Attempts to delegate open-ended research to autonomous agents fail in a recognizable pattern. Promising directions are abandoned on weak or synthetic evidence. Budgets end largely unspent even though usage is fully monitorable. Self-reviews surface the very flaws external reviewers later raise, yet plans do not change. Ambitious targets are retired early and never revisited. Explicit rules on cadence, scope, and output are simply ignored.

The standard reading treats each of these as a capability gap and asks how quickly better models will close it. This article proposes a different reading: the five failures share a single condition. Research is institutional work, and the processes performing it were never constituted as institutions.

Research is an institution, not a workflow

A workflow consumes inputs and produces outputs; when it ends, nothing of it needs to remain. An institution maintains a state that outlives every trajectory through it.

Open-ended research has the second character. A hypothesis is not a passing thought: it has identity, an owner, and a lifecycle, and abandoning it is an event, not a silence. Evidence is not text: it is a claim whose weight depends on where it came from and how it was produced. A budget is not a number to observe: it is an obligation someone answers for. A review is not commentary: it binds the plan it addresses. An instruction is not a preference: it constrains what may happen next.

None of these facts can live inside a single reasoning trajectory. They are properties of the institution’s state — persistent, shared, and normative. A process that holds them only in its working memory does not merely risk forgetting them; it never possessed them in the institutional sense at all.

A hypothesis that can be silently abandoned was never a hypothesis. It was a thought.

Preference and constraint

The deepest distinction the failures expose is between what is expressed to a process and what is evaluated on it.

An instruction in a prompt is a preference expressed to the process: the entity being governed also judges its own compliance. A norm in a governance substrate is a constraint evaluated on the process: judgment is rendered by a layer the process cannot argue with, defer, or reinterpret. Prompting compresses the entire institution — its ontology, its obligations, its authority — into prose, and then hopes the process will simulate the enforcement it was never given.

Systemics has held this point for decades: regulation requires a regulator distinct from the process it regulates, carrying its own model of that process. In agentic research today, that regulator is the harness — the scaffold of prompts and loops around the model — and it is typically invisible, unversioned, and unenforced. The executive function exists; it is simply not an object anyone can inspect, govern, or reproduce.

The same asymmetry governs enterprise assistants: a model that decides for itself which retrieved passage is authoritative is judging its own compliance. Making that judgment external is what Better context beats bigger models calls disposition — models propose, governed systems dispose. Why disposition can never be delegated to the model — because concerns are owned by stakeholders, and a model can hold no institutional standing — is stated as the full Poesis position in Governance owns the concerns.

This is why the five failures are one. Judgment, stewardship, responsiveness, reconsideration, and compliance are not five skills a process lacks. They are five obligations that had no layer to live in.

What a governance substrate must provide

If the diagnosis is constitutional rather than cognitive, the remedy is a substrate with a specific set of properties — stated here independently of any particular technology.

An explicit ontology. What a hypothesis, a budget, a review, or a deadline is must be defined before any process can be held to it. Semantics carried only in prose bind no one.

Governed identity and lifecycle. The objects of research must persist across trajectories, with state changes that are events — attributable, ordered, and revisable only by governed means. What was retired can then be reconsidered, because the record of its retirement exists and the identity it belonged to survives.

Evidence as qualified claim. An observation enters the institution with its confidence and its provenance attached, or it does not enter at all. Certainty is a fact; everything else is a contribution whose weight is inspectable. Qualification does not make a claim true — a source can be confidently wrong — but it makes the claim governable.

Evaluable norms with temporal tolerance. Institutional obligations are rarely instantaneous. An hour of idleness violates nothing; chronic under-use of a mandate does. A substrate must distinguish momentary deviation from sustained violation, and render the difference as a verdict rather than an impression.

Separated measurement, judgment, and enforcement. The layer that observes a fact, the layer that judges it, and the layer that acts on the judgment must be distinct — because a process that measures, judges, and enforces itself is precisely the ungoverned condition being remedied. This is the same separation that keeps description from conferring its own legitimacy, argued in Definition versus description.

