Myceliation Stability at Scale: Regime-Aware Learning

Scaling Before Stability (Scaling as Experiment)

There is a distinct and under-discussed moment in bioprocess development where a system has been scaled far enough to carry real weight (real cost, real opportunity, real consequence) but not long enough to be demonstrably understood or stabilized at scale. You have something like an MVP: a process that can produce the right kind of outcome, sometimes, under conditions that are not yet fully bounded. You are running successive operations not only to manufacture, but to learn what the process even is at scale. Each run carries stakes. Each outcome becomes a verdict.

This is the paradox of scaling before stability: the system must operate in order to reveal what it needs to become stable, yet operating under instability attracts pressures that resist the kind of learning stability requires. Technical pressure mounts because scale reveals constraints that do not exist at bench. Economic pressure mounts because proof-of-concept must show up as repeatable yields, not anecdotes, and capital burn becomes visible in ways it never was during benchtop exploration. Psychological pressure mounts because once operations begin, the system starts issuing verdicts run by run, and it becomes difficult not to interpret each verdict as a diagnosis, each deviation as a crisis, each improvement as a solution.

The conventional model treats scaling as a downstream activity: you stabilize a process through bench and pilot scales, lock in a control strategy, and only then translate it upward into production. That model fits many engineered systems, and some biological systems once they are mature, particularly where scale-up principles such as oxygen transfer, mixing, and bioreactor design can be modeled and controlled with reasonable confidence (García-Ochoa & Gómez, 2009). It does not fit early-scaled myceliation as cleanly, particularly under solid-state fermentation, where scale-up is strongly shaped by heat transfer, mass transfer, substrate heterogeneity, and process-control challenges, and where the act of scaling is itself part of the experiment (Sánchez et al., 2024). In these systems, stability is something discovered through scale, under operational pressure, while the process is still becoming what it is.

Most standard improvement frameworks assume that stationarity is either already present or is quickly achievable. Statistical process control presumes a stable distribution around which control limits are meaningful, typically derived from historical “in control” behavior (American Society for Quality, n.d.; Montgomery, 2012). EVOP presumes that small, structured perturbations reveal local gradients while the process remains in a satisfactory operating state (Box & Draper, 1969). Classical root-cause analysis generally presumes that failure can be investigated through structured tools for identifying underlying causes within a defined problem context (Andersen & Fagerhaug, 2006). In early-scaled myceliation, these assumptions are, by definition, false. When the baseline itself is drifting, control limits can become either a false comfort or a constant alarm, neither of which improves understanding. When the operating regime is changing, local optimization can become a precise form of overfitting. When multiple pathways can produce the same failure, root-cause narratives can become a kind of story completion that drives action without increasing understanding.

Initial scaling as part of the learning process itself requires a different posture. Scale forces non-stationarity. Non-stationarity forces regime awareness. Regime awareness forces multi-causal reasoning. These are structural realities of early myceliation systems when they are pushed into meaningful operation before their behavior has been fully mapped.

A few concepts underpin that spine, used operationally as practitioner-facing learning concepts rather than formal statistical definitions:

A regime is a run-level behavioral state expressed consistently across multiple successive operations. It is a pattern of behavior that persists across runs, even as individual runs contain phases and dynamics that contribute to that pattern.

Coherence is the practical signature of a regime. Within a coherent window, variance has structure and relationships between inputs and outputs remain at least partially intelligible.

A manifold is the connected shape of coupled, co-varying system characteristics that move together and jointly drive outcomes, as observed operationally. In multi-causal myceliation systems, the relevant “cause” is often not a single lever but an interacting configuration: a set of conditions that is stable enough to be exploited, fragile enough to be crossed, and only partially observable through the measurements available.

This essay offers a framework for experimental learning when scaling is part of the development problem space. At its core it does not assume stationarity. It treats stability as a hypothesis that must be continuously tested, and uses regime awareness to distinguish isolated deviation from sustained rule change. It holds local and global models simultaneously to reconcile short-horizon coherence with long-horizon structure. And it treats multi-causal pathways as the default, resisting the temptation to force early myceliation into monocausal narratives. The goal is to keep human judgment well-calibrated while the process is still becoming what it is.

