The confidence figure attached to a completion date is the probability that one path finishes on time, not that the project does — and every parallel stream added to save time quietly reduces the true number without changing the reported one.
A delivery director inherits a date that cannot move. She does what competent directors do: splits the remaining work into concurrent streams, brings in a second contractor for one of them, and overlaps two stages planned in sequence. The forecast finish comes back earlier, the stated confidence in the committed date comes back higher, and the board is reassured.
Every one of those moves has made the project less likely to finish on time than the number now says. Not because the plan is worse, but because the number never measured the project. It measures one chain of activities inside it, and each intervention added chains it does not see.
That figure does more work than almost any other number in a delivery estimate. It sizes contingency, decides whether a date is offered as a commitment or a forecast, gates capital release, and determines which programme is defended when there is not enough of anything to go round. A board that believes it has the probability of an outcome, when it has the probability of a component, is mispricing the enterprise's exposure.
The mechanism is neither obscure nor hidden. Durations are set to their expected values, the longest resulting chain is designated critical, uncertainty is aggregated along that chain alone, and the spread becomes a probability of finishing by a chosen date. Everything off that chain is treated as certain to be ready. It never was.
The Strategic Context
Two things have changed since this method became the standard answer to a customer asking how confident you are. The first is that enterprises now buy time with concurrency as a matter of routine. Overlapping stages, splitting scope across suppliers, running qualification alongside production readiness: these are the accepted moves when a date is fixed and scope is not. Each converts a single chain into several. The method's blind spot is not an edge case; it is the normal condition.
The second is that these figures now aggregate. A portfolio office collects a confidence level from each delivery unit and reports a portfolio position, though each contributing figure was computed on a different network, with a different number of unseen paths, by a team with different habits. Whether such an aggregation can be arithmetically meaningful at all is not examined here; that is the subject of [Related article: Three Scales, One Word].
The stakes rise with the size of the delivery unit. A project may survive an optimistic figure; it has a sponsor close enough to see the truth. A programme coordinating a dozen interdependent projects has more merge points than any of its components, and a correspondingly larger gap between reported and real. A portfolio built from programme figures compounds that gap twice.
What Leaders Take the Figure to Mean
Four readings are common in senior forums, and each is wrong in a way that changes decisions.
The first is that the figure is about the project. It is the probability that one named sequence completes within the target, conditional on nothing else mattering; every other sequence has been assigned an implicit probability of one.
The second is that it reflects estimating quality. Leaders press for better estimates, tighter ranges and more experienced estimators, all of which improve the inputs and none of which touch this defect. The gap is created by the shape of the network, not by the honesty of the durations fed into it.
The third is that protecting the date harder raises the number. Adding a parallel stream shortens the expected duration and improves the arithmetic on the chain being measured. It also creates a new chain, a new merge point and a new way to be late, none of which enter the calculation. The reported figure improves while the real one falls, which rewards precisely the behaviour that widens the error.
The fourth is that whatever the method omits errs on the safe side. The omissions run one way: unseen paths, merge points and correlated durations each make the true spread wider than the computed one. The method errs not towards caution but towards comfort.
Reframing the Issue
A confidence level is not a statement about a team's realism. It is a statement about a topology: how many independent ways there are to be late, and how many of them the calculation looked at. The governance question changes with it. It stops being whether we believe the estimators and becomes how many paths there were and how many the figure covered — a question with an answer, cheap to obtain, and almost never asked.
It also changes what a rising figure means. A confidence level that improves after a replan which increased concurrency is a warning: the calculation has been re-run on a network with more ways to fail and returned a better result, which is possible only if the new ways to fail went uncounted.
How a stated range then hardens into a single committed date is treated elsewhere in this series and is not reopened here. The concern here is prior to it: what the range was ever a range of.
Three Ways the Reported Figure Departs from the Real One
The paths nobody counted
A network of any size contains many routes from start to finish. Only the longest, calculated on expected durations, is designated critical. Others sit within a reporting period of it, and those near-critical routes often carry more uncertainty than the critical one, because the activities with the widest ranges are frequently the novel ones off the main chain. Once uncertainty is admitted, the designation is provisional: a route shorter on average but far more variable will finish last often enough to matter. The reported figure treats a provisional designation as a permanent fact.
The slack that keeps those routes off the critical path is spent, silently, by whoever touches those activities first. Who is entitled to spend it, under what rule, and how its consumption is recorded is not covered here; that is the subject of [Related article: Float Is the Only Resource With No Owner]. What matters here is that as slack is consumed, the number of routes able to control the finish date rises, and the reported confidence does not move.
The merge points where the streams wait for each other
Wherever two or more streams converge on a successor, that successor waits for the slowest. It cannot start early because one stream was early; it can only start late because one was late. A merge point is an asymmetric filter: it passes lateness through and absorbs earliness.
Splitting work into parallel streams therefore shortens the expected path while multiplying the chances that one stream is late when the others need it. Consider a metro operator renewing signalling on an open line, with track possession, systems integration, driver retraining and station works running as separate streams that must all complete before service testing begins — hypothetical, but recognisable. Each may be well managed. Service testing starts when the last finishes, and the probability that all finish on time is materially lower than the probability that any one does. The standard calculation contains no term for this. It measures one stream and stops.
