estimation-under-uncertainty git:20261016.c769bd8A

Immutable. This exact content never changes and is served at /api/v1/blob/eae3d21ab489f5b4.

---
name: estimation-under-uncertainty
description: >
  Producing a software estimate that carries its own uncertainty instead of hiding it: a range
  with explicit probability assumptions when supported, decomposition, PERT and the limits
  of summing task estimates, calibrating against what
  this team has actually done, and keeping estimate, target and commitment as three separate
  things. Use when asked how long something will take, when a single date is being requested
  for work that has not been broken down, when an estimate is being treated as a promise, when
  padding is being added silently, when a plan is slipping and the message has not gone out
  yet, or when someone asks for a number before the requirement is clear. Does not cover
  clarifying the requirement itself (requirements-and-acceptance), how to deliver bad news
  (engineering-communication), or trading quality for time (technical-debt-decisions).
---

# Estimation Under Uncertainty

## Purpose

An estimate predicts work or completion under assumptions. A single number without its
uncertainty can be planned against as certainty. Even a numerical range can mislead when
its coverage, scope or calendar assumptions are invented.

The job is to support the planning decision with a forecast and its uncertainty when evidence
permits, or identify what must be learned before a number is useful.

## Workflow

1. **Identify the decision and forecast event.** Reuse supplied scope, history and constraints;
   ask only for missing information that changes the decision (requirements-and-acceptance).
   Give a conditional range only when its bounds have a basis; pressure for a number does not
   supply one. Otherwise state what prevents forecasting and the next useful investigation.
   Define effort versus elapsed time, working versus calendar days, forecast date, start point,
   total versus remaining work, completion criteria, capacity and dependencies.
2. **Decompose where it exposes omitted work or a material dependency.** Comparable whole-change
   history may already be sufficient. Stop when more detail would not change the decision;
   unfamiliar work needs explicit assumptions or investigation, not endless subdivision.
3. **Choose inputs supported by evidence**: comparable cycle-time samples or elicited
   quantiles with explicit assumptions. Before using project history, follow the
   [history procedure](references/methods.md#reference-class-forecasting-use-your-own-history)
   to trace delivery records and check what their timestamps measure. A commit or merge is
   not automatically the work's start or completion. If using PERT, distinguish bounds and
   modal inputs from P10/P50/P90; they are not interchangeable. Without calibration, label
   the range provisional and do not fabricate a coverage percentage.
4. **Combine them with their dependencies.** For a sum, expected values add; variances add
   when covariances are zero, as under independence. Shared uncertain productivity, outages or
   integration risks can correlate durations. Shared people also constrain parallel work even
   when task durations are independent. Model precedence, capacity and calendars separately
   from statistical dependence; simulation does not infer them. A sum of effort is not
   automatically elapsed project duration.
   Task percentiles do not generally add, and the largest branch P80 is not automatically
   the project's P80. Derive the completion-time distribution for the actual schedule first;
   marginal task ranges alone do not determine it. Use the exact counterexample in
   [Methods and arithmetic](references/methods.md#task-percentiles-are-not-project-percentiles)
   when reviewing a deadline assembled from task percentiles.
5. **Calibrate against comparable history.** Record sample count, start/end definitions,
   changed team conditions and omissions. Three examples can anchor discussion, not establish
   reliable tail percentiles. Preserve dated forecasts and compare later outcomes at comparable
   horizons. Count known deadline misses even when delivery is unfinished; do not claim learning
   from a correction fitted and evaluated on the same outcomes. Use the
   [feedback procedure](references/methods.md#learn-from-outcomes-without-rewriting-history)
   when checking calibration or changing the estimation method.
6. **State the forecast event and assumptions**: a central P10–P90 interval has nominal 80%
   coverage, while "by P80" is a model completion percentile, not an interval.
   Discrete durations/rounding can change coverage; calibration requires actual outcomes.
   Both require a stated model/basis. Otherwise report an uncalibrated scenario range.
7. **Re-estimate when evidence arrives**, and say so immediately. An estimate is a statement
   about what you knew when you made it. Update when new evidence changes the decision or
   risk materially; retain the original forecast and explain the change. Use observed progress,
   remaining scope and current capacity, not percentage of the original time budget consumed.
   Updating one forecast does not demonstrate that the method is better calibrated.

Deliver scope and completion criteria, estimate units/range and its basis, key assumptions,
target versus commitment if supplied, and the next evidence/re-estimation trigger. When using
history, retain source identifiers, selection criteria and material exclusions so the basis
can be checked. Scale detail to the decision; do not invent a date, owner or probability
merely to fill this shape.

## Rules

- Avoid a bare single number when uncertainty affects the decision. If supported, give the
  range and explain the planning value selected from it for the decision's risk tolerance.
  Keep assumptions and residual risks visible; an unsupported broad range is still a guess.
- Distinguish a scenario span, a conditional model interval and an interval with coverage
  checked against comparable outcomes. State calibration evidence or its absence; a nominal
  model probability is not observed reliability. Say whether a quoted percentile is a deadline
  or interval endpoint. A confidence interval for a mean does not describe the uncertainty of
  one future delivery.
- Estimate, target and commitment are three different things and must be named separately. An
  estimate is what you predict, a target is what someone wants, a commitment is what you have
  promised. Deriving an estimate from a target is how a project becomes late on day one.
- Do not pad silently. Hidden padding obscures assumptions and calibration. Put contingency
  where it can be managed: at the plan level, visible, owned.
- The estimate covers the whole change, not the coding: tests, review, the review round trip,
  migration, deployment and anticipated rework. Avoid counting these twice if historical
  elapsed-time samples already include them.
- Name the uncertainty an investigation could resolve and its timebox. A spike can reveal
  more work and widen a range; do not promise a specific reduction without evidence.
- Report a slip when you believe it, not when it becomes undeniable. The information is worth
  most while there is still time to change something; overdue updates still matter for
  downstream plans (engineering-communication).
- Never resolve a schedule problem by silently lowering quality. If something must give, name
  it and let the trade be decided (technical-debt-decisions).

## References

- **Methods and arithmetic** — `references/methods.md`. Three-point estimation and PERT with
  worked numbers, when variances add, decomposition, reference-class
  forecasting from your own history, the cone of uncertainty, and the failure modes of each
  method. Read when selecting a method, combining work, forecasting remaining time or checking calibration.
- **Estimates, targets and commitments** — `references/commitments.md`. Keeping the three
  apart, what to do when a target is presented as an estimate, negotiating scope rather than
  dates, communicating a slip, and the agent-specific version of the same discipline. Read when
  an estimate is about to become a promise, or a plan is slipping.