Decisions Are Fields,
Not Forecasts.

Most organisations do not suffer from a complete absence of information. They suffer from a failure of connection.

Customer feedback sits in one report. Revenue sits in another system. Competitor knowledge remains anecdotal. Operating constraints are understood by staff but rarely represented in the model. Strategic alternatives appear late, usually after the analysis has already been shaped around a preferred answer. The final recommendation may be polished, yet the route from evidence to action remains difficult to inspect.

Marres was built around a different premise: a strategic decision is not a forecast waiting to be discovered. It is a field of evidence, actors, constraints, uncertainties and possible interventions. A defensible recommendation emerges only when those elements are connected without concealing where observation ends and judgement begins.

This is the canonical Marres method:

Frame the decision. Understand the field. Model its relationships. Simulate what could change. Respond within the evidence. Learn from what follows.

The method is technically pluralist. It may use qualitative evidence, financial relationships, state-space models, structural-break detection, sensitivity analysis, competitive simulation and decision criteria drawn from economics or philosophy. But technical sophistication is not the objective by itself. The objective is disciplined judgement: making a consequential choice while remaining honest about uncertainty, causality and moral consequence.

From a retention model to a decision system

Marres did not begin as a general-purpose strategic platform. Its conceptual origins can be traced to Viktoriya Doncheva's 2025 applied research thesis on improving customer relations at Pahat Pahit Teahouse. The study began with a bounded management question: how could the teahouse improve retention among its target customers?

The research combined reservation and point-of-sale records, surveys, in-store observations, social-media classification, customer reviews, competitor visits and relevant literature. This triangulation produced a view of the organisation that was simultaneously commercial, behavioural and operational. Customer experience was not treated as an isolated branding variable; it was connected to service consistency, retention, product dependence, price perception and the rituals through which customers experienced the venue.

An early analytical workflow documented in the thesis translated this evidence into four customer-facing domains: atmosphere, food and drinks, service, and value for money. The prototype compared Pahat Pahit with peers, examined variation and cross-domain correlations, estimated domain elasticities, simulated a change in each score, and translated the result into indicative revenue opportunity and risk. It then selected the domain with the greatest apparent intervention potential.

This established the enduring proposition behind Marres:

Evidence should not terminate in description. It should be connected to a decision.

But the prototype also revealed the limits of intervention ranking. Its target organisation and four domains were fixed. It could fall back to an assumed revenue value when direct data was absent. Elasticities were normalised within predetermined bounds, small correlations were discarded, simulated impact was capped, and the workflow searched for a single highest-upside move. These were pragmatic choices for an applied research prototype, but they compressed uncertainty and organisational context into one recommendation.

The system could ask:

Given these customer-facing levers, which intervention appears to offer the greatest opportunity?

It could not yet fully ask:

Which action remains defensible when evidence is incomplete, operating capacity is constrained, alternatives compete, conditions change and consequences unfold over time?

The next methodological step was the Marres–Ramjet Fusion Model, formalised by Jonathan Edbert Muliadi in 2026. This changed the object being modelled. The organisation was no longer represented only as a collection of domain scores. It became a dynamic allocation system in which outward activation interacted with structural readiness, conversion capacity, referral propagation, market visibility and accumulated goodwill.

This introduced a crucial constraint: attention is not equivalent to realised growth. An organisation can generate demand faster than it can absorb, serve or convert it. Under weak readiness, additional activation may create operational congestion, inconsistent service and diminishing marginal return. Strategic allocation therefore had to consider both the production of attention and the infrastructure required to receive it.

The contemporary Marres architecture extends the method again—from a model that recommends an intervention into a persistent investigation that governs the path from question to learning. In the current system:

• a research question frames the investigation;
• evidence is preserved as a traceable artefact rather than transient workflow input;
• alternatives are explicit objects rather than implied recommendations;
• each analytical capability receives only the context it requires;
• model preparation, simulation, robustness and decision have separate readiness states;
• success and failure produce semantic history events;
• readiness gates prevent unsupported findings from appearing in the final Brief; and
• the decision remains connected to its assumptions, evidence and subsequent learning.

