Pho Registered Audit

Public Health Observatory — Registered Algorithmic Audit

This skill completes any PHO registered algorithmic audit task. Every such task has the same shape: a read-only portal is the sole evidence source; an analysis_request.json declares the protocol (geography, years, outcome/exposure, cohorts, audit modules, gates, decision rule); an answer_template.json fixes the response contract (keys, orders, lengths, types, precision, enums). The domain varies (state longevity, county mediation, country burden, state diabetes, county diabetes dynamics) but the operating procedure is constant. Follow it exactly.

What you receive

• prompt.txt — restates the gates and the submission constraint (one JSON object, no narrative).
• payloads/analysis_request.json — the protocol (the "what to compute").
• payloads/answer_template.json — the contract (the "how to format").
• environment_access.md — the portal base URL and the only allowed GET endpoints.

Governing principles (non-negotiable)

1. Portal is the only evidence source. Fetch everything from <TASK_ENV_BASE_URL> via the allowed GET endpoints. No external data, no fabricated values, no recalling outcomes from memory. See references/portal.md.
2. Resolve releases and revisions independently and reproducibly. Final replaces provisional; the highest applied final revision wins; pending/withdrawn revisions never replace; declared revision_priority tie-breaks. See references/portal.md.
3. Never zero-fill. Suppressed or missing values are unavailable. The cohort drops the jurisdiction/year, never imputes zero. Use JSON null only when a statistic is mathematically undefined (e.g., a degenerate fit), never for missing source data.
4. Preserve every declared order. State order, feature/coefficient order, grid order, division order, checkpoint-replicate order, source-group order, subset order. Do not re-sort an array that is positionally aligned to another (e.g., delete-coefficients align to state_order; PCA scores align to state_order). Sort only where the template says so.
5. Deterministic numerics. Use the exact RNG (PCG32 / XORSHIFT32), seed, stream, replicate count, and weight scheme declared. Record checkpoint PRNG states + t-statistics at the declared checkpoint replicates, plus-one p-value, requested quantiles, and the terminal generator state. Compute in full precision; round only at output.
6. Match the contract exactly. Required keys, array_lengths, cardinality_rules, types (integer vs number vs boolean vs enum), precision, identifier casing, enum values. See references/contract.md.
7. One JSON object, no narrative. Submit exactly one JSON object conforming to answer_template.json. No markdown, no commentary, no surrounding text.

Procedure

0. Read the contract first

Open answer_template.json and enumerate, per top-level key: required sub-keys, array lengths, ordering, types, precision, and enum/allowed values. Then open analysis_request.json and map each module to its cohort, ordered features, grids, seeds, replicates, checkpoints, and required_evidence. Reverse-engineer the deliverable shape before computing anything — the contract tells you which arrays must align with which.

1. Pull the portal

• GET /catalog — column schemas, available filters, row counts, measure dictionary.
• GET /geographies/{states,counties,countries} — jurisdiction, region, division, rucc, canonical-name/alias reference.
• GET /download?dataset=<d>&format=csv&<filters> — machine-readable records (CSV with header). Filter query params mirror the catalog's documented filters (state_abbr, year, measure_id, value_type, source_type, release_status, revision, region, county_fips, iso3, country_label, indicator_id, …).
• GET /data/revisions — revision events (status = APPLIED/PENDING/WITHDRAWN, reason_code, effective_year, old/new values).
• GET /methodology and GET /methodology?doc=<slug> — resolution policy. Read the CURRENT docs; ignore DRAFT and SUPERSEDED docs for publication decisions.

Pull the full datasets you need (filter per measure/year/value_type/source_type to keep payloads small), then resolve locally. Do not trust a single filtered view to be the universe — cross-check against /geographies for the jurisdiction universe (e.g., 50 states + DC = 51).

2. Resolve the evidence

For each requested (jurisdiction, year, measure, value_type, source_type):

1. Keep release_status = FINAL only (final replaces provisional).
2. Among final records, keep the highest applied revision; break remaining ties with the declared revision_priority (typically revision, then released_at, then observation_id / record_id).
3. Ignore revision events whose status is PENDING or WITHDRAWN — they do not authorize replacement.
4. Drop any record whose quality_flag is in the request's invalid_quality_flags (e.g., INVALID_SCALE, INVALID, WITHDRAWN). Other quality flags (e.g., REVIEWED, STALE) are retained — they describe review state, not release precedence.
5. Treat suppressed (suppression_flag) or blank value as unavailable; never zero-fill.
6. Respect source/value-type semantics: DIRECT_SURVEY is the primary state series; COUNTY_ROLLUP is a parallel coverage estimate that must not silently replace direct records. AGE_ADJUSTED for state comparisons; CRUDE for observed burden (primary vs parallel exposure).
7. For countries: reconcile portal_label/alternate_labels to stable iso3 via the countries geography file (alias resolution).

