The Locus of Competence: Composable, Governed Intelligence and the Architecture of Defensible Decisions

Machine competence does not have to live inside a model's parameters. A capable system can be split into a general policy that decides what to do and a library of specialized, inspectable, executable components that do it. Small models now show this in practice, solving technical problems they have not memorized by calling external tools. This paper makes a narrower claim about where that approach stops being enough. For decisions that carry institutional weight, the difficulty is not running the right component but composing several of them without losing what each one guarantees. Representation, evidence, causal inference, calibration, and governance each give a partial guarantee, and those guarantees are fragile under composition, because a guarantee added late cannot repair an admissibility condition that was broken early. I argue that the trustworthiness of a composed decision is bounded by the weakest condition it violates anywhere in the pipeline, that learning from outcomes has to preserve auditability to be admissible at all, and that the right test of such a system is whether it can close a single measured, governed decision loop.

For most of the modern history of machine learning the answer has been simple. It lives in the weights. Competence is whatever a model has absorbed from data and stored in its parameters, so to make a system more capable you enlarge the model, widen the data, and train for longer. Sutton's “bitter lesson” turned this into something close to doctrine. General methods that lean on computation, mainly search and learning, tend to overtake systems built around hand-encoded human knowledge, and the field keeps relearning this at the expense of whatever scaffolding it had grown fond of [1].

There is now an obvious counter-current. A small open model running on a laptop can solve problems it has plainly never memorized. It can compute the deflection of a loaded beam, trace a crack through an atomic lattice, or find the lightest shape of a structural member that still carries its load. It does none of this by holding the physics. It selects an external component that holds the physics, runs that component, reads the result, and decides what to do next. The competence sits in the tool. What the model supplies is the judgment about which tool to reach for and what to make of its output.

I do not read this as a refutation of the bitter lesson so much as a clarification of where generality is worth spending. The policy that orchestrates stays general and learned, which is exactly the thing Sutton was defending. What moves outside the model is not human heuristics smuggled back in through a side door, but competence that can be inspected and rerun and that carries its own provenance. So the old question has a more interesting answer than weights or rules. Competence can live in a composition: a general policy operating over a library of specific, governed components. The rest of this paper is about what that composition has to satisfy before it can be trusted with anything that matters.

Tasks differ in the one property that matters here, which is the shape of their ground truth.

In the first kind of task there is a checkable answer and it arrives quickly. A beam deflects by the amount mechanics predicts. A simulation reproduces an experiment, or it does not. An optimized part comes out measurably lighter. Feedback is cheap and almost immediate, and a run either met the criterion or failed it. Here, externalized competence comes close to telling the whole story, because if the tool is right and the model calls it correctly, the system is right. The science demonstration belongs to this kind of task, and a good part of its force comes from operating where checking is easy.

The second kind of task has no answer in the back of the book. When a bank tightens credit policy, when a ministry moves a multi-decade investment programme from one purpose to another, when an insurer pulls out of an exposed coastline, the right answer is contested, the outcome that would settle it lands years later if it lands at all, and someone has to defend the decision in front of a board, a regulator, or the public. Truth is partial and slow. The cost of a mistake falls on people who had no part in making it. And the act of deciding is itself something that has to be authorized and later reconstructed.

These two situations ask for different architectures, and treating them as one is the mistake I most want to name. In the first, calling the right tool is most of the job. In the second it is a small fraction of the job, because the hard part has moved. It is no longer computing an answer. It is assembling a justification, and a justification has structural requirements that a correct number does not.

It helps to be concrete about the decomposition. A system of this kind has a policy, call it the orchestrator, and a library of components. Each component can be executed rather than merely consulted, can be inspected down to its assumptions and logic, and can, in a well-built system, be authorized and logged when it runs. The orchestrator maps a request and its context to a sequence of component calls and stitches the results together.

