Lessons require original forecasts and actual outcomes

Delivery lessons require the original forecast, its evidence and the changes that occurred. A final completion date alone cannot explain why a project differed from its plan. The actual scope may have changed, resources may have become unavailable, an acceptance obligation may have been omitted or the original estimate may have been poorly supported. These causes have different implications. Preserve the contemporaneous basis so the review can distinguish them rather than judging every decision solely from the final result.

A useful learning record connects forecast, actual event and accepted scope. The forecast should have a version, cutoff, task boundary and estimate basis. The actual should state the event and evidence. The scope comparison should identify whether they describe the same obligation. Without this mapping, a duration comparison can be numerically correct but substantively meaningless. A larger package cannot fairly be treated as the same task simply because its identifier stayed unchanged.

Outcome quality and decision quality differ. A well-supported decision can have an unfavourable outcome because an acknowledged uncertainty materialises. An unsupported decision can have a favourable outcome through luck or an unrecorded intervention. Learning should examine the reasoning and evidence available at the time as well as what happened. This protects against copying successful but unsound practices and abandoning reasonable controls merely because a particular outcome was poor.

The project should retain rejected alternatives and challenges. They reveal what options existed and why the authority chose one. A recovery decision may look unnecessary after the original supplier delivers early, but the contemporaneous record may show a credible threat and an expiring fallback. Conversely, an acceleration can look successful while another unrelated change actually produced the saving. The record should preserve the causal assumptions so later analysis can avoid attributing the whole outcome to the most visible action.

The frozen studies reproduce fictional assumptions and exact outputs; they are no historical performance dataset. A lessons programme needs comparable accepted events, contextual conditions and a transfer-evaluation method. The paper supplies that record structure, without observed rates or task-distribution calibration from invented mechanism examples. Numerical transfer must therefore await an appropriate empirical account.

A variance taxonomy should be specific enough to guide improvement. Scope change, estimate-basis change, execution departure, resource conflict, interface failure and decision latency can all move completion. These are original analytical categories for the record, not a universal attribution standard. The review should permit more than one cause where evidence supports it. Forcing every delay into a single code can obscure interactions and encourage responsibility arguments before the facts are understood.

The reviewer should compare the planned and actual release conditions. If the supplier task was forecast from design release but actually began while the configuration remained unstable, the duration difference may reflect a start-basis departure. If the receiving scope expanded, the original duration may no longer be comparable. If the task completed physically but waited for acceptance, the difference may concern release evidence rather than productivity. These distinctions help the next estimate and interface design.

Record interventions that changed the outcome. Additional resources, partial releases, changed logistics or scope decisions can make actual duration shorter or longer than the original basis. A lessons dataset that records only the final duration loses this information and may produce misleading analogies. The next planner could copy a short duration without knowing it depended on exceptional capacity or deferred scope. Context is part of the evidence, not an optional narrative beside the number.

A review should also examine reporting behaviour. Did the forecast change when evidence changed? Were completion states used consistently? Did an unaccepted submission release a successor in the chart? Did rebaselining remove the comparison needed for diagnosis? These controls affect whether the team saw the problem in time. Improving them can be more valuable than adjusting a generic contingency factor. The lesson should identify the point where better evidence or a clearer decision could have changed action.

Responsibility and learning can coexist, but their purposes should be explicit. A contractual assessment may need causation and notice evidence under actual terms. A learning review may seek process improvement across organisations. Combining them without clear rules can inhibit candid evidence. The control record should preserve facts for both while avoiding unsupported legal conclusions. The reviewer can state an operational mechanism without claiming that it determines liability.

A fictional archive records a short supply duration from an earlier installation. The headline says the supplier performed exceptionally. The detailed record shows that the project paid for additional capacity, accepted a partial lot and used a different receiving arrangement. A planner preparing the next baseline wants to reuse the short duration. The learning review asks whether those conditions exist in the new undertaking before treating the outcome as a transferable rate.

The archive's original forecast and change record are essential. They reveal that the short duration was not the supplier's ordinary basis under unchanged scope. It resulted from a specific intervention with cost and interface consequences. The next planner can consider a similar intervention, but it needs current capacity, admissibility and acceptance evidence. Copying the outcome without the mechanism would turn a genuine success into an unsupported estimate.

The scope comparison also matters. The earlier supply event may have been partial acceptance, while the new task means full availability for assembly. The same package name does not make the events comparable. The learning record should map accepted outcomes and residual work. If the archive contains only a final date, the planner should state that the evidence is insufficient for a strong numerical transfer. The absence of detail is a data limitation, not a reason to invent equivalence.

Review unsuccessful and ordinary cases where available alongside celebrated recoveries. Selection can overstate normal capability; an empirical account must define population, context and missingness. This paper provides no such dataset or calibration. Its qualitative archive rule is to retain interventions and conditions with outcomes so future estimates expose survivorship and selection effects rather than inherit them unnoticed.

The organisation turns the review into a concrete archive rule. Supply records will distinguish original basis, authorised changes, accepted event and material interventions. They will retain configuration and receiving boundaries. The rule is checked on a subsequent closeout rather than declared successful immediately. This makes learning an operational control with evidence, not a collection of optimistic anecdotes.

The next baseline can still use the earlier project as an analogy, but the record explains the adjustment or qualification. A mechanism-based question may be more defensible than a rate: is the required supplier capacity actually secured, and does partial delivery release useful work under this configuration? The lesson preserves the value of history while preventing a favourable outcome from becoming universal proof.

A reusable lesson can take the form of a question or control rule rather than a numerical rate. For example, verify that supplier readiness and receiving acceptance use the same configuration before joining their milestones. This rule follows from a mechanism and can be adapted across settings. It does not imply that every interface review takes the same duration. Qualitative lessons are valuable when they direct attention to a causal condition without pretending to offer a universal empirical parameter.

A second archive problem arises even when the organisation keeps every project. A package still open at the collection cutoff has no observed accepted-completion duration. Treating it as zero, discarding it from a completed-duration average or inventing its eventual date can bias the apparent performance. Preserve its start, cutoff, current scope and unfinished state. The elapsed observation is real; the completion outcome is not yet known. A later analytical dataset should state how it handles these incomplete observations using a method suitable for its intended estimate. This paper supplies no fitted duration distribution or numerical accuracy claim.

Cancellation creates a different observation again. A cancelled obligation can carry real cost and elapsed time without producing the accepted event used by a completed delivery forecast. A scope replacement may close the old package administratively while another package carries its remaining service duty. The archive should retain those dispositions instead of treating administrative closure as physical acceptance. The appropriate comparison depends on the question: elapsed exposure, completion timing, resource use and achieved service require different outcome definitions. Combining them under one completed flag loses precisely the mechanism that the learning process needs to preserve.

A fictional lesson-retirement record tests a previously useful rule: inspect every supplied module at site before assembly. Its original basis was a mismatch between factory inspection and the installed interface. A later receiving method supplies a different configuration-linked acceptance sequence. The reviewer retains the objective, compatibility before reliance, but retires the universal site-inspection wording when the new sequence addresses that objective. If the new record omits interface evidence, the lesson remains applicable. This disposition shows how a lesson can be narrowed without erasing its history. The transferable object is the failure mechanism and controlling claim, rather than the organisational location where an earlier project happened to perform its check.