Using Regret Minimization Framework to enhance PDSA Cycle

What each one is actually for

Regret Minimization Framework

The Regret Minimization Framework is a filter for one specific kind of stuck. You have two options, you have been round the arguments enough times that no new information is arriving, and both cases still look about equally good. The framework refuses to arbitrate on those terms. Instead you move the seat: picture yourself at eighty with the whole story finished, look back at this fork, and ask which way you would wish it had gone. What that does is change which considerations carry weight — this quarter's discomfort shrinks, the shape of a life gets bigger — and for many people the deadlock breaks within a few minutes. Two things it does not give you, and both matter here. It does not tell you how to do the thing. And it produces a verdict with no expiry date and no review, which is a strange property for something you are going to spend the next decade acting on.

PDSA Cycle

The PDSA Cycle — Plan, Do, Study, Act — comes out of the quality-improvement tradition of Walter Shewhart and W. Edwards Deming, and its whole discipline lives in one instruction most people skip: write down what you expect to happen before you run the change. Plan means specifying the small change and the prediction together. Do means running it at genuinely small scale, one week, one team, one instance, and recording what happened including the parts you did not plan for. Study means putting the real result next to your written prediction and looking at the difference — Deming argued for the word study over the older word check precisely because checking implies inspecting whether things went to plan, while studying means learning from the gap when they did not. Act means one of three choices: adopt the change, adapt it and run another cycle, or abandon it. Then you go round again with what you just learned. It is deliberately small and deliberately fast, because the point is to be wrong cheaply and often rather than right once.

What you get out of pairing them

The thing you get here is permission to be wrong about the plan without reopening the decision. Most people who make a big directional call and then hit a rough first year face a single undifferentiated question — was this whole thing a mistake — and answer it with their mood. Separating the two questions gives you somewhere else to put the trouble: the direction can hold while the second version of the plan replaces the first. What the long view adds on its own is the end of deliberation, which is worth more than it sounds when a decision has been eating a year of your attention. What the improvement cycle adds is the ability to find out, in writing, that your expectations were off by a factor of three, in month one rather than year three, while the cost of adjusting is still low. Together they cover the two ways this typically goes wrong: carefully optimising something you never checked was worth doing, and abandoning something worth doing because the first attempt at it was clumsy.

What Regret Minimization Framework rests on

• Self distancing improves judgment randomised trials
  Evaluating a situation from a distanced, third-person vantage point instead of an immersed first-person one improves the quality of judgment about it.
• Long term framing surfaces values practitioner report
  Framing a decision against a long time horizon surfaces which option better matches a person's deeper values and cuts through short-term ambiguity.

What PDSA Cycle rests on

• Feedback against prediction corrects belief plausible mechanism
  Comparing the actual result of a change against the predicted result before adopting it corrects mistaken beliefs faster than adopting or rejecting changes based on intuition alone.
• Small scale testing precedes scaling observational studies
  Running a change through a small planned test cycle before wider adoption reveals real-world effects while limiting the risk of being wrong.

Where they actually meet

Both practices work by dragging something into the open that would otherwise stay implicit, and then treating the exposed thing as a question rather than a fact. The long-horizon question exposes a ranking of values you have been carrying without stating: you thought you were weighing salary against commute, and what surfaces from the eighty-year-old's seat is that neither was ever the real term. Writing a prediction before running a change exposes a forecast you were already relying on: you thought you were just trying something, and the moment you write "three of the eight will reply within a week" you can be shown to be wrong, which is precisely the property that makes learning possible. The seam is what each leaves untouched. The regret question answers should-I and then stops, leaving the whole of the execution as an untested assumption sitting quietly behind the decision. The improvement cycle answers is-this-working with real rigour and is completely indifferent to whether the thing being improved was ever worth doing — a well-run cycle will optimise a pointless process very efficiently. What the regret question leaves as an assumption is exactly what the cycle turns into something you can look at directly: not a decision that gets treated as permanently right or silently abandoned, but a specific prediction about a specific version, checked against what actually happened.