The advantages are real. New capability arrives by adding a component rather than retraining the model. A component's behavior can be audited in a way the interior of a large network cannot. And capability accumulates, since components are permanent and shared while the policy stays put. For tasks of the first kind this is most of the architecture. My claim is that for tasks of the second kind it is the easy half, and that mistaking it for the whole produces systems that look impressive in a demonstration and turn out to be unusable inside an institution. The reason is everything a defensible decision needs beyond a correctly executed tool.

A justified decision in the second regime rests on five things, and each supplies something the others cannot.

The first is representation: a typed and consistent account of the entities, relations, rules, and units of the domain, so that the symbols the system manipulates actually denote what they are taken to denote. What this buys is semantic admissibility, the modest but easily lost property that the question being computed is the question that was asked.

The second is evidence: a record of what was observed and what was decided, with provenance, and with the ability to reconstruct what was known at any earlier moment. Call this epistemic admissibility. Claims rest on traceable, time-correct evidence rather than on whatever happens to be convenient at the moment of asking.

The third is causal inference: the ability to represent interventions and counterfactuals, to say what changes if we act rather than what merely tends to occur alongside what. The property at stake is interventional validity, that the quantity being estimated corresponds to the decision actually under consideration.

The fourth is calibration: a statement of uncertainty that means something operationally, together with a willingness to abstain when the evidence is too thin to support a claim. Confidence has to be earned, and the system has to be able to decline.

The fifth is governance: authorization, purpose limitation, approval, and a record around each action. This is institutional admissibility, the property that the decision was allowed to be made and can be shown afterwards to have been made properly.

Each of these is partial, and the partiality is the whole problem. A perfectly calibrated estimate of the wrong quantity is precise and useless. A valid causal effect computed over inadmissible evidence is rigorous in form and wrong in fact. A flawless audit trail around an unjustified inference documents the wrong thing carefully. The five are not options on a menu. They are joint preconditions. The assumption that fails, quietly and often, is that assembling components which each carry a good guarantee yields a system that carries one.

This is where the real content of the position sits. Competence composes. Guarantees do not, unless the architecture is built to carry them.

Writing a defensible decision as a composition makes this easier to see. In the vocabulary we use at Helios, the chain of operations runs from harvesting how the institution works, through encoding it, recalling time-correct evidence, inferring causally, calibrating honestly, all orchestrated by a learned policy and stewarded by authorization and provenance. A sixth operation, LEARN, closes the loop by feeding measured outcomes back into the evidence and the calibration. I come to it in the next section.

Setting it out this way makes the failure modes visible.

Selection quietly destroys coverage. Distribution-free calibration, conformal prediction for instance, holds its guarantee under conditions such as exchangeability and a target that was not chosen by looking at the same data used to calibrate it [3]. A pipeline can break those conditions without anyone noticing. If the orchestrator picks which question to answer by glancing at outcomes, or runs many analyses and reports the one that stands out, the guarantee that justified the calibration step is already gone. So composition has to account for selection and multiplicity in the architecture itself, rather than leave it to the good habits of whoever is driving.

Inadmissible evidence breaks identification. A causal quantity such as P(Y | do(X)) is identified only under structural conditions, an admissible adjustment set and the absence of open confounding paths [2]. The inference step inherits its validity from representation and evidence. If the representation leaves out a confounder, or the evidence used to adjust is not admissible, the estimate is biased no matter how faithfully it is computed or how well it is calibrated afterwards. A guarantee added late cannot repair an admissibility condition that was violated early.

A single unrecorded step makes the whole thing unauditable. Governance is end to end or it is nothing. If any link in the chain has no provenance, a data pull, a transformation, a human override that went unlogged, the composed decision cannot be reconstructed, and the institutional guarantee fails for the whole even when every other step was sound.

What ties these together is a bound. The trustworthiness of a composed decision is set by the weakest admissibility condition it violates anywhere along the way, not by the strength of its best component. A system earns trust to the degree that its architecture carries each component's preconditions forward and refuses to let a later, stronger-looking guarantee paper over an earlier broken one. This is why the second regime cannot be reached by collecting better tools. Better tools improve the parts. Only an architecture improves the composition.

