1. Define the decision and what will change
Name the concrete intervention, delivery model, beneficiary geography, affected population, decision unit and time horizon. Describe realistic business as usual: existing services, spending and outcomes continue unless evidence says otherwise. Compare the policy with that baseline, including take-up, timing, capacity and displacement. A new programme name or administrative category does not mean activity was previously zero.
2. Review evidence before choosing a model
Prepare a short review of impact and CBA evidence. Assess research quality and relevance together: local evidence is valuable, but weak local evidence should not override stronger relevant studies. Search comparable countries, systematic reviews and transparent syntheses. Administrative data can establish quantities or costs without establishing causal effects. Separate observed impacts, appraisal forecasts, evidence transfers and assumptions. Inspect the original source; record the exact page or table supporting each material input.
3. Decide whether a numerical estimate is justified
An intervention needs a specified delivery model and a credible benefit mechanism supported by evidence or an explicitly accepted, source-backed transfer. If those are missing, report that BCR and NPV are not estimated and explain the present decision implication. Do not fill missing evidence with low, base and high numbers. Illustrative proxies require explicit participant acceptance, visible assumptions and a clear distinction from an evidence-supported appraisal.
4. List every cost and benefit across everyone affected
Review government, providers, companies, beneficiaries, users, households, caregivers and others affected. Record incremental public and private spending, time, travel, opportunity costs, setup, operating costs, maintenance and replacement. Count administrative time saved as a benefit. Children have no opportunity cost for their own school time; caregiver time savings can be benefits, while adult trainees' counterfactual earnings can be costs. Record material effects that cannot be monetized, with reasons.
5. Keep transfers and real resource effects visible
For a public-money decision, count government transfers as costs and the corresponding value received by recipients as benefits, unless a documented welfare reason changes that treatment. Include induced non-public costs. Keep each transfer separate from real resource effects; a transfer-excluded resource result is a diagnostic. Grants, subsidies, debt relief, taxes and financing amounts are not automatically new production or welfare gains. Avoid counting the same contribution, output, income or saving twice.
6. Use one real currency and value year
Put monetary inputs in spendable real local currency or real US dollars, with a common price year. Preserve the original amount and show inflation and exchange conversions. PPP international dollars are comparison inputs, not the final CBA currency. Source country inputs and use consistent wage or income definitions, exchange conventions and geographic coverage. Label household consumption or another substitute for wages as a proxy.
7. Derive the beneficiary-geography discount rate
Use the latest available five to ten years of real GNI, income or GDP per capita growth for the beneficiary geography. Exclude negative-growth observations, average the remaining observations, double the trend and round half up to a whole percentage point. Apply a 5% floor for low- and lower-middle-income beneficiary geographies; no floor applies to upper-middle- or high-income geographies. Use suitable subnational data or a transparent local-income analogue. Show the series, years, exclusions, income classification, candidate and final rate. These filtering, rounding and floor choices are Academy conventions.
8. Calculate dated incremental streams
For each year, calculate the quantity changed relative to BAU, multiply by the unit value, and discount by (1 + r) raised to the number of years from the base year. Sum the annual present values. Put setup costs in the year incurred; include only genuinely recurring costs later. Use a documented annual-equivalent method only when that matches the decision. Do not mix annual and lifetime amounts or treat an annual-equivalent total as an observed annual expense.
9. Show the result and what could change it
BCR equals present value of benefits divided by present value of costs. NPV equals present value of benefits minus present value of costs. Show the totals and major components, plus natural per-unit quantities where useful. The executive answer states the decision supported by current information and shows only the base-case BCR. Put numerical scenarios and alternative discount rates in a later sensitivity section. Identify the inputs that could reverse the conclusion; consider implementation capacity and diminishing returns before scaling.
10. Preserve sources, assumptions and count definitions
Each material input needs a source ID or a visible assumption, plus units, geography, population, date, price year, method and limitation. Record missing, blocked or restricted data access. Realized counts of people, devices, facilities, incidents and cases must be whole numbers. Fractional expectations are allowed only when labelled as modeled expectations or averages. Keep service volumes, successful outcomes and resource-use units distinct.
11. Make the report understandable and reconstructable
Lead with the decision, then explain the programme, BAU and mechanisms before detailed results. Show formulas, input definitions, substitutions, intermediate values and final results. Separate benefits, costs, transfers, setup and recurring components. Include a literature summary, source links and visible limitations. Use accessible charts from structured data, readable currency rounding and the shared BCR formatter: two decimals below 1, one decimal from 1 through 10, and none above 10. Keep exact values in supporting JSON.
12. Check the analysis and its actual delivery format
Run the maintained Academy compliance preflight, preserve its output and resolve blocking failures. Check the calculations against the saved inputs and inspect the rendered report. The existing preflight checks a manifest: declarations alone do not prove that a formula, source claim or visual presentation is correct. Automated checks support evidence and human review; they do not certify causal validity or confer Copenhagen Consensus endorsement. Record the toolkit version used.