Continuous Probabilistic Foresight

On March 9, 2023, depositors pulled more than $40 billion out of Silicon Valley Bank in a single day. Regulators closed the bank the next morning. The Federal Reserve's own postmortem on the failure found that its supervisors had already identified what killed the bank: heavy exposure to rising rates, a deposit base that was concentrated and largely uninsured. The risk had been flagged inside the institution responsible for catching it. What never happened was an update to the standing judgment fast enough for anyone to act on it before the run started.

Most analysis fails the same way, only slower. A question gets assigned, research gets collected, a report gets written, decisions or policy get made based on that report, and the analyst moves on to the next research question. The report can be rigorous on the day it publishes, but the world keeps changing after that while the reasoning inside a finished document does not. Continuous Probabilistic Foresight, or CPF, is Cultivate's methodology behind Hinsley: a way of keeping the evidence, the assumptions, the alternative futures, the forecasts, and the record of what changed and why connected and active for as long as a research question remains relevant.

This guide describes the discipline CPF asks for at each stage of an analysis, independent of Hinsley or any other tool. Two companion guides cover parts of the method in more depth, with a worked example: Authoring Good Analysis Questions covers the first step below, and Developing Good Scenarios covers building the alternative futures a CPF analysis depends on. More guides are on the way.

CPF is a disciplined and structured way to state uncertainty honestly, change a judgment when the evidence changes, and learn from an outcome without rewriting history once it is known.

CPF is also the infrastructure other work gets assessed against. A set of decision paths, a portfolio of programs, a register of future risks, or a body of ongoing assumptions all rest on some view of how a research question is likely to develop, and CPF is what supplies and keeps that view current. What sits on top of it then inherits evidence and reasoning that stay traceable.

Why analysis needs a continuous method

A one-time report creates two problems that compound over time. As new evidence arrives, a reader has no way to tell which parts of the original assessment still hold. And the organization loses the chance to learn whether its judgments were sound, whether warning arrived early enough, or whether the way the issue was defined caused it to miss what happened.

Organizations rarely notice this happening in real time. A 2016 study of Nokia's collapse in the smartphone market, published in Administrative Science Quarterly, found that Nokia's own middle managers were aware the iPhone had changed the competitive picture well before the company's strategy caught up with that reality. Fear of contradicting senior leadership's existing view kept that evidence from reaching the judgment that mattered. The information existed inside the organization. The assessment built on top of it did not move until the market had already moved past it.

Peloton, 2021 to 2022. Peloton's connected fitness sales grew an average of 168% year over year across the four quarters through March 2021, and the company built its plans around that rate continuing, committing to a $400 million US manufacturing facility that May. Growth was already falling by the time that announcement landed, down to 54% the following quarter and then 6%, but the investment case behind the factory went unrevisited. By January 2022, Peloton was pausing Bike and treadmill production, shelving the factory, and had recorded roughly $182 million in inventory impairments.

It is the same shape as the Silicon Valley Bank story that opened this guide. The evidence needed to catch the drift already existed somewhere inside the organization. What was missing was a process that forced the standing judgment to confront it before the gap became the story.

CPF treats every published assessment as a snapshot of a body of work that stays active, which is what keeps drift from setting in unnoticed. Sources, assumptions, alternative futures, indicators, forecast questions, probabilities, reviews, and outcomes remain connected after publication, so a change in one part can be traced through the rest of the work.

What Continuous Probabilistic Foresight is

Continuous Probabilistic Foresight is a disciplined process for building and maintaining an evidence-based view of how an important issue may develop, including what is uncertain and how likely its different outcomes currently appear.

Continuous means the analysis stays active after it first publishes. Evidence gets refreshed, estimates get updated, assumptions get monitored, and the question and its working model get reconsidered once reality stops fitting them.

Probabilistic means material uncertainty gets stated on purpose. A probability can describe whether an event will occur, which future is gaining ground, or whether a proposition is true. It does not mean every sentence needs a number, and a precise-looking estimate is not automatically a reliable one.

Foresight means the analysis holds more than one plausible future in view at once, and identifies what would make each one more or less likely. CPF is a method for noticing change earlier, understanding the alternatives, and knowing when the current view needs to change.

Likelihood and confidence answer different questions

CPF keeps two questions separate that get blurred together constantly, the same distinction the US intelligence community's analytic standards require analysts to draw. Likelihood asks how probable the judgment or outcome is. Analytical confidence asks how strong the evidence and reasoning behind that judgment are.