Deterministic verdicts. Whether an obligation was met must have exactly one answer, computable by anyone holding the definitions and the record. Governance that cannot be reproduced is opinion.

The record as a primary artifact. Reconstructing what happened from raw logs is archaeology. In an institution, the history of decisions, evidence, and verdicts is the substrate’s own state, queryable the moment it forms.

None of these properties makes a process more intelligent. They make it accountable — and, just as importantly, re-enterable: any capable process, human or artificial, can pick up the institution’s state and continue, because that state does not live in anyone’s context window.

The Poesis stack as one such substrate

The Poesis stack is our attempt to build this layer — not an agent framework, but the governance substrate agent frameworks lack.

GSM, the specification underneath the stack, supplies the ontology and the generative governance grammar: definitions with stable identity and versioned, lifecycle-managed content; Directives that open scopes of governing intent; Norms that make selected consequences of that intent evaluable, with explicit temporal tolerance; and a strict separation between the primitives that measure, the primitives that judge, and the primitives that act.

The Definition Manager hosts that model and refuses incoherent states — identity cannot mutate, lifecycles cannot skip. The Operator evaluates typed rules deterministically over it: same definitions, same inputs, same verdict. The Definition Blackboard, in development, admits contributions only with confidence and provenance envelopes attached, over an append-only audit ledger. An observability floor, also in development, is designed so that governance-relevant behavior is never unobserved.

As elsewhere in these pages, implemented and designed are kept distinct: the Definition Manager and the Operator exist; the Definition Blackboard, the observation and derivation planes, and the observability floor are in development. The argument of this article is architectural, not a product demonstration.

Under such a substrate, the five failures change status. Abandoning a hypothesis becomes a lifecycle event with an evidential basis, not a silence. A budget becomes an owned obligation whose chronic neglect is a verdict, not a log entry. A review becomes a governed coupling whose persistent disregard is measurable. A retired direction remains a re-openable identity rather than a lost branch. An instruction becomes a constraint whose violation prevents the act instead of annotating it. Nothing in this list makes the researching process smarter; every item makes the epistemic process governed.

Bottlenecks become governed objects

The larger stake is recursive self-improvement. If autonomous processes cannot yet conduct open-ended research, improvement loops are bottlenecked — and Amdahl’s law caps the payoff of accelerating only the parts that can be verified.

A governance substrate does not remove those bottlenecks. Its contribution is more austere: it makes them explicit, attributed, and measurable. When research is a governed system, which capabilities are delegated and which are reserved to human roles is a definition, not a habit; where the process stalls is a verdict, not an anecdote. How far automation actually reaches stops being a forecast and becomes something the system’s own record answers.

That reframes self-improvement itself: from a mystical loop of agents improving agents into an institutional process whose degree of automation is, at every point, a defined, versioned, auditable fact.

The reflexive note

One reason we hold this thesis with some confidence is that the stack applies it to itself. GSM’s criterion for systemness is reflexivity — a structure becomes a system when its own mechanisms produce its own governance — and the Poesis stack is built that way, from the self-typing archetype at the root of its type system to the governed portfolio process through which agentic teams deliver it. That delivery process runs on SAF, whose separation of a deterministic harness from the agentic organization it executes is described in Better context beats bigger models. The failure pattern catalogued at the outset is the one our own delivery process is constituted against, for the same reason and with the same machinery.

Conclusion

The path to open-ended machine research is not only better models. Judgment, stewardship, responsiveness, reconsideration, and compliance are not independent capabilities to be hill-climbed; they are one property seen from five angles — the property of being governed. A scaffold that matters is a governance layer, and a governance layer worth trusting is explicit, typed, deterministic, and auditable. That is what the Poesis stack is: not a replacement for agents but the substrate that governs them; not a replacement for human institutions but their formalization.

The agents didn’t fail at research. They failed at being an institution — because nobody gave them one.