Use Estimation Prover with your AI.

Validate estimation

The Planning Fallacy (Kahneman 1979) proves humans systematically underestimate by 25-50%. even when they KNOW about the bias. The only defense is structured estimation. You must: (1) DECOMPOSE scope into units ≤2 days each. each with its own estimate. A single estimate for a multi-day task is a guess. Decomposition forces specificity, (2) MAP unknowns. technical risks, knowledge gaps, dependency uncertainties. For each: name, likelihood (low/med/high), impact on timeline if it materializes. The unknowns will blow the estimate, not the knowns, (3) cite a SPECIFIC historical precedent. not "based on experience." Reference Class Forecasting (Flyvbjerg 2006): name the project, how long it took, how accurate the original estimate was, and how THIS task compares, (4) apply a CONTINGENCY buffer with a NUMBER. ≥20% for familiar work, 40-60% for novel work with unknowns. The Cone of Uncertainty (McConnell) shows early estimates are 4x off. A buffer without a number is not a buffer, (5) state EVERY assumption. team availability, scope stability, API reliability, reviewer responsiveness, infrastructure readiness. "No assumptions" means you have not examined. If rejected, your estimate has a blind spot. Structured reflection capability for project estimation. forces decomposition, unknown mapping, historical grounding, contingency buffers, and explicit assumptions BEFORE committing to a timeline. Based on Reference Class Forecasting (Kahneman/Flyvbjerg), Cone of Uncertainty (McConnell), and Planning Fallacy research (universal 25-50% underestimation). Catches Scope Vague ("auth work" instead of "migrate JWT to OAuth 2.1 with refresh token rotation, update 3 API endpoints, update React auth context, write migration tests". vague scope produces vague estimates), Unknowns Hidden (estimating without mapping technical risks, knowledge gaps, and dependency uncertainties. the things you do not know will blow the timeline, not the things you do know), No Precedent ("based on experience" instead of "the auth migration at Company X took 3 sprints, estimated at 1.5 sprints, because OAuth discovery flow testing took 2x longer than expected". Reference Class Forecasting requires SPECIFIC historical precedent), No Buffer (estimates without contingency. Kahneman proved humans underestimate by 25-50%. A buffer is not padding. it is correcting for a known cognitive bias), and Assumptions Implicit ("2 weeks" without stating that this assumes full-time allocation, stable scope, available API documentation, responsive code reviewers, and working CI pipeline). Call once per estimation