Non-stationarity as the Default (Stability as a Hypothesis)

Myceliation systems, particularly those operated under solid-state fermentation, are intrinsically non-stationary. Certainly non-stationarity is a property of immature processes, and a consequence of poor control, but beyond this it is a structural feature of working with living organisms embedded in material, environmental, and operational contexts that evolve over time. Any framework for experimental learning at scale must begin by acknowledging this reality.

At early scale, the expectation that a process distribution will remain fixed across successive operations is rarely justified. Instead, the system should be assumed to be drifting unless evidence suggests otherwise. Stability is something to be discovered, characterized, and continually re-validated, rather than assumed to be achievable as a short-term goal as part of the initial scaling effort.

There are a multitude of dominant sources of drift that can occur across runs, even when procedures appear unchanged.

Feedstocks derived from crops carry nutritional and physical properties that vary with season, geography, and upstream handling. Protein content, carbohydrate composition, mineral availability, and moisture behavior can drift gradually over time or shift discretely when sourcing changes. Even nominally identical formulations may present meaningfully different substrates to the organism across months or facilities.

Myceliation systems depend on inoculum that has itself undergone propagation, passage, and handling. Passage number, growth conditions during propagation, storage duration, and preparation protocols all shape the physiological state of the fungus entering each operation. Fungi are historical organisms: past conditions influence present behavior. As inoculum volumes scale, these historical effects intrinsically become more pronounced rather than less.

Environmental seasonality influences solid-state systems both directly and indirectly, interacting with facility infrastructure, insulation, airflow, and control authority in ways that are not always fully compensated by setpoints alone. Seasonal cycles can also shape raw material handling, storage conditions, and shipping timelines, affecting moisture content, physical integrity, and biological availability before substrates ever enter the process. In parallel, raw material availability and composition often track agricultural and logistical seasonality, introducing correlated shifts in input properties over time. Operational and human cadences may shift as well, as staffing patterns, maintenance schedules, and procedural attention vary across seasons, further coupling environmental rhythms to system behavior.

Together, these factors introduce systemic drift across successive operations that can manifest as slow and progressive, or sudden and jarring. The system’s behavior may change gradually, without any single run appearing anomalous, yet the operating regime after dozens of runs may differ meaningfully from the one that came before. Or conversely, the system’s behavior may shift dramatically in a single run, representing a step-change in performance regime.

Overlaying this broad drift is a second class of variability: local, run-specific noise that does not necessarily indicate a change in the underlying regime.

Sensors drift, actuators lose authority, and partial failures emerge that degrade performance without fully interrupting a run. The resulting variance is real, but often short-lived. Human operations contribute in similar ways. Training depth varies, procedures are interpreted rather than executed verbatim, and turnover introduces subtle shifts in practice. Even when documentation is stable, execution is not; processes are enacted by people, and that enactment changes over time. Biological systems themselves exhibit stochastic responses. Localized stress events, micro-environmental heterogeneity, and fluctuating contamination pressure can produce acute deviations that resolve without leaving a lasting imprint on subsequent operations.

This class of variance increases noise at the run level, but does not necessarily imply that the system has entered a new behavioral state. Stability must therefore be tested, not assumed. 

The coexistence of broad drift and local noise makes a simple assumption of stationarity untenable. In this context, stability is the persistence of structured behavior across runs. Classical control tools must therefore be reframed. A control chart tracks a process metric over time against an expected baseline and control limits, allowing shifts and trends to be detected across successive observations (American Society for Quality, n.d.; Montgomery, 2012). In this case, rather than serving solely as enforcement mechanisms within a fixed process, control charts become tests of stationarity across operations. Their role shifts from policing deviation to asking whether the system continues to behave as it has recently behaved.

Rather than being fixed historical averages, baselines are explicitly provisional. They are defined over known windows of coherent operation and revisited as the system evolves. A baseline is a reference point against which current behavior is evaluated, not a static goal.