An analyst can assess a 70% chance of an outcome and hold low confidence in it, because the evidence is thin and conditions are moving quickly. A different judgment can sit at 55% and carry high confidence, because strong evidence consistently points to a genuinely close contest. The probability states current belief. Confidence describes the quality of the evidence and reasoning behind it - information gaps, reliance on assumptions, the strength of the working model, and how much disagreement remains. Sensitivity, how easily the estimate could move, is a related but separate question again.

The traditions CPF brings together

CPF draws its method from several professional traditions built up over the past seventy years. Each one solves a different part of the problem, and none of them supplies the whole continuous method by itself.

CPF's distinctive contribution is the connective tissue between these practices, maintained over time rather than assembled once. An assumption can lead to an indicator. An indicator can become a forecast question. A forecast can supply evidence for changing a scenario's probability. A material change can reopen the whole assessment, and an eventual outcome can show whether a forecast held up or whether the issue was framed too narrowly from the start.

Starting conditions get rechecked as the issue develops. The lifecycle below assumes a meaningful issue and enough evidence to begin. As the issue develops, CPF keeps checking whether the question still matters and can still be answered, whether the evidence supports a defensible judgment, and whether the working model still covers the ways events could plausibly unfold. When one of those conditions fails, the method changes, pauses, or retires the analysis.

The twelve-step CPF lifecycle

CPF follows twelve logical steps. Every analysis uses the same underlying process, and the depth given to each step should track the question, the time horizon, the available evidence, and the cost of being wrong.

One method, several levels of effort. A direct forecast can move through the first nine steps quickly. A high-stakes geopolitical assessment can call for several ways of decomposing the issue, multiple scenario sets, independent forecasters, and a formal review. The logic stays the same either way: a clear question and scope, traceable evidence, recorded assumptions, answerable forecast questions, uncertainty stated with care, meaningful challenge, and regular updating.

Step 1: Define the issue

State what the analysis is about without forcing it to serve one particular decision. The question and its scope should be broad enough to reveal genuine alternatives, and bounded enough to guide research.

• State the question or topic in neutral language.
• Set the scope, the time horizon, and the relevant definitions.
• Identify the intended audience and use, when they are known.
• Clarify what would make the analysis useful, and what kind of error would matter most.

Challenge at this stage: ask whether the question already contains its answer, favors one actor's point of view, or excludes an entire kind of outcome.

Running example. This guide carries one question through every step below: "How is Russia likely to use or threaten military force beyond its internationally recognized borders during the first three years after Vladimir Putin no longer exercises effective control of the Russian state?" It states a topic rather than a policy position, and sets an explicit time horizon triggered on a condition: loss of effective control.

Step 2: Build the evidence base and map the terrain

Research begins before conclusions get drafted. The first goal is understanding the terrain: the actors, the events, the live disagreements, the gaps in current knowledge, and the sources available.

• Collect relevant internal and external sources.
• Evaluate source relevance, quality, and independence.
• Identify known unknowns and unresolved disagreements.
• Keep a record of the evidence base so the analysis can be reconstructed later.

The cold-start rule: the first research output maps the terrain; a bottom line comes only after that. An early conclusion becomes a powerful anchor, especially once later work gets generated by AI.

Example. The terrain here spans force posture and nuclear custody, the informal rules governing elite succession in Russia, precedent from past transitions in personalist systems, and the areas where open reporting is thinnest, such as loyalty networks inside the security services. None of this yet produces a view on what a successor would do.

Step 3: Decompose the issue

Decomposition breaks a broad issue into the smaller forces and relationships that could shape how it develops, turning an overwhelming topic into a working model that can be examined and updated piece by piece.

• Identify the drivers, actors, incentives, constraints, and capabilities at work.
• Describe how the drivers relate to one another.
• Distinguish structural forces from short-lived developments.
• Keep more than one decomposition when different ways of organizing the issue would lead to meaningfully different analysis.

Example. A decomposition here might trace how elite cohesion, control of nuclear forces, the state of the military after prior campaigns, economic strain, and Western alliance posture interact with one another.

Step 4: Surface assumptions and uncertainties

Make explicit what the analysis is treating as true, what it expects to persist, and where uncertainty is doing the most work.

• Write the consequential assumptions as clear propositions.
• Record the evidence behind each one and the analytical confidence it carries.
• Show which judgments, drivers, or futures depend on it.
• Define what would weaken or disprove it, who is watching it, and how often it gets reviewed.