The progression can therefore be summarised as follows:

This was not simply an increase in computational complexity. It was a change in the philosophy of the system. The early prototype asked a model to recommend a move. The current approach asks an organisation to construct, contest, test and remember a decision.

1. Begin with the decision, not the available data

Conventional analysis often begins by asking, "What data do we have?" Marres begins one level earlier:

• What decision must be made?
• Which alternatives are genuinely available?
• Who is affected by the decision?
• What constraints cannot be wished away?
• Which uncertainties could reverse the preferred answer?
• What evidence would count against our initial intuition?

This framing is not administrative preparation. It establishes the object of analysis.

A revenue dataset, for example, does not contain a decision. It can describe transactions, trends and discontinuities, but it cannot determine whether an organisation should expand, retrench, reposition or wait. That requires an explicit relationship between evidence and available action.

At HeySara, commercially sensitive information was deliberately bounded, including a redacted revenue figure. This prevented false numerical precision, but it did not eliminate the strategic question. HeySara's ambition was to develop from a set of incorporation and corporate services into a more integrated relationship-led offering. Marres's analysis approached the organisation from another direction and reached a complementary conclusion: the company's network of clients, companies, directors and services could itself become commercial infrastructure.

The resulting proposition was not simply to produce more retrospective reporting. It was to test whether relationship mapping could help answer a more operational question: where should commercial attention be directed next?

This is an important principle. Incomplete data should reduce the strength and specificity of a claim. It should not automatically prevent useful reasoning.

2. Understand the decision field

Once the decision is framed, Marres constructs an evidence model around it. The purpose is not to accumulate everything available. It is to identify what bears on the decision and what kind of evidence each item represents.

The field may include:

• customer language and market signals;
• transaction and operating data;
• price, cost and capacity constraints;
• competitor identities and observed behaviour;
• organisational capabilities;
• stakeholder incentives;
• prior interventions;
• explicit assumptions; and
• material absences in the evidence.

These inputs are not treated as interchangeable. A verified transaction, an analyst assessment, a customer review and a simulated competitor response have different epistemic status. Marres therefore separates four categories:

Observed — recorded directly in supplied or collected evidence.

Modelled — produced by a formal representation of relationships and assumptions.

Inferred — a reasoned interpretation consistent with the evidence but not directly observed.

Proposed — an action or mechanism that has not yet been implemented and validated.

This separation creates provenance. A reader should be able to trace a recommendation backwards and see which parts rest on measurement, which on assumptions, and which on judgement.

Pahat Pahit: bounding attribution to the intervention

Pahat Pahit illustrates why this matters. Service improvements were introduced after 11 August 2025. They acted principally in the middle and at the end of the commercial funnel: the quality of service, conversion of an existing visit and the conditions for a return visit. No corresponding top-of-funnel marketing programme was conducted.

It would therefore be methodologically incoherent to attribute total subsequent revenue movement to the service intervention. Revenue depends on the whole commercial system. Better service may improve conversion, average spend, retention or resilience, but it cannot independently replenish awareness and acquisition.

The available transaction file begins on 1 August, leaving only a short pre-intervention window. From August through December, recorded monthly net sales moved from approximately Rp67.3 million to Rp44.5 million. Aggregate revenue did not provide evidence of a sustained uplift.

Yet the headline concealed an important compositional change. Average net sales per positive transaction rose from approximately Rp166,000 in August to Rp176,000 in September and Rp216,000 in October. It subsequently remained around Rp190,000 in November and December—still above the August level—even as the number of transactions declined. Customers who completed a purchase were therefore spending more per ticket, while fewer tickets were being generated overall.

This pattern is compatible with, but does not prove, an improvement within the middle or end of the funnel. A service intervention could plausibly affect conversion quality, basket formation or the experience of customers already present. It could not compensate indefinitely for weakening acquisition or lower visit volume. The evidence therefore supports a more granular diagnosis: transaction volume weakened while realised ticket value strengthened relative to the initial month.