Treating non-stationarity as the default and stability as a hypothesis sets the stage for regime awareness. It explains why learning must be organized across sequences of runs, why drift must be actively detected, and why models should remain provisional. Without this foundation, distinctions between regimes, coherence, and local versus global structure lack a meaningful basis.

Regime Awareness Through Model Coherence

You have built an MVP myceliation process and moved it into a meaningful scale for the first time. You might call it pilot or first production, but either way you are now running successive growth operations as part of characterization, stabilization, and proof-of-concept, with the added reality that scale itself is being de-risked in real time. Outcomes arrive at the end of each run as a verdict: yield, conversion efficiency, material quality, pass or fail. The verdicts are noisy. Run-to-run variability is far larger than anything a coherent commercial operation can tolerate.

At the same time, almost everything around the organism is moving. You are operating a new physical system. You have new operators and evolving procedures. Feedstocks vary across lots, seasons, sources, and substitutions. The inoculum pipeline is being pushed into volumes it has never seen before, with passage history, propagation conditions, and handling steps that now matter in new ways. The question is no longer whether variation exists, but where it lives, how much of it is structured, and how much is contingent. What should be addressed first? What can be ignored? Which “bad runs” are isolated deviations, and which indicate that the system itself has changed?

In early-scaled myceliation systems, the most important changes are rarely interpretable at the level of a single operation. Instead, they emerge gradually across sequences of operations, as the system reorganizes over time and sequence of operations. Recognizing these shifts requires treating the run as the unit of observation and the run sequence as the unit of inference. Each completed growth cycle represents a single realization of the system under a particular configuration of biological history, material inputs, environmental conditions, and operational practice. Taken individually, any one run may appear noisy or idiosyncratic. Taken together, sequences of runs reveal whether the system is still operating under the same behavioral rules.

A regime is defined by coherence. Within a regime, endpoint outcomes cluster, variability has structure, and relationships between inputs and outputs remain at least partially intelligible. The system may still be inefficient or variable, but it behaves in a recognizable way across runs.

Even though regimes are identified across operations, within-run dynamics still matter. A single growth cycle contains phases, transitions, and timing effects that often carry the biological meaning relevant to detecting and interpreting broader systemic regimes. These within-run dynamics are captured indirectly through feature engineering. Signals are summarized into descriptors of trends, variability, timing, coupling strength, lag, and other phase-sensitive quantities that compress what happened during the run into interpretable run-level features. This makes within-run structure available as explanatory material for inter-run regime shifts. When the system enters a new operational regime across runs, the shift often appears first as a reorganization of within-run timing, responsiveness, or phase structure, which then surfaces in the engineered features.

In myceliation systems, regime shifts may be driven by accumulated biological history, gradual drift in feedstock composition, seasonal environmental effects, or slow changes in the physical system. Because many of these drivers are only partially observable, regime detection relies on indirect evidence: changes in consistency, responsiveness, and explanatory structure over time.

Run-to-run variance is unavoidable in biological manufacturing, especially under early scale. A single anomalous run may reflect local noise such as a transient mechanical issue, a handling error, or a stochastic biological response. Treating every deviation as a regime shift leads to reactive thrashing and poor learning. At the same time, treating every run as an isolated event can obscure fundamental systemic problems and delay recognition of meaningful change.

Regime change is characterized by loss of structure across runs. Outcomes stop behaving like recent history, and relationships that recently explained variation no longer do so reliably. Signals that tracked performance weaken, reverse, or fragment into conditional patterns. Detecting this requires looking beyond magnitude and into structure. The question shifts from whether an individual run was bad to whether it still belongs to the same behavioral population as recent runs.

In non-stationary systems, empirical models are provisional. Any model built to explain run-level outcomes carries assumptions about biological state, material properties, and system dynamics, and those assumptions hold only as long as the system behaves in a compatible way. Recognizing this, greater learning value is often obtained by treating models as measurement instruments. Their role is to test whether the system’s current behavior remains consistent with its recent behavior, given the information available.