An assumption does not retire once it becomes forecastable. It stays in the analytical record as an assumption, and can also link to an indicator, a collection requirement, or a forecast question at the same time.

Example. One assumption worth stating explicitly: nuclear command and control stays intact through a transition. Another: a successor's early moves say more about consolidating domestic power than about intentions toward any particular border.

Step 5: Represent alternative futures

Describe the main ways the future could unfold, so a single expected future does not become the unstated default. This can range from two clear outcomes to several formal, fully worked scenarios.

• Build alternatives that differ in kind.
• Keep them consistent in scope, level of detail, and time horizon.
• When probabilities get assigned across scenarios, make the set mutually exclusive and collectively exhaustive: the scenarios should not overlap, and together they should cover the full range of outcomes, with an explicit other-outcomes category added where needed.
• Give every probabilistic scenario observable criteria for occurrence, non-occurrence, or invalidation at the stated horizon.
• When scenario narratives are meant to overlap, assess how plausible or well supported each one is instead of treating them as shares of a single probability distribution.

The scenario set is a model that gets retested as the issue develops. Its structure, and the rule for deciding what occurred, changes when the current alternatives stop covering the outcomes that matter. Evidence like that is a reason to rebuild the set.

Example. A scenario set here might include an emboldened successor testing NATO's resolve directly, a besieged successor turning outward to unify a fractured elite, an internal struggle for control that leaves little capacity for external action, and continuity under a chosen heir. Each implies a different posture toward the border, which is what makes them different in kind.

Step 6: Choose useful indicators

Determine which observable developments would separate the alternatives, test an important assumption, or provide warning that the outlook is shifting.

• Favor indicators that clearly separate one future from another and provide useful lead time.
• Weigh whether an indicator can be observed and assessed clearly, what it costs to track, and how often it produces a false alarm.
• Remove redundant indicators that respond to the same underlying change.
• Forecast selectively. More questions do not automatically produce more insight.

Review the working model before forecasting. Before forecasting begins, test the decomposition, the assumptions, the alternatives, and the indicators as one system. Ask what future might be missing, whether the analysis favors one point of view, whether the indicators genuinely separate the alternatives, and whether a different decomposition would produce meaningfully different questions.

Example. Candidate indicators include changes in nuclear forces' alert status, redeployment from active theaters, public statements from regional military commanders, and the pace of departures or purges among senior officials. Each gets tested against whether it would also hold true under more than one of the four scenarios above.

Step 7: Write forecast questions that can be answered

Convert the selected uncertainties into questions that can be estimated, updated, and eventually settled by evidence.

• Define the outcome space and the resolution date.
• Write objective resolution criteria - clear rules for deciding the answer - and name the sources that will be used.
• Specify thresholds, edge cases, and cancellation or void rules.
• Link each question to the driver, assumption, indicator, or alternative future it informs.

Example. "Will Russia conduct a cross-border military strike against a NATO member state within 12 months of a confirmed transition in effective control?" is answerable. "Will Russia become more aggressive?" is not.

Step 8: Collect and combine forecasts

Collect independent probability estimates from qualified contributors - people, computational models, or both - and combine them in a way that can be explained afterward.

• Preserve every individual estimate, its rationale, and its update history.
• When people and models both contribute, collect their initial judgments independently before comparison, to limit anchoring.
• Show the distribution and the disagreement alongside the consensus number.
• Record the evidence behind each estimate, the contributor or model version, any prompt used, and the method used to combine them.
• Weight a contributor's track record only once enough past results exist to support doing so.

An AI-generated rationale is a reasoning artifact to inspect. The evidence itself has to stay traceable to the underlying sources.

Example. A question like this benefits from combining regional specialists, military analysts, and computational models, since no single vantage point covers succession politics, force posture, and alliance dynamics equally well.

Step 9: Explain the current outlook

Bring the evidence, the drivers, the assumptions, the alternatives, the indicators, and the forecasts together into a clear account of where things currently stand.

• State the major judgments and their likelihoods.
• Explain analytical confidence separately from likelihood.
• Represent the supporting evidence, the contrary evidence, the gaps, the alternatives, and any dissent.
• Name the indicators and developments that would change the assessment.
• Say plainly when the available evidence cannot support a defensible estimate, and identify what evidence would make one possible.
• Explain the implications without being forced into a recommendation.

Example. The resulting account states the likelihood of a strike against a NATO member separately from the likelihood of action against a non-NATO neighbor, explains the confidence behind each, and names what would change the picture, such as a confirmed successor's first six months of public statements and troop movements.

