Institutions often have to act before they know enough. A new policy may be introduced because an existing problem is urgent, because available evidence is incomplete but promising, or because decision-makers need to test an intervention under real conditions. There is nothing inherently irrational about this. Governing under uncertainty sometimes requires provisional action. The difficulty begins later, when the provisional character of that action disappears from institutional memory even though the uncertainty that originally surrounded it has never been fully resolved.

Implementation has a powerful normalising effect. Once a policy acquires a budget, procedures, staff responsibilities, reporting systems and legal or administrative routines, it starts to look settled. People learn how to operate it, organisations adapt around it and future decisions increasingly take its existence for granted. What began as an uncertain intervention can therefore acquire the appearance of established knowledge simply because it has become operationally familiar.

This creates a dangerous confusion between two different statuses. A policy can be fully implemented while remaining only partially justified. It can be administratively mature while epistemically provisional. Implementation status tells us whether the institution has organised itself to deliver an intervention. Epistemic status tells us how confident the institution should be that the intervention actually works as intended, under which conditions and through which mechanisms. The two can evolve at very different speeds.

Consider a programme introduced because early evidence suggested that it might improve access to a public service. The institution decides that the evidence is good enough to justify action, but not strong enough to close the question. At the beginning, everyone understands that the intervention is still being tested in practice. Several years later, however, the programme has its own team, budget line, performance indicators and established routines. New staff encounter it as part of the normal institutional landscape. The original uncertainty is no longer visible. Yet nothing may have happened during those years that actually converted the initial hypothesis into stronger knowledge.

The policy has become more established without becoming more certain.

That distinction matters because institutions can mistake institutionalisation for validation. Repetition creates familiarity, familiarity creates expectation and expectation can gradually be interpreted as evidence. A programme that has existed for ten years may feel more legitimate than one introduced last month, but duration alone does not tell us whether the causal assumptions behind it have been tested. The institution may simply have become better at implementing something whose underlying effectiveness remains uncertain.

This is different from the problem of evidence expiry. Evidence expiry occurs when a result that was once sufficiently informative becomes less transferable because relevant conditions have changed. Forgotten provisionality describes another path. The institution may never have possessed strong enough evidence to treat the question as fully settled in the first place. The intervention was introduced under uncertainty, but implementation gradually erased the uncertainty from institutional awareness.

Both processes can eventually produce the same outward condition: a durable policy supported by weaker knowledge than its institutional confidence suggests. Yet the diagnosis differs. In one case, confidence was initially justified and later became stale. In the other, confidence was always conditional, but the condition was forgotten.

This suggests that policies need an epistemic status as well as an operational one. Institutions already classify interventions in many ways: pilot, programme, regulation, service, temporary measure, permanent scheme. These labels describe administrative form, but they do not necessarily describe the quality of the knowledge supporting the intervention. A permanent programme can still rest on provisional evidence, just as a temporary pilot can test a mechanism already supported by substantial research. Administrative category and evidential status should not be treated as synonyms.

An epistemic status would make explicit what the institution believes it knows. Some policies might be supported by well-established evidence across multiple contexts. Others might be conditionally supported, with important assumptions still uncertain. Some could remain explicitly experimental, while others might rely largely on plausible theory, expert judgement or necessity in the absence of strong empirical evidence. The purpose of such distinctions would not be to create a simplistic hierarchy of good and bad policies. It would be to preserve institutional awareness of which claims remain open.

That awareness changes how a policy should be governed. If an intervention remains experimental, implementation should generate opportunities to learn. Data collection should reflect unresolved questions rather than only operational performance. Variation across locations or populations may become analytically useful. Unexpected outcomes should be investigated rather than treated solely as delivery problems. Where feasible, institutions may preserve comparison groups, phased implementation or other structures that allow stronger inference.

Without this awareness, implementation can unintentionally destroy learning opportunities. A programme introduced everywhere at once may eliminate useful comparison. Procedures can become standardised before alternative mechanisms are understood. Performance metrics may focus on whether activity occurred rather than whether the intervention produced the intended effects. By the time someone asks whether the policy actually works, the institution may have made the question much harder to answer.

Epistemic status also matters for communication inside institutions. Operational teams need to know whether they are delivering an established intervention or participating in an ongoing inquiry. Senior decision-makers need to know which programmes rest on strong evidence and which depend on assumptions that remain uncertain. Analysts need to understand where further learning has the greatest value. If those distinctions disappear, institutional attention can become poorly allocated: heavily validated programmes are repeatedly studied while deeply uncertain ones continue largely unquestioned because they have become routine.

The same problem affects institutional memory. Organisations often preserve records of what decisions were made but lose the degree of confidence attached to them. A decision document may show that a programme was approved while failing to preserve whether the underlying evidence was considered conclusive, promising, weak or merely the best available under urgent conditions. Later generations inherit the decision but not the uncertainty. What was once recorded as “we think this may work well enough to try” gradually becomes “this is how the institution knows the problem should be addressed”.

That transformation is particularly important in public institutions because policies can persist across changes in leadership and personnel. The people who originally understood why an intervention was provisional may leave. Political attention moves elsewhere. The policy remains embedded in administrative systems that no longer contain the context of its creation. Unless uncertainty itself has been institutionalised, it can disappear even while the programme it belonged to survives.

Artificial intelligence makes this issue especially visible because institutions are increasingly deploying systems whose effects, limitations and interactions with human judgement may still be uncertain. A model can move rapidly from pilot to operational infrastructure, and once embedded, its outputs may become part of routine decision-making. Yet operational adoption does not resolve questions about bias, distribution shift, human reliance, behavioural adaptation or long-term institutional effects. The fact that a system is deployed tells us that the institution has chosen to use it, not that every important question about it has been answered.

The principle extends far beyond technology. New regulatory approaches, behavioural interventions, organisational reforms and service models can all move from trial to routine while important uncertainties remain open. The challenge is therefore not to label everything experimental indefinitely, but to avoid allowing institutional familiarity to answer an epistemic question that evidence has not answered.

Eventually, some policies should graduate from experimental status. Repeated evidence can strengthen confidence, mechanisms can become better understood and uncertainty can narrow enough that continued experimental treatment adds little value. Other interventions may remain conditional because their effects vary substantially by context. Some may produce evidence strong enough to justify revision or abandonment. The important capability is to make these transitions deliberately rather than allowing administrative permanence to decide them implicitly.

This requires institutions to preserve uncertainty as actionable information. A policy record should contain not only what intervention was chosen and why, but also what remained unknown when the decision was made. Those unresolved questions should be capable of surviving changes in personnel, organisational structure and political attention. Otherwise the institution can accumulate policies while progressively losing track of which parts of its own operating model are well established and which still amount to working hypotheses.

There is a deeper implication here for institutional intelligence. An intelligent institution does not merely distinguish knowledge from ignorance. It can also represent degrees and forms of uncertainty within the things it is already doing. It knows that action does not always imply certainty, and that commitment does not have to erase curiosity. In fact, some of the most valuable institutional questions are precisely those attached to policies already in operation, because those policies create continuing opportunities to learn from the world they are affecting.

Institutions should know which policies are still experiments because implementation can make an uncertain intervention look more settled than the evidence actually warrants. Operational maturity is not the same as epistemic maturity, and repeated delivery does not by itself transform a hypothesis into knowledge. A cognitively disciplined institution therefore preserves the uncertainty attached to its decisions, tracks which questions remain open and ensures that implementation continues to generate opportunities to answer them. The purpose is not to keep policy permanently provisional, but to make sure that confidence grows because knowledge has improved rather than simply because the institution has become accustomed to what it already does.