Requirements Nobody Wrote Down: Emergence in a Public-Sector ML Pipeline
B · Useful, limited A careful 15-month insider case study from the Federal University of Pará and partners, with artifact analysis, six interviews and a replication package; single case, first author was on the team.
Summary based on the full text·Original: about 40 minutes to read
In short
An arXiv paper (5 November 2026) by Ferreira, Pinto and colleagues on how requirements came about for a machine-learning classifier that sorts citizen proposals by theme on Brazil's national participation platform. Almost nothing was specified up front. Demands arrived as broad goals; AI was the delivery team's offer, endorsed but never sponsored; the key requirement, the category set, was settled by adopting a government taxonomy chosen for legitimacy. Acceptance rested on metrics, domain judgement and politics, with no shared threshold and no test of usefulness in practice.
Key numbers
- over 60,000citizen proposals on Brazil's national digital participation platform
- 40%share of the 95 reconstructed requirements and criteria that were ever formalised (38 of 95)
- 0 of 17success criteria formalised
- 43project artifacts analysed (issues, commits, notebooks, presentations, meeting notes)
As reported by the source.
English·Web page·opens arxiv.org
What it reports
- 95 requirements and criteria reconstructed: 21 functional, 12 non-functional, 23 acceptance, 22 evaluation and 17 success criteria. Only 38 (40%) were written in an issue or meeting note; no success criterion was.
- Interviews added 14 more requirements and 7 criteria beyond the 95.
- The method moved from BERTopic clustering to seeding with the government vocabulary (VCGE) to supervised classification, then to a service whose deployment depended on a state IT company.
- Seven patterns, including: classification requirements surfaced in reaction to concrete model outputs; operational requirements surfaced only during integration.
- A recurring need to validate LLM-produced labels was raised but its systematic execution could not be confirmed.
How they got it
Exploratory single-case study, January 2025 to March 2026. The authors catalogued all 43 artifacts, normalised requirements, reconstructed 17 episodes, then interviewed six participants (mean 68 minutes) using the episodes as prompts. Coding was deductive and inductive with a second-author review. Claude Opus 5.5 assisted with coding review and episode consolidation, with all outputs reviewed by the first author.
Limits
- One case; the first author is a team member and conducted all interviews.
- Small interview set; two codes appeared only in the last two interviews, so no saturation is claimed.
- Artifacts and transcripts cannot be released; LLM-assisted consolidation adds a layer the reader cannot fully audit.
Why it matters for companies
The finding generalises well beyond government: in ML and AI projects, requirements surface through experiments and integration, not workshops. Teams should log requirements as they appear, assign an owner to the category scheme or taxonomy, and agree acceptance thresholds and an in-use test before anyone calls the model done. Endorsement without a sponsor is a predictable failure mode for AI projects started by the delivery side.
Summary based on the full text. Read the original for the full argument and data.Read the original on arxiv.org