Step 10: Test the whole argument before publishing it

Every earlier step has already been challenged individually. This step asks a broader question: does the complete argument hold together? The review needs enough distance from the original work to test the argument rather than defend it.

• Test whether the forecasts and the other evidence support the scenario probabilities and the key judgments.
• Check for probability inconsistencies, assumptions that shifted mid-argument, and alternatives that got left out.
• Examine whether contrary evidence was acknowledged and then effectively ignored.
• Record reviewer independence, the responses given, unresolved dissent, and what changed as a result.
• Approve a dated snapshot for publication while leaving the underlying analysis active.

Example. A review here checks whether a low near-term probability against a NATO member follows from the evidence on nuclear custody and alliance deterrence, or whether it assumes deterrence holds without testing that assumption directly.

Step 11: Monitor and update continuously

Once the first assessment publishes, monitoring continues inside the current question and working model. The analysis stays active, and each published version is a dated snapshot of it.

• Refresh the evidence base, and monitor the indicators and the assumptions.
• Update forecast probabilities and, where a valid scenario distribution exists, scenario probabilities.
• Explain every material change when it happens.
• Record scheduled review cycles, even the ones where no material change is warranted.
• Trigger a targeted re-challenge and a revised assessment once a threshold is crossed.

Example. The assessment sits largely dormant until a transition begins, then updates quickly as succession events unfold: named indicators get checked daily, and probabilities move as reporting on the new leadership's posture accumulates.

Step 12: Rethink, resolve, retire, and learn

CPF has to improve both the current analysis and the method that produced it. This step covers what happens once the question or working model no longer fits, once the issue can no longer be assessed usefully, or once an outcome finally becomes available.

• Reopen the drivers, the assumptions, the alternatives, and the indicators once new evidence stops fitting the working model.
• Rebuild the scenario set once it no longer covers the plausible outcomes, while preserving the earlier version and the reasons for the change.
• Rewrite the question once it no longer captures the main uncertainty, or can no longer be answered as written.
• Suspend an assessment when the issue still matters but the present evidence cannot support a defensible estimate, and state what would let it resume.
• Retire an assessment once it has resolved, been overtaken by events, lost relevance, or stopped producing useful learning.
• Resolve forecast questions and scenarios against the outcomes that occurred.
• Compare forecast accuracy, calibration, warning time, and the ability to separate outcomes against simple benchmarks.
• Use the misses and the false alarms to improve how issues get defined, how questions get written, how forecasts get combined, and, where relevant, how models and prompts get designed.

Example. If effective control changes hands through a negotiated arrangement none of the four scenarios anticipated, that is a reason to rebuild the set.

When reality no longer fits

CPF responds at the lowest level that restores a defensible analysis.

Challenge is continuous and layered

CPF spreads challenge across four distinct moments, rather than concentrating it into one red-team meeting held after the intellectual work is already finished.

• Checks at each stage. Every part gets tested before later work depends on it: sources for coverage, assumptions for support, scenarios for structure, indicators for how well they separate the alternatives, and questions for whether they can be answered clearly.
• Review before forecasting. The working model gets tested before the team invests time and credibility producing forecasts against it.
• Independent review of the full argument. The complete argument gets tested for consistency before it becomes the approved snapshot.
• Review triggered by change. Important new evidence reopens the judgments it affects. Evidence that no longer fits the working model triggers a broader rethink.

Not every check is red teaming. Validation asks whether a piece of work meets a stated standard. Challenge asks what might be wrong or missing. Peer review examines the evidence and the method. Red teaming applies a genuinely independent and adversarial point of view to the argument itself. Keeping these apart matters, because it stops a routine or automated check from getting presented as independent challenge.

CPF requires meaningful challenge without prescribing who provides it. Depending on the stakes and the capacity available, a check can be performed by the original analyst, a team, an independent reviewer, an automated system, or some combination. Human and independent review earn their cost most clearly when consequences are high, the evidence is ambiguous, or the assessment leans heavily on judgment. Either way, the analysis should state plainly what kind of review happened.

Four loops that keep the analysis honest

The evidence loop: update the current view

The evidence loop asks what the newest information does to the current working model. It refreshes research, forecasts, assumptions, indicators, and, where appropriate, scenario probabilities. Small changes stay in the history. A significant change needs an explanation, a review, and a revised assessment.