Record the resolution in the cohort section of the answer (release counts by year, excluded jurisdictions, yearly complete-case counts, etc.) — the template requires it.

3. Build the cohorts

Construct each cohort independently from resolved records, per the request's definitions:

• PRIMARY — reference-year complete-case from the jurisdiction universe.
• BALANCED — complete across all study years (intersection across years).
• BROAD / REFERENCE — reference-year complete for the outcome and every ordered ridge/elastic-net feature.
• ML / machine-learning — primary cohort members also complete for the extra ML fields.
• STRICT_DUAL_SOURCE — complete for outcome + primary exposure + parallel exposure + adjustments in every study year.

"Complete" = every requested value present, nonsuppressed, nonmissing, and not carrying an invalid quality flag. Track the complete exclusion set (every universe jurisdiction absent from a cohort, and no others) — templates demand e.g. primary_excluded_state_codes, core_balanced_excluded_state_codes, strict_excluded_state_codes.

4. Run the six audit modules

Each request declares six modules. Execute each exactly as specified — method, cohort, ordered features, grids, seeds, replicates, checkpoints — and produce the required_evidence. The six recurring families and their deterministic execution notes are in references/modules.md. Summary:

1. Delete-one-cluster fixed-effects / two-step GMM / jackknife — full coefficient, every per-deletion diagnostic, bias-corrected inference, influence summary.
2. Nested leave-group-out ridge / elastic-net CV — outer + inner folds, grids, selected hyperparameters, per-fold inner grids, outer RMSE, pooled metrics, worst group.
3. Restricted-null wild cluster bootstrap-t — observed CR1 SE/t, exceedance/tail count, plus-one p-value, requested quantiles, checkpoint PRNG states + t-stats, terminal state.
4. Grouped split conformal — per-group thresholds/coverage/widths/MAE, aggregate calibration, worst group.
5. Trajectory PCA + deterministic k-means — spectrum, loadings, scores, deterministic centroids, Lloyd update count, labels; leave-year-out and/or delete-state ARI stability.
6. Exhaustive source/year/group perturbation — all subsets/groups, coefficient + p-value vectors, percent shifts, same-sign summary, worst subset; exact Shapley when declared.

5. Apply the gates and decision

Evaluate every gate against its declared threshold (in robustness_gates / decision_rule / controlled_conclusion — do not invent thresholds). Report per-gate PASS/FAIL (or boolean), the passed/supported count, first_failed_module (or NONE), and the classification per the declared precedence rule (e.g., all-pass → primary; ≥N → partial; else → none). The decision is mechanical once the gates are computed.

6. Emit and self-check

Serialize one JSON object. Before submitting, run the pre-submission checklist in references/contract.md:

• Exactly the required top-level keys; every required sub-key present.
• Every array_lengths / cardinality_rules satisfied.
• Shared state_order is byte-identical across the modules the template says must match.
• Aligned arrays align positionally; sorted lists sorted exactly as specified.
• Precision correct (4 dp default; 6 dp for computed reals where declared; integers/booleans/ enums as natural JSON types); no NaN/Infinity; null only for mathematically undefined statistics.
• Identifiers: uppercase state codes, portal division/region names verbatim, ISO3 uppercase.
• No narrative, no markdown fences — just the JSON object.

Anti-patterns (do not)

• Do not substitute a map/set for an ordered list the template declares as a list.
• Do not re-sort a positionally-aligned array "for readability".
• Do not zero-fill, mean-impute, or forward-fill missing source data.
• Do not silently swap COUNTY_ROLLUP for DIRECT_SURVEY (or CRUDE for AGE_ADJUSTED).
• Do not change the seed, replicate count, RNG family, or checkpoint schedule.
• Do not round intermediate results; round only the final reported values.
• Do not emit multiple JSON objects, trailing prose, or markdown code fences.
• Do not apply DRAFT or SUPERSEDED methodology docs to publication decisions.

References (deeper detail, read as needed)

• references/portal.md — full endpoint list, dataset schemas/filters, methodology doc map, and the release/revision/quality/suppression/alias resolution rules in full.
• references/modules.md — the six module families: inputs, deterministic execution, RNG, and the exact evidence each must produce.
• references/contract.md — output-contract rules and the pre-submission checklist.