How MERIDIAN works The method behind every assessment, written out in full

1What MERIDIAN is

MERIDIAN is an intelligence product covering eleven Southeast Asian markets. It answers questions of the form will this hold, and how would anyone know if it did not.

What separates it from a research note is not a theory. It is a procedure. Every serious risk house publishes the same four things: what it looks at, how it scores what it finds, how much weight it gives each part, and where the data stops and the analyst's opinion starts. This note is those four things.

The claim in one sentence: five ways of looking, one published method, every call scored afterwards.

2The one rule that governs everything else

What counts as working. The sources by name, with the date each was published. What was done to them. A method that can be named. An entry in the record MERIDIAN keeps of every figure it prints — the Backend Computation Log — holding all of it. The test is simple: a reader must be able to arrive at the same number without asking MERIDIAN a single question.

What does not count. A list of references attached to an opinion is not a calculation. Turning a made-up number into a made-up range does not make it real. Calling a figure a judgement does not earn it the right to be a figure. An account of how the analyst was thinking is a record of thinking, not of arithmetic.

What happens to a claim that has no working. It stays in the report, without a number — as a sourced fact, a named limit, or a judgement in plain words. The analysis stays. The number goes.

The reason this comes first, rather than sitting in a quality checklist: MERIDIAN's whole offer is that its numbers can be checked. One invented figure destroys that offer across the entire product, because a reader who finds it has no way of telling which of the others are real. No amount of good analysis makes up for a single fabricated number.

A second rule sits beside it. A gap is never filled with the nearest number to hand. Using a state figure where a national one is needed, a city for a country, or a forecast where an actual was required, is the most common way a solid document becomes an indefensible one. Where the figure does not exist, the report says so.

3What is covered

Eleven markets, every time. Malaysia, Indonesia, the Philippines, Thailand, Vietnam, Singapore, Cambodia, Laos, Myanmar, Brunei and Timor-Leste. Where a report cannot cover one of them properly, it says so on the page rather than quietly leaving it out.

Thirty subject areas, checked every time. These run from macro-economics and capital flows through governance, trade, technology, energy and water, demography, and public narrative. All thirty are examined for every report, whatever the subject, and the report publishes the full list showing which mattered and which did not.

An indicator register of 1,076 entries, as it stood on 13 August 2026. Each entry names what is measured and the official source it comes from. The number is always published with the date of the register behind it, because a register that never grows is a register nobody is maintaining.

Careful language about the register. An indicator is either defined — it exists, with an official source named, and no reading has been taken; retrieved — a reading has been taken, with the date and the source; or assessed — a named analyst has judged it because no reading can be taken automatically. MERIDIAN does not describe an indicator as verified, because that word implies a reading that has not been taken.

Where a report claims coverage of all eleven markets, it publishes the limits of that coverage just as prominently.

4How evidence is handled

Every reference carries one of four labels, printed at the end of the reference where the reader can see it.

The rule this produces. Where no source can be reached, first-hand or reported, the claim does not appear as a claim. An honest gap beats an unsupported assertion every time, without exception.

Showing the reasoning, not just the label

A label says what kind of claim something is. It does not say how the claim was reached. Every judgement in a MERIDIAN report therefore carries a short entry in a reasoning log, kept alongside the Backend Computation Log — the facts it actually rests on, how those facts became this judgement, what came out, and how strong the reasoning is. A judgement with no entry does not appear.

The strength marking is the part that does the work. It has three settings: strong, moderate and weak — and it has to be willing to say weak about MERIDIAN's own reasoning, because a log that marks everything moderate is a formality. A single past example is a pattern of one, and the entry says so.

The alternative is to print every judgement as though it carried the same authority. A reader who then finds one weak claim has no way of telling whether the rest are any better, and discounts the whole document. The log allows a reader to discount one claim instead.

Marking a claim as weak is not permission to keep it. Where striking the claim out changes nothing, it does not appear at all.

Grading the sources before the writing starts — the Quality of Information Check

Before any drafting, every input is listed with its publisher, its date, and whether it is a measurement, a model output, an announcement or a claim. That table publishes with the report. For infrastructure subjects the difference between what has been announced and what has actually been built is decisive, and the two are never added together into one figure.

5How MERIDIAN says how sure it is

Two different things get confused constantly in this industry, and MERIDIAN separates them by rule.

A judgement can be highly likely and rest on thin evidence. It can be a close call resting on excellent evidence. The two are never put in the same sentence, and a confidence figure is never dressed up as a probability.