When models are trained over short, rolling windows of runs, they encode the local rules of the current regime. As long as those rules hold, model performance remains stable and error patterns remain interpretable. When the regime changes, the model becomes mis-specified in a diagnostic way.

A few terms are worth defining explicitly.

Model performance refers to how well the model predicts or explains outcomes on data it was not trained on, using a quantitative error score.

Residuals are the differences between observed outcomes and model-predicted outcomes for each run. Residuals summarize what the model failed to explain.

Coherence describes the consistency of the model’s behavior over time: similar performance, similar residual patterns, and similar “what matters” rankings across adjacent windows. Coherence indicates that learned relationships remain internally consistent across recent runs.

A workable way to operationalize regime awareness is to treat each completed run as a single training example and build a rolling model that is only asked to understand the recent past. The workflow is straightforward in concept, even when the system is complex.

First, define each growth operation as a single observation with an endpoint outcome metric. That outcome can be yield, conversion efficiency, a quality score, or any run-level result that captures what “good” means in the current phase of development. This outcome serves as the anchor signal.

Next, engineer run-level features from within-run signals. In solid-state fermentation, the most informative evidence often lives in dynamics rather than static values, so features should compress how signals behaved over time. Feature selection should prioritize what is operationally relevant and reflect explicit hypotheses about where system fragility or robustness is most likely to reside, given the current scale, physical constraints, and biological unknowns. Examples include slopes and curvature over defined phases, measures of variability and intermittency, simple frequency content, and lag-aware coupling descriptors between an actuator-like variable and a response-like variable. The objective is to capture within-run phase behavior in a form that can be compared across runs: timing, responsiveness, stability of control authority, and the shape of biological progression.

Then train a simple, interpretable model on a rolling window of recent runs. “Rolling” means that the training set advances through time: the model is fit using the last N operations and evaluated on the next operation or small group of operations. The exact model family is less important than the discipline. The model is local in time and intentionally provisional.

From there, three outputs are tracked across the run sequence.

Performance drift. Out-of-sample error is monitored as runs progress. Error functions as a coherence metric: sustained elevation indicates that relationships learned from recent history no longer describe the system well.

Residual structure. Residuals are examined across time. Residuals that remain small and patternless suggest limited unexplained variation. Residuals that grow, cluster, or shift sign systematically suggest new drivers or constraints. Residuals also help prioritize attention by identifying which runs were most surprising given recent history, and whether outcome drivers are likely inside or outside the current feature space.

Feature-importance volatility. Within each rolling window, feature importance is estimated using an interpretable method. Importance represents a hypothesis about what currently explains differences between runs. Stability in rankings suggests a stable sensitivity structure. Rapid reshuffling indicates that the system is reorganizing which factors dominate outcomes.

None of these outputs establishes causality on its own. Their value lies in triangulation across the run sequence. Rising error indicates loss of coherence. Residual structure shows how assumptions have become misaligned. Importance volatility points toward where explanatory power is moving. Together, these signals distinguish isolated deviations from sustained regime change and provide practical guidance about where to look next.

Loss of model coherence is often one of the earliest signals of regime transition. Predictive degradation reflects misalignment between model assumptions and system behavior. Residuals develop structure. Feature importance reorganizes. From an operator-in-the-loop perspective, these changes are valuable observations. The model is functioning as a detector.

Feature-importance volatility is especially informative in this context. When rankings reshuffle rapidly across successive windows, causal relevance becomes conditional on regime, revealing how the system’s sensitivity structure is reorganizing.

The purpose of regime awareness is to guide inquiry. When coherence holds, learning can focus on refinement within the current regime. When coherence breaks, learning shifts toward identifying what changed and where the next explanation is likely to live. Declining model validity becomes a prompt to interrogate the system rather than to optimize the model. Something has changed in the organism, the materials, the environment, the operational practice, or their interaction structure. The model has done its job by making that change visible.

