Plan and decision scoring
Achievability and Fragility
How likely the plan is to hold, and how hard you would have to push it.
Achievability scores a plan before you commit to it: replay its commitments against every recorded error pattern on file, and Achievability is the share under which it would hold. It runs from zero to one, higher is better, and it is read off the record rather than assumed: a rank in a sorted list, which at decision time costs a lookup rather than a solve.
Fragility is the rate at which that number falls as the recorded error is scaled up. It exists because Achievability alone cannot separate two plans that score the same today. One of them may give up eleven points when errors grow by half; the other may give up fifty-eight, and a scorecard reading only the first number ties them.
Both are built on the same machinery as conformal prediction, carried up from a single quantity to a whole plan. And both inherit its honesty condition: distribution-free is not assumption-free.
One Question, Two Scopes
The same quantity asked of one decision and of a whole plan. They are not the same number, and the second is not an average of the first.
Per decision
Decision Achievability
The calibrated probability that a single agent decision delivers the outcome it predicted. It is field four of the Inspect contract, it carries on the audit row beside the prompt, the decision and the expected outcome, and it is the axis the Decision Stream triages on. Below the threshold, the agent informs instead of acting.
Per plan
Plan Achievability
The probability that a whole plan would hold every commitment it makes at once: the fill rate, the capacity ceiling, the customer date, the cash position. Every one has to survive. This is a joint question, so the answer is lower than any single commitment's own probability, and often much lower.
Plan Achievability is not computed by combining decision scores. It is measured directly, against error patterns that occurred together, so the correlation between them survives into the answer.
Why One Number Is Not Enough
Two plans, both scoring 0.92 against the error the business actually makes. Then ask what happens if the errors get larger.
Illustration. The figures are invented to show the shape, not measured. Plan A gives up eleven points by the time the error is half again as large; Plan B gives up fifty-eight and is below any tolerance a business would set.
There is a second reading of the same curve that planners argue with more easily, and it is the one we prefer on a screen: the multiplier at which the plan drops below tolerance. Plan A survives errors about eighty per cent larger than the business has been making. Plan B survives about thirty per cent larger. That is a sentence a planner can take into a supplier review, because it is denominated in something they have an opinion about.
How It Works
Four steps, no sampling, and no distribution anyone had to choose.
Keep the Errors Together
One row per period, holding every quantity's realised error as it actually occurred. Demand running hot in the same week a supplier ran late is one row, not two independent draws.
Apply Each Row to the Plan
Ask of each stored row whether the plan would have held every commitment it made under that pattern of error. One pass, every row used exactly once.
Count, Do Not Simulate
The share that survived is the Achievability. Nothing is sampled and no random number is drawn, so the answer is identical every time it is computed and reproducible from the stored record months later.
Scale, and Re-count
Repeat at larger error, a few multipliers out. The curve through those points is the Fragility, and it costs about as much as computing the first number a handful of times.
This is not Monte Carlo, and the distinction is load-bearing
We do not Monte Carlo a plan, and nothing here changes that. Sampling futures at plan time produces a tight distribution over your own assumptions, and its tightness is misleading. This counts what actually happened instead, which is why it is reproducible and why it can be audited long after the plan was executed.
Scaling the errors is the one place that claim needs care, because a multiplier does manufacture error magnitudes the business has not seen. It is a dial, not a forecast. The curve says what would happen to the plan if errors were half again as large, and it is silent on how likely that is. We do not attach a likelihood to a disruption scenario, and a stress multiplier is not one.
Fragile, Resilient, Antifragile: a Different Level
Gartner's supply chain practice sorts supply chains into three states, taking the term from Taleb's Antifragile (2012): fragile loses value when exposed to uncertainty, resilient holds its value, and antifragile gains from it. Many readers now arrive carrying that frame, so it is worth being exact about where these two numbers sit inside it.
They sit at a different level. Gartner's states describe an organisation. Achievability and Fragility describe one plan. A resilient organisation can hold a fragile plan, and most do, routinely, with no way of knowing. That gap is most of the reason a per-plan measure is worth building: an organisational diagnosis does not reach down to the thing anyone actually executes.
No plan can be antifragile
Antifragility is gain from disorder: a positive response to stress. For a fixed plan against fixed commitments, the slope cannot be positive. Scaling errors up cannot raise the share of patterns under which every commitment holds, because a pattern that breached a threshold at a smaller scale still breaches it at a larger one. The best a plan can manage is a flat line. An antifragile plan is not difficult, it is unavailable.
The loop can be, and it is measurable
What can gain from disorder is the loop that replaces plans. Hold the stress multiplier fixed and plot Achievability across successive planning cycles. If the line rises, the system is producing more achievable plans under the same stress than it did last cycle, which is gain from disorder in a sense Taleb would accept and a CSCO can audit. That is a falsifiable claim rather than a posture.
Gartner publishes figures for how many supply chains sit in each state. We do not quote them, because they come from a survey of what respondents expected rather than of what was measured.
What These Numbers Are Not
Four limits that belong on the same page as the claim, because each of them is a way the number can mislead someone who trusts it.
Not finer than the record
Achievability is quantised by how many patterns it was computed from. With two hundred stored patterns it cannot resolve below about half a percentage point, so a screen showing 0.996 from a two-hundred-row record is inventing precision. The number travels with its row count, and the interface refuses digits it has not earned.
Not valid through a regime change
The guarantee rests on exchangeability: that the recorded past and the thing being predicted come from the same process. A rewritten tariff regime or a new competitor violates that, and no amount of calibration reaches it. What the measure can do is report the day it stopped holding, by checking its own promises against outcomes.
Not safe to pick on blind
Choosing the highest-scoring plan from a set biases that score upward, because part of what made it win is that it happened to suit the stored patterns. The chosen plan is re-scored against a held-out slice of the record before it is committed. That check costs nothing to run and is easy to skip, so it is a gate rather than a report.
Not a probability of a shock
The stress multiplier is a sensitivity of the plan, not a forecast of the world. Achievability at 1.5x says what the plan would do in a rougher environment. It says nothing about how likely that environment is, and the moment someone writes a percentage next to the multiplier, the instrument has been misused.
See how a plan gets scored
Achievability and Fragility sit on top of the same calibration layer that governs every agent decision.