The components approach promises that competence accumulates. In the second regime the thing most worth accumulating is not the count of tools but the system's grip on its own track record: what it predicted, what was decided, and what actually happened. That is the work of the LEARN step, and it comes with a demand of its own.

Learning from outcomes must not corrupt the record that makes decisions auditable. Overwriting beliefs as new information arrives is the obvious way to do it and the wrong one, because it destroys the ability to answer what was known and when, which was the point of epistemic admissibility in the first place. The way out is to treat correction as bitemporal and non-destructive. Valid time, meaning when something held in the world, is kept separate from transaction time, meaning when the system came to record it, and corrections add new records instead of erasing old ones [4]. The learned state is then a deterministic projection of an append-only, governed ledger, so that getting better and staying auditable stop being in tension [5].

This gives the property I want to insist on. A system updated by an outcome has to be the same kind of governed object it was before, reachable by replaying the ledger, with no step of learning able to move it outside the space of auditable states. An approach in which competence accumulates while accountability decays with every update is the wrong fit for the institutions that need it most. This closure is what lets a model be living and governed at the same time.

A position should be possible to put at risk. My claim is not that some particular system is better than another. It is that the second regime is reached only by a certain kind of architecture, and that claim can be tested against a single observable.

The test is whether the system can close a loop on a real decision. A consequential choice is rehearsed before it is made, with its outcomes simulated and its uncertainty quantified. It is then made. It is then measured against what actually happened. And the result is fed back so that the next decision is demonstrably less wrong, with every step of the loop authorized and recorded. A benchmark will not do, because a benchmark lives in the first regime. An architecture diagram will not do either, because it asserts rather than shows. One decision that completes the circuit under governance is the thing that counts. A system that cannot close this loop has not yet earned the second regime, however capable its tools, and the size of the model and the ingenuity of the component are beside the point with respect to this test.

The position comes down to this. Competence does not have to live in the weights. It can live in a composition of a general policy over inspectable, executable, governed components, and that is a sound place for it to live. The decomposition is enough where verification is cheap and not enough where decisions carry institutional weight, because there the difficulty is composition rather than execution. The parts of a defensible decision, representation, evidence, causal inference, calibration, and governance, each give only a partial guarantee, and those guarantees survive composition only if the architecture is built to carry them, with the trustworthiness of the result bounded by the weakest condition it breaks. Learning that accumulates competence has to satisfy a closure property, non-destructive and ledger-projected, so that a system can be living and governed at once. The standard by which any of this should be judged is plain enough. Whether it can close one measured, governed decision loop.

The intelligence is moving out of the weights. Whether it lands in tools that merely run, or in systems that represent, remember, reason, quantify their own uncertainty, and answer for themselves, is what separates a capable demonstration from infrastructure an institution can rely on. The formal development of each operation named here, the composed causal world model, the governance ledger, the differentiable simulation layer, the neurosymbolic representation, and the bitemporal evidence substrate, is set out in the Helios Technical Report Series [6].

[1] R. Sutton. The Bitter Lesson. 2019.

[2] J. Pearl. Causality: Models, Reasoning, and Inference. 2nd ed., Cambridge University Press, 2009.

[3] V. Vovk, A. Gammerman, and G. Shafer. Algorithmic Learning in a Random World. Springer, 2005. See also A. N. Angelopoulos and S. Bates. A Gentle Introduction to Conformal Prediction and Distribution-Free Uncertainty Quantification. 2023.

[4] R. T. Snodgrass. Developing Time-Oriented Database Applications in SQL. Morgan Kaufmann, 2000.

[5] Helios Brain Technical Report HB-TR-2026-05. The Living World Model: A Bitemporal Evidence Substrate for Memory, Provenance, and Continuous Learning. 2026.

[6] Helios Brain Technical Report Series, HB-TR-2026-01 to HB-TR-2026-05. 2026.