MERIDIAN publishes a likelihood only where the likelihood can be computed. A likelihood figure needs either a counted class of comparable past cases, or a rule fixed in advance that turns named signals into a figure. Where neither exists, MERIDIAN publishes no likelihood — not a percentage, not a band, and not a word standing in for one. The call is stated instead as a dated claim, with the reading at which it would not hold, and it is marked in public on its due date.

Confidence is reported separately, as low, moderate or high, with the reason it is not higher named in the same breath. Section 10 explains how that grade is arrived at.

6How a call is made checkable

An assessment nobody can disprove is an opinion. Every MERIDIAN report therefore states in advance what would show it to be wrong.

The assumptions are surfaced first — the Key Assumptions Check

Before the analysis is written, the assumptions the conclusion depends on are listed — the ones that, if wrong, change the answer. For each, the report states what follows if it turns out to be wrong. That list publishes with the report as the Key Assumptions Check. It is not a summary of the argument; it is everything the argument takes for granted.

The failure is imagined before it happens

Before publication the author writes the sentence it is the resolution date, this call was wrong, why? and answers it with the three most likely ways it could have failed. Any failure route not already covered becomes a warning condition. This takes fifteen minutes and it is the cheapest quality step in the whole method.

At least three warning conditions, each written so it cannot be argued about

Every report carries a minimum of three conditions that would signal it is going wrong. Each one has to name five things.

A condition that cannot be written this way does not count. Sentiment deteriorates fails. A general election is called before the second quarter of 2027, seen in the Election Commission announcement passes.

Scenarios, and why they carry no probabilities

Every report sets out at least three scenarios, and they must be genuinely different — watched by different signals, not described with different adjectives. Two scenarios watched by the same signals are one scenario with two names, and they are merged.

Scenarios carry no probabilities. MERIDIAN states no figure it cannot compute from named inputs. A probability weight needs either a reference class with a stated number of cases, or a conditioning rule set in advance. Where neither exists, no weight is published — not a point estimate, not a band, not a hedged one. The scenarios are told apart by triggers that share no indicator, and every trigger can be seen in a named source on a stated date. A reader watches the trigger rather than a number.

Where a future report can meet that test, the probability publishes with its working, in steps of five percentage points, and is never reused in another report.

When a view changes

Where a judgement moves between issues, the report states what moved, by how much, and what evidence moved it. Where a judgement does not move despite new evidence, that is stated too. This is how a new entrant can be credible without a twenty-year history: a reader does not have to take MERIDIAN's word for its own consistency, because the changes are printed.

7How MERIDIAN decides what actually matters

Thirty subject areas are examined. Three are put forward as the ones that decide the answer. What happens in between is where most research quietly stops explaining itself.

An analyst who picks three subjects owns that choice. It carries a name, and if it is wrong, that person is wrong. A product built with AI assistance cannot borrow that ownership — nobody's reputation stands behind a choice nobody recorded. The choice has to hold up as a procedure instead: the same evidence must produce the same three subjects whoever runs it, on whatever day. Without that, thirty areas were examined is decoration on a decision made some other way.

First, what kind of force is it

Before anything is ranked, each subject area that matters is assigned one of three roles, decided by asking what would happen if it disappeared.

Anything fitting none of the three is background and cannot be a headline signal. Anything appearing to fit two is assigned to whichever role its absence would affect more, and the contest is recorded.

The distinction between a limit and an opposition is about origin, not hostility. An opponent who was already there and would be there anyway is a limit. A body that organised because the thing arrived is an opposition, and it behaves differently — it grows as the thing it opposes grows.

Then, rank within each role — and this is the part that matters

Each subject area is scored on three things, all of them countable rather than felt: how much sourced evidence was found for it; how much the conclusion depends on it, tested by striking its evidence out and seeing whether the headline still stands; and whether it can be checked inside the report's time horizon. The arithmetic is in the appendix.

One feature of that arithmetic is worth stating in words, because it is a deliberate design choice: a subject with a mountain of evidence that changes nothing when removed scores zero. Volume of material is not importance.

The ranking happens separately inside each of the three roles, never across all of them at once. This is the whole point of assigning roles first.

Ranking everything together can easily return three subjects that all describe the same force. Three drivers is not an assessment — it is an enthusiasm with citations. Three limits is a doom story. Three oppositions is a campaign document. Each one is internally consistent, each survives a check of its sources, and each is wrong in the same way: it describes one face of a situation and presents it as the whole of it. Taking the top of each role instead gives a complete account — what makes it happen, what holds it back, what pushes against it — and completeness is not something better sourcing can substitute for.

The whole table publishes, not just the winners

The published table is called the signal derivation table, and it is the record of how thirty subject areas became three.

Every subject area that mattered appears in the published table, with its role, its three scores and its total. Ties are broken by a stated order, so the procedure never returns an ambiguous answer.