The scheduled review loop: inspect the model as a system

Not every weakness announces itself through a crossed threshold. On a regular schedule, the analysis gets reviewed for aging evidence, weakening assumptions, duplicated indicators, connections nobody can support, gaps opening in the scenarios, and disagreement that keeps growing. A review can conclude that nothing important needs to change. Human review earns its keep here, though CPF does not prescribe who performs the check.

The rethink loop: ask whether the question and model still fit

The rethink loop asks whether the current question and working model are still the right ones. A scenario set can be internally consistent and still miss the future that is emerging. A failed assumption, growing evidence that none of the current alternatives fits, or a newly important cause are all reasons to revisit the decomposition and rebuild the scenarios. The practical question is simple: is this still the right problem, analyzed the right way?

The outcomes loop: learn from what happened

The outcomes loop resolves questions and measures performance. It distinguishes a genuinely good forecast from a persuasive explanation written after the fact. Scoring should show whether the stated probabilities matched how often events occurred - the property known as calibration. Comparison against simple benchmarks, plus an honest review of the successes and the failures, should feed back into changes to the method.

Why all four matter. The evidence loop keeps the estimates current. The scheduled review loop catches quiet deterioration. The rethink loop keeps the question and the working model honest. The outcomes loop keeps the method itself accountable.

Common failure modes

The CPF assurance profile

A reader taking on someone else's analysis needs a fast way to see whether the method got followed, and where the resulting judgment stays vulnerable. Build the record so that question can be answered without reopening the whole file: call it a CPF assurance profile, or think of it as an analytical nutrition label. It works whether it lives in a single memo, a shared document, or a dedicated review, and it does not depend on any particular piece of software.

Assurance describes discipline, and it has limits. A strong profile shows whether the team followed the method, how healthy the analysis is right now, and what performance evidence exists from past outcomes. Disciplined analysis can still miss, and a weak process can occasionally get lucky, so the profile cannot guarantee a judgment is accurate.

Did the team follow the method?

• Clear question: stated neutrally, with a defined scope and time horizon.
• Grounded: major claims trace to an evaluated evidence base.
• Decomposed: the decomposition explains how the forces shaping the issue relate to and act on each other.
• Assumptions visible: the assumptions the analysis depends on are supported, linked, and monitored.
• Alternatives complete: where scenario probabilities are used, the scenarios avoid overlap and together cover the outcomes.
• Indicators useful: the selected indicators separate the alternatives or test an important assumption.
• Questions answerable: forecast questions and probabilistic scenarios can be settled using stated criteria and named sources.
• Uncertainty clear: the major uncertain judgments are stated clearly and supported by reasoning.
• Confidence explained: analytical confidence is stated separately from likelihood, and explained.
• Challenged: the individual stages, the working model, and the full argument were each tested at the right moment.
• Traceable: important changes, disagreement, and the review history are preserved.

How healthy is the analysis right now?

• Freshness: when were the evidence, the assumptions, the probabilities, and the scope last reviewed?
• Coverage: do the current scenarios still cover the plausible outcomes?
• Assumption strain: which important assumptions are weakening or disputed?
• Evidence sufficiency: does the available evidence support a defensible estimate now?
• Sensitivity: which evidence or assumption could move the assessment the most?
• Dissent: what important disagreement remains unresolved?
• Status: is the analysis active, suspended, resolving, or retired, and when is the next review due?

What performance evidence exists?

Once enough outcomes accumulate, the profile can summarize forecast accuracy, calibration, warning time, how well the resolution process held up, how well the forecasts separated the outcomes, and performance against simple benchmarks. Where no track record exists yet, say so directly. An absent track record is neither a point in the method's favor nor against it, and turning its absence into a score either way misrepresents what is known.

What CPF must prove

CPF draws on decades of established practice in forecasting, structured analysis, scenario development, warning, assumption testing, and learning from outcomes. How CPF brings those practices together into one continuous method is newer, and still being proven out. Its credibility should rest on transparent evidence of how it performs.

The test is practical. Can its questions and its scenarios be answered clearly? Are its forecasts accurate and well calibrated over time? Does it beat simple benchmarks, provide useful warning, leave a record other people can reconstruct, and improve after a failure? Some of what it offers - clearer assumptions, more visible disagreement - is a quality of the process, and should get assessed on its own terms rather than through whether one particular forecast happened to land. CPF's discipline is a way to stay less wrong for less time, and to know why the view changed when it did.

Run enough analysis this way, and the operating question stops being whether last quarter's report still holds up. It becomes whether the view is still current, and whether you can show exactly why it changed.

This guide was written with the help of AI, but was reviewed and edited by a human.