The principle underneath both

A commitment you never wrote down cannot be corrected. Held vaguely, both a values ranking and a plan behave the same way: they quietly reshape themselves to fit whatever happened, so you can never be shown to have been wrong and therefore never learn anything. Stating either one in advance, in words specific enough to fail, is what makes it correctable. Deming's tradition applies that to a process change and the regret question applies it to a life direction, and neither of them thought of it first. The uncomfortable corollary is that the value of both comes from their ability to embarrass you later. A prediction you cannot fail and a verdict you never revisit are equally useless, for the same structural reason.

What that makes the real goal

A big commitment that stays correctable is what both halves are aiming at: direction and current plan kept as two separate written claims, each specific enough to be checked, and neither allowed to hide behind the other. Put that way it detaches from both traditions and can be built with whatever tools you like. Nobody needs a quality-improvement vocabulary to write "here is what I expect the first three months to produce, and here is the date I will look" on a page. The thing to protect against is the two failure modes this arrangement is aimed at: a direction that survives on sunk cost because it was never re-examined, and a plan revised so continuously that no version of it was ever actually tested.

Other practices that serve the same goal

A decision journal is the cheapest version and it has no trials behind it at all, only a long practitioner track record: at the moment you decide, write the decision, what you expect it to produce, what would tell you it is going badly, and a date to reread the page. The reread is the entire mechanism, and it works because memory of your own past expectations is far more flexible than the page is. A premortem goes at it from the other end: before you start, assume it is a year later and the thing has failed, and write the story of how. Experimental work on imagined hindsight suggests that assuming an outcome has already happened helps people generate more, and more specific, reasons for it than asking them to forecast risks in the usual way. And mental contrasting paired with if-then planning, tested under the name WOOP in a number of randomised studies, does something adjacent for personal goals: you picture the outcome you want, then name the concrete internal obstacle in your way, then write an if-then plan for it. Where a PDSA cycle checks a prediction about the world, that one checks a prediction about yourself.

A tradition that already does this

The improvement tradition worked out this exact problem decades ago and built the answer into its own front end. In the Model for Improvement, the framework PDSA cycles are usually run inside, three questions come before the first cycle turns, and the very first is what are we trying to accomplish. Only then does it ask how you will know a change is an improvement, and what change to try. That first question is doing the job the regret question does here — checking that the aim is worth the effort of testing — and the tradition put it there because teams kept running immaculate cycles at targets nobody had justified. It is worth noting what does not carry over: an improvement aim is usually set by a group against an organisational purpose, not by one person against their own life, and the regret question has no equivalent of a shared aim statement other people can argue with.

Using the Regret Minimization Framework to enhance the PDSA Cycle

A cycle can run beautifully and still be testing the wrong thing: small changes, honest measurement, tidy comparisons, all pointed at a goal nobody ever stepped back to check. The move goes before the Plan step, and it takes five minutes. Take the aim the cycle is serving — not the change, the aim — and ask the long-horizon question about the aim itself: at eighty, looking back, would you regret never having pursued this at all. The answer sorts your cycles into two piles fast. Some aims survive it and you run the cycle with more appetite for the tedious measurement part. Others do not, and what usually surfaces is that the aim was chosen because it was easy to plan and easy to measure rather than because it mattered, which is the failure mode measurable work is prone to. One warning about scope. Do not run this question on every cycle; on a two-week test of a new invoicing step it produces a shrug and wastes the frame. Reserve it for the aim above a series of cycles — the thing you will still be running tests against in a year — and ask it once, when the aim is set, and again if the aim drifts.

