| name | pho-registered-audit |
| description | Solve "Public Health Observatory" (PHO) registered algorithmic-audit tasks: a business prompt plus a formal analysis_request.json (protocol: cohorts, statistical modules, seeds/grids, robustness gates, decision rule) and a strict answer_template.json (output contract), backed by a read-only public-health data portal. Use this whenever a task hands you an analysis_request.json / answer_template.json pair that references a health-observatory Web portal and asks for one JSON object of resolved releases, cohorts, multi-module diagnostics, and a controlled decision. Covers state / county / country geographies and the recurring six-module audit anatomy (fixed-effects & jackknife, nested ridge / elastic-net CV, wild cluster bootstrap-t, grouped conformal, trajectory PCA + k-means + ARI, source/year perturbation & Shapley). |
Public Health Observatory — registered algorithmic-audit playbook
1. Recognise the task family
Every task in this family gives you three inputs (names may vary slightly):
prompt.txt — business framing (who is deciding what). It is context only; the
authoritative spec is the JSON payloads.payloads/analysis_request.json — the registered protocol. It declares the
geography scope, analysis years / reference year, outcome / exposure / adjustment
variables, evidence & cohort definitions, an ordered set of audit modules
(each with its exact method name, cohort, feature order, seeds, grids,
checkpoints), the robustness gates, and the decision rule.payloads/answer_template.json — the output contract: required top-level keys,
per-field types, array lengths, ordering rules, identifier rules, precision, and
cardinality constraints. Your deliverable is exactly one JSON object conforming
to it, with no narrative outside the JSON.
Access to the evidence portal is described in the task's own access file (an
or equivalent that lists the base URL and the allowed
read-only routes). — do not assume routes; the portal is read-only and exposes a catalog,
geography references, per-domain data tables, a revisions table, a methodology
library, and a bulk CSV/table export. Pull the full tables once and compute locally.