Publishing all of it is the point. A reader who disagrees is not reduced to saying I do not accept these three. That reader can see that the runner-up came within one point of the leader and argue about that one line, and can see everything MERIDIAN classified as background and dispute the classification rather than the conclusion.

Where a person overrules the procedure, that is allowed and it is published — what the procedure returned, what was substituted, why, and by whom. That is the point at which a human takes the ownership back.

How many are published is a decision about product, not about analysis. Three — one per role — is the shortest complete account and is the minimum. Publishing more means going further down the list inside each role: a second and third limit, never three more drivers, so the account stays balanced as it goes deeper.

8Where the warning conditions come from

A warning condition is not something an analyst thinks up. It is the visible form of a specific way the assessment could be wrong, and there are only two.

1. One of the three signals stops working. Each published signal produces one condition: what would show it has stopped operating? A driver fails by weakening at its source. A limit fails by biting harder — or by being lifted. An opposition fails by winning, or by never turning into anything with teeth.
2. A judgement underneath the numbers is wrong. Every place where the report relied on a judgement rather than a measurement produces one candidate: what would show that judgement was mis-set?

The number of conditions therefore follows from the structure of the report rather than being chosen. Three signals plus one load-bearing judgement gives four candidates — not because four is a good number, but because that is how many ways the assessment can fail.

Candidates then face two tests before they are ranked at all: all five fields must be writable, and the event must be capable of happening inside the report's time horizon. The survivors are sorted on two questions. Does it tell the scenarios apart? A condition that would occur in the good case, the bad case and the base case alike carries no information, however dramatic it sounds. How early does it fire? A condition that resolves on the same day as the outcome is a scorekeeper, not a warning.

9The five ways of looking

Five analytical lenses. Each one answers a different buyer's question and each one has to produce a specific piece of work before a report is allowed to say it was used.

10The confidence grade

Every report carries a grade of low, moderate or high for the strength of the evidence behind it. It is built from six things: how well the important claims are sourced; how many printed figures carry a computation or a dated source behind them; how much rests on figures read from a named source on a named date; how many of the five lenses returned a reading, and how strong each was; how tightly the warning conditions are written; and how complete the published working is — the domain scan, the derivation table, the assumptions check, the premortem, the source grading and the rejection log. The arithmetic is in the appendix, published so it can be re-run and argued with.

What the grade is not. It is not the probability that the report is right. A well-sourced report can reach the wrong conclusion and a thin one can be right. The grade is a floor under the evidence, and it appears as a disclosure rather than as the headline number.

That is what makes publishing the method worth anything. It is not a self-audit for strangers to repeat; it is a shared instrument for grading rival views. A reader whose three subjects rest on thinner sourcing scores lower and can see precisely which part did it. A reader who scores higher than MERIDIAN has produced something MERIDIAN ought to publish.

Who checks the checker

A grade given by one party with nobody checking is an opinion. Where two independent people grade a report and differ materially, they reconcile and the reconciliation is recorded.

Independence is what makes this work, and it is easy to fake. Running the same AI system twice does not produce two opinions. It produces one opinion twice, because both runs share the same training and the same blind spots. Their agreement measures consistency, not accuracy — it will confirm the same mistake with high agreement and no warning. A genuine second opinion means a different person, or a different AI system built by a different company, or the same system explicitly instructed to attack the grade and find the claim that fails.

MERIDIAN currently has no independent reviewer. Every grade therefore publishes marked single rater, unreconciled, and that marking appears on the face of the report. Appointing an independent reviewer is the only thing that would lift it.

11Scoring, and what is being promised

Every research house in the world has been right and wrong in roughly equal measure. Almost none of them can show which. MERIDIAN keeps a public record of both.

What a prediction log is

It is a public register of every call MERIDIAN makes. Each entry is written at the moment a report publishes and it fixes four things: the call itself, in words precise enough to settle an argument; the date it comes due; what would count as right and what would count as wrong, named in advance; and where the outcome will be looked up.

Nothing in an entry may be edited after publication. On the due date the outcome is written beside it, whichever way it went, and the running accuracy record is updated in public.

That is the whole mechanism, and everything MERIDIAN claims rests on it. Any research house can say it was right. A house with a prediction log can be checked — including by somebody who wants to prove it wrong.

This is worth stating plainly because it cuts both ways. It means MERIDIAN has no track record to show, and a reader is entitled to weigh that. It also means the record, when it exists, will be the only one of its kind in this region — because a track record cannot be assembled after the fact. It can only be accumulated, one dated call at a time, in public, including the ones that go wrong. Anyone can copy the format in an afternoon. Nobody can copy three years of resolved calls.