That does not prove that service improvements caused the increase in ticket value. It means the intervention must be evaluated against outcomes it could plausibly influence, with the limits of the baseline stated openly.

In this case, the correct response to ambiguity is not to manufacture a success story. It is to narrow the claim.

3. Model relationships, not decorations

A model is useful when it makes a consequential relationship explicit. It is not useful merely because it produces a number.

Marres models are constructed around questions such as:

• What constrains the organisation's ability to convert demand?
• Which investments reinforce one another?
• Where might additional activation meet diminishing returns?
• Which assumptions materially change the recommended allocation?
• How quickly does strategic momentum decay without reinforcement?
• How might competitors or market conditions alter the result?

The model is not presented as causal truth. It is a testable representation of how the decision is believed to work.

Atlas Incorporated: from micro-test to a new operating regime

Atlas Incorporated began as a deliberately small market test. Early revenue was low and intermittent, followed by an extended period that was operationally inactive or otherwise non-comparable with the later business. The concept was then renovated and expanded.

This history matters because a forecast trained indiscriminately on the entire series would combine different data-generating processes. The early micro-format and the expanded operation did not have the same capacity, visibility or commercial conditions. Treating them as one stable process would create numerical continuity where the business itself had changed.

At week 37, weekly revenue rose to Rp1.493 million and began a persistent acceleration. A sequential regime-shift method, STARS, detected an upward structural break around that point under the selected parameters. "Ignition" became the operational description of the event; it was not a formal output of the statistical method.

Revenue then developed rapidly:

After a sharp negative observation at week 43, revenue recovered and accelerated again, eventually reaching Rp21.112 million at week 47.

Two analytical questions now had to be kept separate. First: had the business entered a genuinely different regime? Second: what was happening to the underlying level and momentum within that regime?

STARS addressed the first question by requiring accumulated evidence before declaring a structural shift. A local linear trend model, estimated through a Kalman filter, addressed the second by updating a hidden revenue level and hidden weekly growth as each observation arrived.

The distinction became useful after the peak. Revenue declined from Rp21.11 million at week 47 to Rp12.31 million at week 51. The Kalman estimates responded quickly, reducing hidden growth and indicating negative acceleration. STARS, however, had not yet accumulated sufficient evidence to confirm a second downward regime.

These were not contradictory findings. They operated at different evidentiary speeds:

• the state-space model asked what momentum appeared to be doing now;
• the regime test asked whether the change was sufficiently persistent to rename the underlying state.

At week 52, revenue rebounded 10.2% to Rp13.56 million. Comparable Monday-to-Saturday revenue increased 16.8%, while Sunday revenue declined slightly. The rebound was therefore not an artefact of an unusually strong Sunday.

A newly introduced baked butter rice product appeared in 19 orders. Its direct revenue was approximately Rp1.10 million, while baskets containing it represented approximately Rp2.32 million, or 17.1% of weekly revenue. This supported a medium-confidence hypothesis that a heavier-meal intervention assisted the rebound. It did not establish causality: the product may have generated incremental demand, substituted for other purchases or coincided with an independent recovery.

The purpose of the model was not to eliminate that ambiguity. It was to locate it precisely.

4. Test the assumptions that could reverse the answer

A single forecast conceals how dependent its recommendation may be on a fragile assumption. Marres therefore treats sensitivity as part of the decision, not as an appendix.

For Atlas Incorporated, a baseline allocation model suggested directing 67% of strategic resources towards activation and 33% towards system-building. Under a pessimistic readiness assumption, the model moved to 55% activation and 45% system-building. The mathematical result changed because the organisation's ability to absorb and convert demand had changed.

