Pho Algorithmic Audit

Public Health Observatory registered algorithmic audit

These tasks look intimidating but share one shape: resolve published records from a data portal → build declared cohorts → run N declared statistical modules → evaluate decision gates → emit one JSON object that conforms exactly to answer_template.json. The portal is the only evidence source; everything is computed deterministically from it (the sole randomness is inside a bootstrap whose PRNG is fully specified).

Inputs you are given (per task)

• prompt.txt — the framing and the portal base URL (a placeholder like <TASK_ENV_BASE_URL>).
• payloads/analysis_request.json — the registered protocol: scope, years, outcome/exposure, evidence filters, cohort definitions, the audit modules (each with its standard_method, feature/coefficient order, grids, seeds and required_evidence), and the decision rule / gates.
• payloads/answer_template.json — the response contract: required keys, array lengths, orderings, identifier and precision rules, enum values.

Read all three fully before computing. The request and template are exhaustive — they tell you every field, order and threshold. Treat them as the spec.

Workflow

1. Map the portal. From the given base URL, read GET /catalog (datasets, columns, filters, CSV export) and GET /methodology (the publication rules). Download whole datasets as CSV and process locally — they are small.
2. Resolve releases. Apply registered-final-release resolution to every requested series: keep FINAL, take the highest revision, drop PROVISIONAL, treat suppressed/invalid-flagged values as unavailable (never zero-fill), honor direct-vs-rollup and crude-vs-age-adjusted. For country tasks, apply the revision-notice / scale-break logic. See references/data_model.md.
3. Build the declared cohorts (complete-case per year, reference/primary, balanced panel, ML/augmented, strict dual-source) and report the audit counts exactly as the template asks — including that resolved-record counts include suppressed rows, while completeness excludes them.
4. Implement each audit module from its declared spec, matching the reproducibility conventions in references/audit_modules.md (SE variants, training-only standardization, the exact named PRNG + seed/stream, covariance PCA + deterministic k-means init, tie-breaks). Preserve every declared order.
5. Evaluate the decision block: turn each gate's threshold expression into a boolean on your computed statistics, count passes, apply the classification / precedence ladder literally.
6. Emit one JSON object matching the template: correct keys and array lengths, declared decimal places as JSON numbers, integers/booleans as natural types, identifiers exactly, null only for mathematically-unavailable statistics. No text outside the JSON.

Reference files

• references/data_model.md — portal datasets, the exact release/revision resolution rules, country scale-break handling, cohort patterns, and the output contract (rounding, ordering, identifiers, null rule).
• references/audit_modules.md — the six recurring module families with the conventions that make each field reproducible, and how to build the decision block.
• references/portal_toolkit.py — a dependency-light Python helper (takes the portal base URL as an argument): CSV dataset loading, resolve_final, suppression handling, geography lookups, and country-label reconciliation. Import and extend it; it hardcodes no task values.

What makes or breaks the score

The graded output is the numbers, so precision and convention-matching dominate:

• Resolution is the foundation. A wrong revision/suppression/direct-vs-rollup choice silently corrupts every downstream cohort and statistic. Verify counts.
• Match declared conventions exactly: SE type (CR1 vs classical vs HC3), standardization scope (training-only), penalty/units (_per_10000, log_*), PCA covariance-vs-correlation and sign rule, k-means initialization, and CV tie-breaks. A plausible-but-different convention scores as wrong.
• Reproduce the specified PRNG bit-for-bit (XORSHIFT32 / PCG32 with the given seed/stream, draw order, and weight family); emit the requested PRNG-state and statistic checkpoints and the terminal state.
• Ordering and identifiers: never re-sort an aligned array; use exact codes, division names, ISO3 and enum strings; round to the declared places; null only when truly undefined.
• Decision logic is deterministic once the module statistics exist — evaluate each gate literally and apply the precedence ladder.

Work in a scratch dir (not the task input); build a small reusable toolkit for resolution + the module families and reuse it across tasks, since the portal schema, resolution rules and module families are stable across the family.