Using the PDSA Cycle to enhance the Regret Minimization Framework

The regret question produces a clear go and then leaves you alone with the hardest part. You decided from the eighty-year-old's seat that leaving was right; nothing in that verdict tells you whether the specific way you left, or the specific first step you took afterwards, produced anything like what you were picturing. Break the execution into cycles instead of enacting it whole. Take the first concrete move — the two weeks' notice, the first three client conversations, the first month of the new routine — and before you do it, write one sentence saying what you expect it to produce, in terms specific enough to be wrong: not "it will go well" but "two of the five will want a second conversation, and I will still have four months of runway at the end of the month." Do it. Then sit down on a date you set in advance, put the real numbers next to the sentence, and study the gap: which part of the forecast was off, and by how much, and what does that suggest about the next version. Then adopt, adapt, or abandon that step — that step, not the decision. Keep each cycle small enough that being wrong costs you weeks rather than years, and keep the original verdict out of the review entirely. It is not what is being tested.

You quit in March, and by June you cannot tell if it is working

You circled the question of leaving a stable job for your own thing for about eighteen months, and in February you finally did the exercise properly: eighty years old, looking back, would you regret never having tried. The answer came back clearly, you would, and three weeks later you gave notice. Now it is June. Two clients, less money than you modelled, and every Sunday evening you relitigate the whole decision from scratch — which is exhausting and produces nothing, because the thing you are re-arguing was settled in February and the actual question is about June. So you stop asking it. You take one thing instead: the outreach approach you have been using since April. You write down what you expect the next four weeks of it to produce, in a form that can be wrong — eight conversations, two proposals, one signed — and you put a date in the calendar for the fifth of July. On the fifth you sit down with the real count: eleven conversations, one proposal, nothing signed. That gap is informative in a way your Sunday evenings never were. Plenty of interest, and something breaking between conversation and proposal. You adapt that step and run it again with a different second call. The February verdict is untouched by all of this, which is exactly as it should be: it was never a claim about your outreach.

When this is worth reaching for

Six months after the call, and you still cannot say whether it is working — only how you feel about it on a given Sunday. That is the condition where this arrangement does the most, because the missing thing is not resolve but a written expectation you can check the world against. It also fits the earlier moment, where a decision is settled in principle and the first concrete step is about to be taken on faith that the first plan is the right one. Two situations where it is genuinely too early. Before the decision is made, there is nothing to run a cycle on — a cycle tests a version of a plan, and you do not have one. And in the first fortnight after a major change, most of what you are measuring is disruption rather than signal; give it long enough that a normal week exists before you start writing predictions about normal weeks.

Where this does not apply

The long-horizon question works on decisions big and directional enough to make sense from an eighty-year vantage point — a career change, whether to have a child, whether to stay in a city — and applying it to which vendor to trial this quarter mostly produces a shrug rather than clarity. The cycle works on changes concrete enough to plan, run small, and measure against a prediction, and applying it to something as large and irreversible as whether to have a child does not fit either, because there is no small-scale version of that to test before committing. The pairing fits a decision that is both big enough to warrant the long view and specific enough, once made, to break into testable pieces; a decision that is only one of those two gets little from adding the other. There is also a failure mode worth naming, because it is common in people who like measurement: running cycles on a decision you have not actually made yet, which looks like diligence and functions as delay.

What the evidence does and does not support

The two halves rest on different kinds of support, and neither is a trial of what this page describes. Judging a situation from an outside vantage rather than from inside it has randomised support for improving judgment quality; the narrower claim that a long time horizon reliably surfaces what someone deeply values is closer to practitioner observation and a well-told business story than to a body of controlled studies. On the improvement side, the central claim — that checking a real result against a stated prediction corrects mistaken beliefs faster than intuition does — is a reasoned, mechanistic case rather than a trial result, and the case for small-scale testing before wider adoption comes mainly from observational work and decades of accumulated practice in manufacturing and healthcare quality programmes. The sequence proposed here has not been studied. It would be awkward to study: you would need to randomise people to run improvement cycles on personally significant decisions or not, and then agree in advance on what a good outcome even looks like years later, which is a definitional problem before it is a measurement one. Read the fit between them as a reasoned division of labour between a should-I tool and a did-it-work tool, not as a finding.