In early-scaled myceliation systems, where stability is still being discovered rather than enforced, this posture is essential, and it demands learning frameworks that can distinguish what is coherent in the present from what remains structurally true over time.

Local vs. Global Modeling (Reconciling short-horizon coherence with long-horizon structure)

Once we recognize regime awareness as a practical requirement, we can recognize another structural truth. Some decisions must be made in the present tense, using only recent evidence. Other decisions should be made with the full history in view, even if that history spans multiple regimes. A single model cannot serve both purposes well, because the statistical assumptions that make a model useful in the short horizon often fail in the long horizon, and vice versa.

Local versus global modeling is the discipline of holding two complementary views of the same process: one tuned for immediate coherence, and one tuned for persistent structure.

Let's consider local models as rolling or windowed models trained on the most recent block of operations; their defining assumption is that stationarity may hold for a short time across a limited slice of runs, and that temporary stability can be used to learn what matters right now. This matters because the short horizon is where operational learning actually happens. Recent runs share similar feedstock availability, similar operator cadence, similar facility boundary conditions, and similar biological history. Within that band, relationships between features and outcomes often become legible. A rolling model can then encode the present regime’s rules with minimal complexity and with interpretable outputs.

Local models are essential for two reasons. First, they detect emerging shifts early. When a rolling model’s performance degrades, when residuals develop structure, or when feature importance reshuffles rapidly, the system is signaling that the short-horizon rules are no longer holding. This is often an early operational warning that the process has crossed a behavioral boundary. Second, local models avoid overfitting to outdated behavior. A global model trained across many months can appear stable while averaging across incompatible regimes. A local model, by design, stays sensitive to what has changed recently because it is trained only on recent history.

In practice, local models provide a living readout of current process logic. They supply regime-conditioned hypotheses, and they set the pace of inquiry.

A global model is the long-horizon counterpart to the rolling model. Where the rolling model is intentionally local in time and optimized for short-window coherence, the global model is built on the full run history and optimized for interpretability across regimes. Technically, it is trained on the same kind of run-level representation used in the rolling workflow, but with two key differences: it integrates across a much wider span of operating conditions, and it is evaluated on its ability to reveal patterns that hold beyond any single short-term regime. The global model therefore plays a different role; it is less a real-time detector and more a structural reference. It turns the run archive into a map, a way to ask where the system has been, how it moves through feature space, whether “high performance” exists in one region or multiple, and whether the system repeatedly approaches the same unstable edges. The rolling model answers: what explains outcomes right now, under current conditions? The global model answers: which of those explanations persist across time, and which are regime-specific? In early-scaled myceliation, that distinction matters because short-horizon coherence can be driven by transient circumstances. A feature can look decisive over twenty runs and disappear over the next twenty. The global model helps separate durable levers from temporary sensitivities.

The rolling model generates time-local hypotheses through performance, residual structure, and feature importance. Those hypotheses are then projected into the global layer in two complementary ways.

First, the global model provides a way to place recent runs in context. By projecting run-level features into a compressed representation, it becomes possible to see whether the current operating window sits in a region the system has occupied before, whether it is moving gradually through familiar territory, or whether it has crossed into a part of the system’s behavior space that is genuinely new. This context matters for interpretation. When feature importance shifts while the system remains in the same region, it often reflects a reweighting of sensitivities within an existing regime. When the same shift coincides with movement into a new region, it suggests a change in the governing constraints of the system.

Second, the global layer allows local signals to be tested for persistence. A global, interpretable model trained across the full run history can be used to ask whether a feature that appears important in the current window also matters more broadly, or whether its influence is confined to a narrow set of conditions. It can also show whether certain features are consistently associated with stable outcomes, or whether they tend to appear near the edges of performance where variance increases and failures cluster. This makes the global model useful for decision making: it helps distinguish durable levers from temporary correlations, and recurring risks from isolated anomalies.

Used this way, the global model functions to discipline the rolling model. The rolling model preserves sensitivity to the present, and the global model prevents over-commitment to short-term stories by providing an interpretive frame built from the system’s full behavioral record.

