Kabla ya kufungua spreadsheet yoyote, timu ya statistics iliandika protocol ukutani.
Amina, mwanafunzi wa mwaka wa nne wa statistics, alisimama mbele ya whiteboard.
“Tunataka swali moja. Si ‘model ni mbaya?’ Swali ni: katika sample yetu, ni aina gani za context zinazohusiana na adverse flags ambazo policy yenyewe ingeona legitimate need?”
Wema akakubali.
Participants walikuwa ishirini na nane, wote kwa ridhaa. Hakuna aliyelazimishwa kuonyesha income. Walitoa tu decision outcome, category iliyotajwa kwenye notice au export yao wenyewe, na context ambayo wangeweza kuthibitisha kwa njia isiyotambulisha zaidi ya lazima.
Team iligawanyika. Group moja ilipokea context bila kujua outcome. Group nyingine ilirekodi outcome bila context. Baadaye walilinganisha.
“Blind coding,” Amina alisema. “Tusiandike kile tunachotaka kuona.”
Case ya kwanza: purchase ya helmet. Context coders waliweka *mandatory course safety equipment*. Outcome: high-risk band.
Case nyingine: bus fare ya funeral. Context: family emergency travel. Outcome: medium-risk.
Case ya tatu: restaurant expense. Context: participant alikiri ilikuwa celebration ya hiari. Outcome: no flag.
Wema alifurahi kuona counterexample. “Hii lazima ibaki.”
Amina akamwangalia. “Hatutengenezi campaign poster.”
Walitembea student services kuthibitisha baadhi ya process outcomes kwa participants waliokubali kuonyesha letters zao. Wema hakushika copies. Aliona tu fields zilizokuwa muhimu na participant mwenyewe akarekodi.
Kufikia usiku, pattern ilianza kuonekana. Mandatory course costs—equipment, transport ya field placement, printing na materials—zilikuwa overrepresented katika high-risk bands kuliko walivyotarajia kwa sample ndogo.
“Correlation,” Amina alisema. “Si proof ya causal rule.”
“Headline yangu haitasema model targets poor engineering students.”
“Please usifanye hivyo.”
Walicheka.
Timu ilikokotoa confidence intervals na kuandika limitations ndefu kuliko conclusion. Sample haikuwa random. Watu waliokuwa na problem walikuwa likely zaidi kujitolea. Hawakuwa na code ya model. Public media signal ilikuwa difficult verify.
Kelvin, activist wa assembly, alifika na kuangalia report draft.
“Hii ni weak,” alisema. “Mnaandika ‘suggests,’ ‘within sample,’ ‘cannot infer.’ Hivi donor atashtuka?”
Amina akamjibu, “Hatufanyi donor ashtuke. Tunafanya claim iwe testable.”
Kelvin akamwangalia Wema. “Na wewe umekuwa professor sasa?”
Wema akasema, “Nimekuwa wrong mara nyingi wiki hii. Limits zinasaidia.”
Kelvin akaondoka bila kuridhika.
Saa tano usiku, walifunga analysis. Report ilikuwa na table ya counterexamples, si only failures. Pia ilionyesha kwamba baadhi ya flagged cases zingeweza kuwa legitimate fraud concerns; problem ilikuwa auto-sanction bila context, si kwamba fraud review haipaswi kuwepo.
Wema alituma report kwa Grace kupitia contact aliyompa donor showcase.
Dakika kumi na nne baadaye, reply ikaingia.
*Can your team support on-site verification of the process, not individual cases?*
Wema akajibu ndiyo.
Grace akatuma calendar request.
**Monitoring visit: within 48 hours. No staged route requested.**
Amina akasoma na kusema, “Sasa spreadsheet inatembea.”
Wema akafunga laptop. Hiyo ilikuwa ndiyo alichotaka: evidence kutoka anecdotes kwenda test, na kutoka test kwenda process ambayo mtu wa nje angeweza kuona kwa macho.
