difference on the customer.[8]

Access is metered in credits, and the design test applied to every proposed feature is whether it returns value for the credit it consumes. A feature that burns credits without returning anything produces churn dressed as usage, and a usage metric that rises while satisfaction falls is the most expensive kind of dashboard. Upsell is intended to arrive as a consequence of a customer hitting a ceiling they wanted to hit, not as a prompt.

The bespoke report is also the funnel's top: it is the same memo discipline as this document, performed on a name the client chooses, which makes free research the demonstration rather than the advertisement of the paid kind.

The company was founded by Varanchai Yingkhamnueng, previously a Model Risk Manager and Senior Data Scientist inside Thai banking, where the work included an NLP early-warning system for credit risk. He holds an IC Complex Type-1 licence from the Thai SEC.[11]

The relevance is not biographical. Model risk management is the specific discipline of asking what a model does when it is wrong, who finds out, and how fast — and every design decision in §5 is that question applied to a search product rather than to a credit scorecard. Withholding unverifiable rows, combining verdicts worst-of, failing closed on an unreadable status, pinning scope floors so a backfill cannot flatter a headline: these are validation-function instincts, not retrieval-engineering ones, and they are the reason this product is unusually willing to return nothing.

The same fact is a risk, and it appears in §10 in that form rather than being left as an implication here. A methodology that lives substantially in one person's judgement is a methodology with a single point of failure, and an institutional buyer is entitled to price that.

The expansion model is a franchise: one engine, one login, one billing account and one set of legal documents, with everything else addressed to a market. Adding a country is a registry row and a tenant configuration file rather than a second product, and the storefront, the coverage dashboard and the nav derive themselves from that row.

Thailand is serving. Vietnam is at corpus-acquisition stage with coverage of zero, and this is the part worth weighing: the Vietnamese storefront publishes that fact, charges nothing, and offers a waitlist instead of a subscription.[6] A greenfield market that were quietly being sold ahead of its index would be the fastest available way to convert the credibility described in §4 into a refund queue.

The condition for opening a market is therefore not commercial appetite but three technical facts: that the raw documents can be obtained at scale and lawfully, that they survive extraction into passages a gate can verify, and that the gate passes at the same bar the incumbent market is held to. A market opened below that bar does not simply underperform — it imports a lower standard into the one asset the brand has, which is the standard itself.

Each entry names a mechanism rather than a category, and states what would show the concern to be answered. A risk with no falsifier is a worry, and a register of worries is decoration.

1. 01Key-person concentration

   Mechanism

   The extraction methodology, the verification thresholds and the editorial standard all live substantially in one person's judgement. An absence does not degrade the service gradually; it stops the part of it that decides what is allowed to render.

   What would answer it

   A second person with commit rights to the extraction pipeline and a written, exercised handover of the verification standard.

2. 02Single-market revenue concentration

   Mechanism

   All revenue depends on one market's disclosure regime, one regulator's publication behaviour and one country's institutional buying cycle. The second market has coverage of zero and therefore contributes nothing to diversification today.[6]

   What would answer it

   A first paying customer in a second market, on the same product, at the same verification bar.

3. 03Dependence on a single document source

   Mechanism

   The corpus is acquired from one regulator's disclosure site under deliberately conservative crawl behaviour. A change to that site's structure, its access controls or its publication format stalls ingestion, and freshness — not volume — is what a subscription is renewed for.

   What would answer it

   A documented second acquisition path, and a published freshness figure that a customer can watch rather than be told about.

4. 04Translation liability on cited lines

   Mechanism

   The English rendering beside each Thai citation is ours, not the regulator's. A mistranslation on a cited line is the single failure most damaging to a product sold on traceability, because it survives every control that checks whether the passage exists — the passage does exist, and the reading of it is wrong.

   What would answer it

   A sampled bilingual review of cited translations with a published error rate, repeated on a schedule.

5. 05Metric legibility

   Mechanism

   The platform correctly publishes several coverage and verification figures at different scopes and dates.[3][4][9] A reader who quotes one without its scope misrepresents the company in their own committee memo, and the error will be attributed to us rather than to the quotation.

   What would answer it

   Every published figure carrying its scope inline at the point of quotation — the construction this page is itself an attempt at.

6. 06Institutional procurement gap

   Mechanism

   The controls described in §6 are engineering controls with automated enforcement. They are not the artefacts an institutional vendor review asks for — a tenancy model, a retention schedule, a sub-processor list, an answered security questionnaire — and a buyer who cannot obtain those may be unable to approve the purchase however good the product is.

   What would answer it

   A completed standard vendor security questionnaire available on request.

Every figure used above, once, with the population it counts, the date it was measured, its source, and how a third party re-checks it.