Let's consider an example of a practical sequence:

A team applies a rolling modeling workflow across successive operations and watches how its behavior changes over time. For a period, model performance is stable and a small set of features consistently explains differences in outcomes between runs. Later, outcomes remain strong, but the set of features that appear most important begins to shift. Eventually, prediction error increases and residuals show structure, suggesting that something influencing performance is no longer being captured. Local models make these changes visible as they happen, but they leave an important question unanswered: do these shifts represent normal variation within a single operating pattern, or do they indicate that the system has entered a different behavioral state that requires a different response?

This is where a global model becomes useful.

Using the same run-level features developed for the rolling analysis, the team looks at the full history of operations at once. Rather than focusing on recent behavior, the global view asks how all runs relate to one another. By projecting the data into a simplified representation, it becomes possible to see whether recent runs resemble earlier ones, whether the system has moved gradually over time, or whether it has crossed into a region of behavior it has not occupied before. A global, interpretable model built on this full dataset can then be used to understand which factors matter consistently, which only matter in certain contexts, and how performance differs across these regions.

Two outcomes from this global step are especially actionable.

One is the identification of persistent drivers: features that remain influential across the full dataset, even when local regimes change. These provide stable levers and stable narratives, and they often point toward physical constraints or structural process truths rather than transient operational artifacts.

The second is the identification of portions of the global state space where outcomes become unstable or failure clusters. A rolling model may flag an emerging issue in real time; a global model can show whether that issue is novel or recurring, whether it correlates with seasonal context, whether it coincides with specific classes of runs, and whether the system has visited that region before.

Local models keep attention anchored to the present regime and make coherence a continuously tested condition. They protect learning velocity, because they respond quickly to change. Global models prevent reactive thrashing by providing a long-horizon context that stabilizes interpretation; showing what persists, what repeats, and what is structurally distinct rather than temporarily noisy.

Together, they provide situational awareness across time. The workflow becomes a paired instrument: rolling models detect and characterize the current regime, while global models contextualize that regime within the system’s full behavioral landscape. In both cases, explainability is used for reasoning and alignment rather than automation, which will come with process maturity. The goal here is not autonomous control, but to keep human judgment well-calibrated in a system that is still becoming what it is.

Multi-Causal Structure: Why Root Cause Thinking Fails in Early-Scaled Myceliation

At early scale, myceliation outcomes are produced by manifolds: coupled configurations of biological state, feedstock properties, environmental dynamics, physical hardware, and operational practice acting synergistically. The reflex to identify "the root cause" when performance degrades is not merely incomplete—it is structurally misaligned with how these systems fail.

Classical manufacturing carries an expectation that failures trace to singular origins: something broke, something changed, something went wrong. In stable, well-bounded systems, that expectation may hold. In early-scaled myceliation, it becomes a trap. Multiple pathways act in parallel. Components drift or deviate simultaneously. Relevant internal states and histories remain unobservable. Performance loss does not reflect a single cause waiting to be isolated, but rather the convergence of multiple constraints, each tolerable on its own, whose interaction crosses a functional boundary. Their relative importance shifts from one operating regime to another, and the system's sensitivity structure reorganizes as conditions evolve. Under these conditions, monocausal reasoning collapses it prematurely, forcing multi-dimensional failure into single-factor narratives that drive action without increasing understanding.

In solid-state myceliation, the same endpoint result can be reached through different routes. A yield drop might be driven by resource limitation in one period and by stress accumulation in another. A quality failure might trace to environmental mismatch one week and inoculum state the next. Even more often, the failure is not a single pathway at all, but a combination. One factor may be suboptimal but tolerable on its own. Another may be drifting slowly in the background. The run “fails” when those effects overlap strongly enough to cross a boundary. In that moment, single-factor RCA tends to misattribute causality, because the last visible deviation is not necessarily the full explanation. The run did not fail because one thing changed; it failed because several things were already near their limits, and one additional nudge pushed the system over.