The correlations assumed away, and the limit quietly passed
Aggregating uncertainty along a chain is valid only when the activities are independent. Real networks are full of common causes: one design authority signing off multiple packages, one specialist trade rotating between streams, one supplier, one scarce test facility. When a common cause moves, several activities move together and the true spread is wider than the sum of the parts. The teaching material of this discipline states this plainly, in the same documents that then present worked probabilities computed as though independence held.
That pattern is the instructive finding. The material is not ignorant of these defects. It names the unseen paths, the merge effect and the correlation problem; it even states a minimum chain length below which its approximation should not be trusted, then reports a probability from a chain shorter than the floor it has just set, in the same document, without comment. That is what happens when a caveat lives in a pitfalls section and a number lives in the worked example, and only one of them travels. Nobody suppresses the caveats. They are printed. They simply do not accompany the figure into the paper that reaches the board.
Decision Framework: The Path-Count Disclosure
No confidence figure should reach a board, investment committee or customer without six items on the same page. This is the path-count disclosure, and it needs no new tooling.
- Path count. How many distinct routes run through the network, and how many sit within one reporting period of the critical route. If unknown, the figure is not reportable.
- Basis path. Which route the figure was computed on, named, with its activity count. Below the count at which the approximation is held valid, it is an indication only.
- Variance off-path. What share of total estimated uncertainty sits on activities excluded from the calculation. A figure computed on a chain carrying a minority of the uncertainty is a figure about a minority of the project.
- Merge exposure. How many points require two or more streams to converge before a successor starts, and how that count has moved since the last baseline.
- Correlation statement. Which common causes were assumed away — named authorities, suppliers, facilities or approvals shared across streams — rather than a generic assertion of independence.
- Recomputation date. When the analysis was last re-run, and whether any concurrency change has been approved since.
Three rules make this binding rather than decorative. If any route is within one reporting period of critical, the position is reported as a range across those routes, not a single number. If concurrency has increased since the last recomputation, the previous figure is withdrawn rather than repeated. And the disclosure is signed by the person accountable for the schedule, not the person accountable for the date; where those are the same person, the enterprise has a larger problem than this one. The adoption test: ask for last month's figure and see whether the six items are on the page.
From Strategy to Execution
Immediate. Take the three largest commitments carrying a stated confidence level and reconstruct the disclosure for each — a day of work per programme. Expect at least one to fail item three outright: the figure will describe a chain carrying a minority of that programme's uncertainty.
Medium term. Make the six items mandatory reporting fields, and change the replanning procedure so any approved increase in concurrency triggers recomputation before the new date is quoted. Fund full-network simulation where the date carries contractual or public consequence, and treat single-path analysis elsewhere as a declared approximation. Consider a diagnostics laboratory network consolidating testing onto a common platform, where validation, accreditation, instrument installation and staff certification run in parallel at each site and converge before go-live: hypothetically, a network whose merge count alone disqualifies a single-path figure from carrying a licence commitment.
Long term. Move the estimating conversation from durations to network shape. The capability worth building is not better three-point estimates but the habit of asking, before approving a plan, how many ways it can be late and which of them the reported number covers. That habit also changes procurement, because adding a supplier to save time becomes visibly a choice to add a merge point.
Signals to Monitor
A confidence figure that has not moved across two reporting periods in which the network changed is not stable; it is stale. One that rises immediately after a replan which added concurrency is a defect report, not good news. A rising count of routes within one reporting period of critical is the clearest early indicator that the reported number is decoupling from reality, and the scheduling tool already holds that count.
Watch too for the disappearance of the caveat: a paper carrying the number but not its basis has performed the same edit the source discipline performs on itself. And watch the reverse signal — a delivery unit reporting lower confidence after adding resources is not underperforming, it is calculating honestly.
Questions for the Leadership Team
- For our three largest commitments, how many routes are within one reporting period of critical, and what share of estimated uncertainty sits on routes the reported figure excludes?
- When was each confidence level last recomputed, and how many concurrency changes have been approved since?
- Which shared resources, authorities or suppliers appear on more than one stream in each major network, and were they treated as independent?
- Where a date is contractually or publicly committed, can we produce the network the supporting probability was computed on, and name who signed it?
- How many merge points exist in each major programme, and has that count risen since the last baseline?
- When we aggregate confidence levels across delivery units, what evidence do we hold that the contributing figures were computed on comparable bases?
Each requires retrieving something the enterprise does not publish. That is the point.
Closing Perspective
The figure is not a lie and the people producing it are not deceiving anyone. It is a correct answer to a narrower question than the board believes it asked, and the narrowing is invisible at every point where it could have been challenged. Whether probability-times-impact scoring elsewhere in the risk function suffers a related confusion between an average and a decision rule is examined separately in [Related article: An Expected Value in a Ranking's Clothes]; this failure is structural rather than arithmetic.
The remedy is not statistical. It is a disclosure decision, made once, about what a number must arrive with before it may influence a commitment. An enterprise that requires the path count will sometimes find its most reassuring figures cover the smallest part of its exposure. That is uncomfortable in the cycle it lands in and far cheaper than the alternative, which is finding out when the last stream arrives late at the merge point nobody counted.
About EraNorth Insights
EraNorth Insights publishes practical analysis on strategy, projects, operations, transformation and decision intelligence for professional and organisational use. About EraNorth.