A house that claimed an accuracy record on the day it launched would be claiming something it could not possibly show. Scoring is therefore described here as a promise — a commitment with dates attached to it — and it becomes a record on the first resolution date, and not one day before.

What a MERIDIAN report is

Every MERIDIAN report is an Assessment. An Assessment can be defended, but is not mechanical: a different competent analyst working from the same evidence could reach a different view, and the disagreement would be arguable on the record rather than settled by re-running anything. Every calculated figure inside it, however, can be rebuilt from the inputs and the operation printed in its appendix.

12Who is behind it

The method is original work by the founder of aseanx — the thirty subject areas and how they are organised, the insistence on all eleven markets including the three that the standard reference datasets leave out, the view that outcomes come from several forces at once rather than any single one, the accountability model with conditions stated in advance and marks published afterwards, and the decision to publish the method rather than ask to be trusted.

AI is the instrument, not the author. Research, drafting and first-pass grading are AI-assisted, carried out under the method set out here and reviewed before issue. The relationship is the ordinary one between a practitioner and a tool.

There is no outside reviewer and no individual signature. Accountability sits with the publisher. aseanx issues the report, aseanx is named when it is marked on the resolution date, and aseanx carries the result either way. This is the normal arrangement at a research house, and it is stated here so that the absence of a signature is a disclosed structure rather than something a reader has to notice.

On marking its own work. The confidence grade is produced by the same process that produced the report. There is exactly one thing that makes that defensible: the grading method is published, so any reader can re-grade the report and produce a number to argue with. A published method turns a self-assessment into a checkable claim; an unpublished one is worth nothing whoever produces it. The implication runs the other way as well. If MERIDIAN ever stops publishing the method, the grade goes — not softened, gone — because at that point it is an unverifiable number produced by an interested party.

13What this Assessment claims, and what it does not

Precision about scope is not a weakness. A reader who knows exactly which claims are being made can rely on them; a reader left to guess discounts all of them.

Every line above tells a reader which claims to rely on and which to read as judgement. Each is dated when its status changes.

14How to argue with MERIDIAN

This method is published so that it can be used against MERIDIAN's own work. There are four routes, in ascending order of force.

1. Challenge a source. Every important claim names an official source with a date. A source that does not support the exact claim made is a defect, and it is corrected at the next issue with the change stated.
2. Challenge a line in the ranking table. The full table publishes, with every subject area, its role and its scores. A disputed score is a specific, settleable argument rather than a general objection.
3. Challenge the choice itself. Name three different subject areas, one for each role. The same grading is applied to that choice and the two numbers are directly comparable. A choice that grades higher than MERIDIAN's is something MERIDIAN ought to publish.
4. Challenge a warning condition, or supply one. The rejected candidates publish alongside the accepted ones, each with the reason it was rejected. A reader who can show that a rejected one is in fact public and measurable has supplied a condition MERIDIAN will adopt.

A challenge MERIDIAN accepts is recorded, the affected judgement is issued again with the change stated, and the reader who supplied it is credited where they permit it.

AAppendix — the working

Everything in this appendix is arithmetic. It is published because a grade nobody can re-run is worth nothing, and because a reader grading a competing view needs the same instrument. Nothing here is required in order to read the report.

A.1 · Ranking the subject areas

For each subject area that shaped the report, three values are recorded.

The first term multiplies so that heavy evidence which changes nothing scores zero. The second adds, because being checkable is a threshold rather than a magnitude: a subject that cannot be checked in the horizon is not disqualified, and it can never lead. Ties break on dependence, then evidence, then register order.

A.2 · The confidence grade

Six inputs, each scored 0 to 4 against published descriptions of the state a document is in. The figure beside each is the maximum points it can contribute to 100; points earned = ( score ÷ 4 ) × maximum.

The six scores, their maximums, their contributions and a one-line reason for each publish on the face of every report. A grade is only checkable where a reader can see which part produced it — otherwise a reader can disagree with the total and with nothing else.

Where the points came from, stated so they can be disputed. Source quality and computational backing carry twenty each because they are what a hostile reviewer checks first, and what the standing rule of section 2 stands or falls on. Lens application carries twenty because five readings are what MERIDIAN claims to produce. Published working carries fifteen because that apparatus is what separates this from a research note. Evidence depth carries fifteen. The conditions carry ten — least, because in practice they are close to pass-or-fail and are already a hard requirement before publication, so weighting them higher would count the same thing twice. These are a judgement, not a finding, and they are open to challenge.

A.3 · The checks a report must pass before it publishes

Enforcement is mechanical rather than editorial. A report publishes only where all of the following are true.