From working paper to applied decision model

The allocation was grounded in Jonathan Edbert Muliadi's Marres–Ramjet Fusion Model working paper, which formalised the relationship between outward activation and the organisation's capacity to absorb it. The paper divides a finite strategic budget B between structural investment R and activation investment J:

Structural investment expands operating readiness, resilience and conversion capacity. Activation investment generates immediate visibility, traffic or demand. The critical move is that the model does not assume every unit of activation becomes useful growth. Realisable activation is bounded by an absorptive ceiling:

where A(R, J) represents the organisation's capacity to sustain and convert activation under its current conditions. If raw activation exceeds that capacity, the excess may appear as bottlenecks, inconsistent service, weaker conversion or declining marginal productivity.

For application, the working paper specifies a partially observable strategic state consisting of conversion capacity, visibility gap, referral propagation and market power. Atlas Incorporated's state was not treated as directly measurable truth. It was assigned through the paper's evidence hierarchy:

1. direct operating or market observations;
2. proxy-derived estimates; and
3. structured qualitative assessment with declared uncertainty.

Customer response, visibility, operating independence, service consistency and organic propagation were therefore translated into inspectable state assumptions. The baseline and pessimistic scenarios altered those assumptions—especially conversion readiness—and observed how the recommended allocation moved. This made the scenario exercise diagnostically useful: it identified readiness as a variable capable of reversing how aggressively the business should activate demand.

The working paper's validation supports the internal coherence, parameter recoverability and noise behaviour of the canonical system under synthetic experiments. It does not establish that its latent variables are uniquely measurable or that its Atlas Incorporated outputs are externally validated causal estimates. Its proper use here was as an interpretable simulation framework: a way to expose allocation logic, test sensitivity and keep analyst judgement auditable under partial observability.

The practical conclusion was more important than either percentage:

Marketing pressure and operating capacity cannot be optimised independently.

If demand is activated faster than the organisation can serve it, additional visibility can degrade service, weaken conversion and destroy some of the momentum it was intended to create. The final recommendation therefore incorporated observed operating conditions and prioritised capability infrastructure before major activation.

This is the role of human judgement in Marres. Judgement does not arrive after the model to decorate its answer. It adjudicates between modelled futures, operational evidence, values and constraints. The intervention must remain visible, including where an analyst chooses not to follow the numerical optimum literally.

5. Use sparse signals to generate hypotheses, not certainties

Not all strategically important evidence arrives as a large dataset. Some begins as an anomaly: a conversation, an incentive that appears misaligned, an unexpected absence or a behaviour that does not fit the organisation's stated advantage.

In exploratory work involving ERA, Marres encountered an agent seeking clients for a prestigious freehold development near Tanjong Pagar. There was little accessible evidence of her own transaction history, and commission competitiveness formed part of the attempt to begin a sales conversation.

One encounter cannot establish the culture of a large organisation. It can, however, reveal a question worth testing.

ERA possessed collective scale, brand recognition and extensive transactional experience. Yet an individual agent could still face the market as an isolated micro-enterprise, required to establish credibility again and tempted to differentiate through price rather than visible evidence of expertise.

This produced an institutional hypothesis: a large agency network may possess substantial latent reputational capital without consistently transmitting that capital to the point where an individual client chooses an individual agent.

The next questions follow naturally:

• Can verified experience become more accessible without erasing individual attribution?
• Can reviews and reputational evidence be consolidated into a navigable system?
• Can referrals connect a client to the agent most relevant to a particular property or need?
• Can agents compete through demonstrated expertise, service and fit rather than relying primarily on commission concessions?
• Can individual activity accumulate into organisational learning?

Here the method resembles colligation: separate observations are brought together under a possible explanatory structure. But the structure remains a hypothesis until broader evidence supports it.

6. Simulate plural futures without pretending to foresee them

Forecasting usually asks which future is most likely. Decision-making under deep uncertainty asks a different question: which action remains defensible across futures that cannot be assigned reliable probabilities?

Marres uses simulation to vary the conditions that matter: demand, costs, conversion readiness, competitor response, implementation quality, time delays and market state. The aim is not to produce theatrical precision about what will happen. It is to expose which decisions depend on optimism and which remain comparatively robust when the world changes.