This is why traditional root-cause analysis can fall apart at early scale. Classical RCA works best when the process is broadly stationary, failure modes are relatively discrete, and the system can be returned to a known baseline. In early myceliation, the baseline itself is moving. Treating each failure as an isolated event encourages “story completion”: selecting the most plausible recent anomaly and calling it the root cause. That produces action, but it may produce wrong or incomplete action.

As previously discussed, feature importance is a regime-conditioned hypothesis: given what we measured, and given the recent operating window, which signals were most useful for explaining outcome differences across runs. In a multi-causal system, that matters because different pathways dominate under different conditions. A feature can look decisive for twenty runs because it is the active constraint in that regime, and then vanish when a different constraint becomes limiting. Importance should therefore be read alongside coherence and residual structure: when coherence holds, importance changes often reflect real shifts in constraint within the observed space; when coherence breaks, importance can become the model’s best attempt to explain outcomes while an unmeasured driver is pulling the system. The practical question becomes directional: do we tune a lever we already track, or do we expand the feature space because the active pathway is currently outside what we can see?

Once multi-causal structure is acknowledged, the goal of analysis shifts. The objective is to identify which causal pathways appear active, which constraints appear to be tightening, and whether multiple drivers are interacting. In essence, it is to find the manifold—meaning the connected co-varying system characteristics—that is functioning together as a cohesive effector. To this end, multivariate modeling across runs is an operational response to this potential manifold structure of the system. Rolling feature-importance snapshots become regime-conditioned hypotheses about what is currently limiting. Residual patterns and coherence breakdown become signals that the system has moved, or that the current representation is missing a driver. Taken together, these tools support a style of engineering that can act without pretending the system has already been reduced to stable, coherent practice.

In early-scaled myceliation, learning fails when complexity is forced into a single story. It succeeds when complexity is structured to be reasoned about. Multi-causal thinking strives to make uncertainty directional. It helps you choose where to look next, what to test next, and which explanations are likely incomplete.

Learning While the System Is Still Becoming

Early-scaled myceliation occupies an uncomfortable but consequential middle ground. The system has moved beyond exploratory play, yet it has not earned the assumptions that support mature process control. Each run carries real consequence in cost, time, and opportunity, but individual outcomes are still weak evidence in isolation. Learning happens under pressure, and interpretation itself becomes a source of risk. What matters most in this phase is not eliminating variability, but deciding which signals deserve belief and which should be held loosely.

This framework, by treating non-stationarity as the default, rejects the expectation that stability should arrive early or cheaply. By framing stability as a hypothesis, it replaces assumption with continuous testing. Elevating regime awareness shifts attention from individual outcomes to patterns across time. And by holding local and global models simultaneously, it reconciles the need to act in the present with the need to learn from history. And by embracing multi-causal structure, it resists the temptation to collapse complex behavior into a single story simply because a decision must be made. Critically, the goal is to convert the pressure to react into the discipline to learn. When every run feels like a test you're failing, that discipline is what keeps you from mistaking motion for progress, and noise for signal.

References

American Society for Quality. (n.d.). Control chart. Retrieved May 27, 2026, from https://asq.org/quality-resources/control-chart

Andersen, B., & Fagerhaug, T. (2006). Root cause analysis: Simplified tools and techniques (2nd ed.). ASQ Quality Press.

Box, G. E. P., & Draper, N. R. (1969). Evolutionary operation: A statistical method for process improvement. Wiley.

García-Ochoa, F., & Gómez, E. (2009). Bioreactor scale-up and oxygen transfer rate in microbial processes: An overview. Biotechnology Advances, 27(2), 153–176. https://doi.org/10.1016/j.biotechadv.2008.10.006

Montgomery, D. C. (2012). Introduction to statistical quality control (7th ed.). Wiley.

Sánchez, A., Oiza, N., Artola, A., Font, X., Barrena, R., Moral-Vico, J., & Gea, T. (2024). Solid-state fermentation: A review of its opportunities and challenges in the framework of circular bioeconomy. Afinidad, 81(601), 50–56. https://doi.org/10.55815/424209