# How a Cookwala food-rescue pilot is run and judged **Status:** draft protocol, 2026-10-04 (RFC-0003). For a food bank, community kitchen, school-meal or relief program that wants to try the Humanitarian Profile. No pilot has run yet. This document says how one would be run so that its result, good or bad, can be trusted. ## 1. What a pilot tests A pilot tests one claim: *with the Humanitarian Profile (SMS, spreadsheets or an API, plus rule packs), a program rescues more food safely, serves it sooner, and knows its own numbers, compared with how it works today.* It does not test robots, and it does not claim to reduce hunger in a city. ## 2. Hypotheses, written before the pilot starts | # | Hypothesis | Measure | Minimum worth continuing | |---|---|---|---| | H1 | More surplus reaches kitchens | kilograms rescued per week (first-leg `Handover.kgAccepted`) | at least 10 % more than baseline | | H2 | Food moves faster | median minutes from `Offer` to `claimed` | at least 20 % shorter than baseline | | H3 | Rejections are recorded and donors improve | share of handovers with a recorded reason for every rejected line; share of offers arriving within the temperature band | reasons recorded 100 % of the time; the in-band share rises. More recorded rejections in the first weeks is expected and is not a failure | | H4 | Menus meet the nutrition rules more often | nutrition pass rate on `Distribution.menu`, labelled `self_reported` when nutrients are typed and `measured` only when derived from linked recipes and portion masses | rises or stays, never falls | | H5 | Cost per meal does not rise | (food + transport + staff + energy) / meals | no more than 5 % higher | | H6 | Staff and volunteers find it no harder | minutes of recording per 100 kg from the evaluator's weekly time log (`templates/time-log.csv`, one row per recording session; not `volunteerMinutes`, which is kitchen labour); an anonymous survey | not worse than baseline | | Safety | No safety incident is caused by the tools | `safetyIncidents`, block findings acted on | zero attributable incidents | The exact thresholds are set with the partner before the start and pre-registered (a dated file in the program's repository or a public registry such as the Open Science Framework). ## 3. Baseline Four weeks of the program's current practice, transcribed by the evaluator from the program's existing records (phone logs, delivery notes, receipts, waste counts) into the same measures. Staff do not fill Cookwala templates during the baseline, so the baseline is not shaped by the tool it measures. Without a baseline there is no result. What a pilot this size can show: with one program, about five donors and twelve pilot weeks, only large effects are detectable. The pre-registration states the minimum number of weekly observations per measure and the comparison method (week-on-week ranges against the baseline weeks, with the same weeks of the year where seasonal supply or Ramadan would otherwise explain the difference). A 10 % change that falls inside the week-to-week spread is reported as "not distinguishable", not as a result. ## 4. Design - **Duration:** 12 weeks after the baseline, plus two weeks of training. - **Sites:** at least two receiving sites and five regular donors, so one unusual donor does not decide the result. Where possible, one site keeps the old method for the same period (a comparison site); if not possible, say so. - **Level:** start at H0 (SMS and spreadsheets). Move to H1 (API) only if the partner wants it and H0 ran for at least four weeks. - **Rule packs:** `basic-nutrition-food-safety` plus, where children or older adults are served, `care-vulnerable-groups`; adapted to national rules by the program's food-safety lead and recorded in `reviews`. - **Who records:** the program's trained staff, by role. No person is named in any document. ## 5. Data responsibility - No personal data enters any document (profile section 7). The SMS gateway maps numbers to organizations only. - A data-responsibility review against the ICRC Handbook on Data Protection in Humanitarian Action and OCHA's data-responsibility guidelines is done before the start and recorded. - University or institutional ethics approval where the partner requires it. - The staff and volunteer survey (H6) is voluntary, anonymous and reported only in aggregate; its questions are pre-registered with the hypotheses, and no answer is linked to a person or shift. - Hosting in-country where the program or the law requires it; retention set in the `Manifest`; deletion after the pilot unless the partner keeps the documents. ## 6. Analysis - Measures are computed with `tools/humanitarian_check.py --summary` from the documents, never typed by hand. The `ImpactSummary` lists the method and the number of documents behind each value. - Compare pilot weeks with baseline weeks; report ranges, not single numbers, when weeks vary. - An evaluator independent of Cookwala and of the partner reads the raw documents and the summary and signs the result. ## 7. Kill criteria Stop the pilot early if any safety incident is attributed to the tools, if recording time per 100 kg in the time log is more than double the baseline after week four, or if the partner asks. Stopping is a result and is published. ## 8. Publication The protocol, the baseline, the `ImpactSummary` and the evaluator's note are published on cookwala.ai whatever they show, with a "what went wrong" section. Negative results are published with the same prominence as positive ones. The partner is named only with its written agreement; otherwise the site says "a food bank in Cairo" and no more. ## 9. Budget (indicative, assumed) An SMS gateway subscription, printed templates, two weeks of training time for staff, a part-time field coordinator for 16 weeks, a food-safety officer's review time, and an evaluator's fee. Hardware: none beyond phones the program already has. A figure in money depends on the country and is not stated here. ## 10. After the pilot If the thresholds are met: a second site and the H1 API; application for Digital Public Good recognition with the pilot evidence. If not: the result is published, the profile is revised through an RFC, and the next pilot waits for the revision.