This leads naturally to regret-based reasoning. For an action a under future state s, regret can be expressed as:

where a*‑s is the best action had the future state been known in advance, and U is the value produced. A minimax-regret decision seeks the action whose worst missed opportunity is smallest:

This criterion is useful when expected-value calculations would require probabilities we do not honestly possess.

There is also a philosophical dimension. Maximin reasoning asks how alternatives perform for the worst-positioned outcome or stakeholder. Rawls used a related idea in considering the justice of social institutions, although a commercial decision model should not casually claim the moral authority of Rawls's theory. The transferable lesson is narrower: an institution should not be judged only by its aggregate upside if that upside depends on making its weakest position intolerable.

Marres can therefore evaluate more than revenue. Depending on the decision, the relevant objective may include organisational resilience, service quality, employee capacity, customer harm, reversibility or the distribution of downside. Optimisation always embeds a value judgement about what deserves to count.

7. Respond with a bounded recommendation

A useful recommendation states more than what to do. It states why, under which conditions, and what would cause the decision to be reconsidered.

A canonical Marres response contains:

1. the recommended action;
2. the evidence supporting it;
3. the alternatives considered;
4. the assumptions on which it depends;
5. the most important uncertainty;
6. the indicators to monitor;
7. the conditions that would trigger adaptation; and
8. any domain boundary requiring specialist review.

This last element is essential. Models can represent regulated or ethically consequential decisions without possessing the authority to resolve their legal or moral status. Recommendations affecting pricing coordination, credit, employment, health, privacy or regulated advice require the relevant domain review before implementation.

Technical capability does not dissolve responsibility. It increases the obligation to state where capability ends.

8. Learn from the distance between model and world

The decision is not the end of the analysis. It creates the next observation.

After an action is taken, Marres records what changed and compares it with the model's expectation. Divergence may indicate several different failures:

• the original evidence was incomplete;
• an assumption was wrong;
• implementation differed from the proposed action;
• an external event changed the field;
• a competitor behaved unexpectedly; or
• the model represented the relationship poorly.

These explanations should not be collapsed into a generic statement that "the strategy failed." Each implies a different correction.

The Atlas Incorporated sequence demonstrates this recursive character. The first structural break made the old baseline obsolete. Rapid growth then made continued linear extrapolation unsafe. Contraction weakened the earlier equilibrium estimate. A one-week rebound interrupted the decline without yet proving recovery. Each observation changed the state of knowledge, not merely the value in a dashboard.

Pahat Pahit shows the same discipline from another direction. Because the intervention addressed only part of the funnel, a change in total revenue cannot be interpreted without acquisition conditions. Learning begins by respecting the intervention's causal reach.

The system therefore becomes an organisational memory of decisions: what was known, what was assumed, what was chosen, what happened and what the organisation learned.

The moral architecture of a decision system

Every model simplifies. Every recommendation distributes risk. Every claim of confidence can cause another person to act.

For that reason, the canonical Marres approach is not simply technical. It is moral in a practical sense. It demands restraint where evidence is weak, disclosure where assumptions matter, and attention to stakeholders who may carry the downside of an apparently efficient decision.

Its standard is not omniscience. No serious decision system can promise that.

Its standard is answerability:

• Can we explain how the recommendation was produced?
• Can we identify which evidence supports it?
• Can we show what remains uncertain?
• Can we distinguish observation from inference?
• Can we revise the recommendation when reality disagrees?
• Can we identify who bears the risk if we are wrong?

This is why Marres treats decisions as fields rather than forecasts. A forecast gives a view of what may happen. A decision field shows what is connected, what is contested, what could change and where responsibility resides.

The final output may be a recommendation, a strategic allocation, a simulation or a briefing. But the product underneath is more fundamental: a traceable path from evidence to action, built to remain useful even when certainty is unavailable.

Methodological note: Figures in this article are drawn from internal Marres case materials. Modelled and inferred findings are described as such. Simulation outputs are directional decision-support analyses, not guaranteed forecasts or substitutes for regulated professional advice.