Amina alisisitiza sample recruitment isifanywe na Wema peke yake. Walitumia student press notice, welfare clubs na random time slots kwenye student services ili kupunguza self-selection kidogo, ingawa hawangeweza kuiondoa.
Wema hakushiriki coding ya cases ambazo aliwahi ku-interview. Alikaa upande wa documentation.
Case moja ilivunja expectation yao: student aliyekuwa na expensive phone purchase alikuwa green kwa sababu transaction haikupita kwenye merchant category inayofuatiliwa. Case nyingine ya low-cost safety boots ilikuwa red.
“Hii inaonyesha amount si core factor,” Amina alisema.
“Category mapping inaweza kuwa.”
“Careful. Inaweza.”
Walitafuta counterexamples zaidi.
Mwishoni walikuwa na table ambayo haikutoa conviction, lakini ilionyesha mismatch ya policy objective na observed outcomes. Mandatory course expenses zilionekana mara kwa mara kwenye adverse bands. Emergency transport pia. Pure leisure examples hazikuwa uniformly flagged.
Hiyo complexity ilifanya report iwe believable.
Monitoring analyst kutoka consortium alipiga simu na kuuliza methodology. Amina ndiye aliyejibu technical questions; Wema hakujifanya statistician.
“Confidence limits zenu ni wide,” analyst alisema.
“Ndiyo,” Amina akajibu. “Ndiyo maana tunashauri on-site process verification, si population claim.”
Wema aliandika sentence hiyo kwa article.
Kelvin alipoona draft, aliita “academic surrender.” Lakini Wema aligundua kwamba campaign yenye claim kubwa inaweza kushinda siku moja na kupoteza credibility kesho. Report yenye limits ingeweza kusukuma test ambayo institution haingeweza dismiss kama viral outrage.
Walituma appendices zote isipokuwa participant identities. Grace alithibitisha receipt.
Usiku, Wema alitembea student services na kuangalia notice board tupu ya appeals. Alifikiria jinsi report ingekuwa na maana ndogo kama monitoring visit ingefuata route iliyopangwa na PR team.
Hivyo alipomjibu Grace, aliongeza: *Please verify ordinary process points, not only presentation rooms.*
Grace akajibu: *Agreed. I will choose route on arrival.*
Ndani ya saa 48, spreadsheet yao ingeacha kuwa argument ya online na kukutana na desks, staff na system halisi.
Timu ilihifadhi protocol na codebook kwenye public repository bila case data. Hiyo iliruhusu monitoring team kuona jinsi categories zilivyowekwa bila kuona identities. Amina alisisitiza pia waweke cases mbili ambazo hazikuunga mkono hypothesis yao. Wema alikubali, ingawa headline ingekuwa dhaifu. “Counterexample si adui,” Amina alisema. Huo ulikuwa mstari Wema aliandika kwenye notebook. Report ya mwisho ilitenganisha observation, inference na recommendation. Recommendation kuu haikuwa “model ni discriminatory” bali “run controlled verification of context-blind categories and appeal process.” Hilo ndilo lililomfanya Grace aombe ziara ya site, hatua ambayo ingetoka kwenye measurement kwenda kwenye inspection ya process halisi.
Walipomaliza, Amina ali-print ukurasa mmoja wenye summary na kuusoma kwa sauti kwa timu. Kila claim ilikuwa na qualifier yake. Wema alihisi urge ya kuondoa baadhi ya maneno kama “within this sample,” lakini akaacha. Kisha walitengeneza list ya vitu ambavyo monitoring visit ingepaswa kuona: appeal desk, service-propagation rules, public policy notice, na live synthetic test. Hiyo list ndiyo ilibadilisha audit kutoka exercise ya darasani kuwa bridge kuelekea verification ya operational reality.
Lakini alijua administration ingetayarisha campus kabla ya Grace kufika.