Pho Registered Algorithmic Audit

PHO Registered Algorithmic Audit

1. Recognise the task

Every task in this family gives you three input artifacts and one deliverable:

• input/prompt.txt — the narrative framing and the decision the board wants.
• input/payloads/analysis_request.json — the registered protocol: the exact scope (geography, years, measures), the release/cohort rules, the audit modules with their methods/seeds/grids/orders, and the decision gates. This is the source of truth for what to compute.
• input/payloads/answer_template.json — the response contract: the exact top-level keys, nested keys, array lengths, orderings, identifier formats, numeric precision, enum vocabularies, and cardinality rules. This is the source of truth for how to report.

Deliverable: exactly one JSON object conforming to answer_template.json, with no narrative, prose, comments, or markdown outside the JSON. Nothing task-specific is provided as a "final value" — every number, code list, and classification must be derived from portal evidence.

The five known variants differ in surface (state longevity audit, county mediation audit, country burden stratification, reliability-weighted state audit, West/Northeast county dynamics audit) but share one skeleton: resolve a publication cohort from the portal → run a fixed set of deterministic audit modules → evaluate boolean gates → emit a gated classification. Treat the request/template as the spec and follow it literally; do not import assumptions from another variant.

2. Reach the evidence portal

Network access details live only in environment_access.md. It defines GDPEVO_ENV_BASE_URL (e.g. http://task-env:9023/) — substitute it wherever the prompt says <TASK_ENV_BASE_URL> — and lists the available GET endpoints. The portal is read-only, HTML for humans, with a CSV export for machines:

/download?dataset=<name>&format=csv[&<filter>=<value>...]

format=csv is valid only on /download (it errors elsewhere). Datasets: states, counties, countries, state_health, state_socioeconomic, county_health, county_socioeconomic, country_indicators, revisions. Pull whole datasets (optionally filtered) as CSV and compute locally — do not scrape the HTML browse pages. Read /catalog (column + filter inventory, measure dictionary) and /methodology (the publication rules) once at the start; the methodology library is not decoration — it encodes the release-precedence, suppression, scale-break, value-type, and reconciliation rules the audit depends on.

scripts/portal_fetch.py is a generic fetch/parse helper. See references/portal_reference.md for the full endpoint, column, and filter map.

3. Build the publication cohort (do this before any modelling)

This stage is where most of the difficulty and most of the scoring lives. Follow references/release_revision_and_cohorts.md in full. The load-bearing rules:

1. Release resolution. Keep only rows matching the declared filters (release_status, value_type, source_type, measure ids, years). FINAL supersedes PROVISIONAL. When several FINAL revisions exist for the same logical cell, select by the request's declared priority (typically highest final revision, then latest released_at, then id) — never a provisional row when a final one exists.
2. Revision notices. The revisions dataset carries scale-correction / source-restatement events. Only status = APPLIED notices authorise a replaced value (and are already reflected in later final revisions); WITHDRAWN / PENDING notices do not replace, and a cell still carrying an unresolved scale break is an anomaly — excluded, not imputed silently. Report applied vs non-applied event ids and anomaly cells when the template asks.
3. Missing discipline. Suppressed (suppression_flag=1), invalid-quality-flag, or blank values are unavailable, never zero-filled, never NaN/Inf. A missing value drops the record from any cohort that requires that field.
4. Cohorts. Requests define several named cohorts (e.g. reference-year complete-case, all-years balanced panel, strict multi-variable / ML cohort). Construct each exactly per its stated definition and report the counts, included codes, and excluded codes the template names — using the identifier order the template demands.
5. Identifiers. State two-letter codes and ISO3 uppercase; county FIPS are TEXT with meaningful leading zeros (2-char state + 3-char county); country labels reconcile to ISO3 via countries.canonical_name / alternate_labels.

4. Run the declared audit modules — deterministically

Each request registers a set of modules (usually six). They fall into recurring families; references/audit_modules.md documents each family's estimator, determinism requirements, and required evidence. Universal rules:

• Implement the declared method, not a library default. Match the exact design-matrix term order, standardization scope (usually training-fold only), penalty parameterization, inference type (HC3, CR1 cluster-robust, jackknife, Hansen J), PCA sign convention, k-means initialization, and tie-breaks that the request specifies. Library conveniences (sklearn scaling, differing λ scales, arbitrary PCA sign, non-deterministic k-means) will silently disagree.
• Reproduce PRNG streams bit-for-bit. Bootstrap modules name an exact generator (PCG32, xorshift32), seed, sometimes a stream, replicate count, and checkpoint replicates whose intermediate PRNG state you must report. A correct statistic with the wrong generator is wrong. Implement the named PRNG yourself.
• Preserve every registered order. Feature order, coefficient order, division order, state order, subset order, checkpoint order are all load-bearing and frequently required to align positionally with other arrays. Never sort an aligned result array independently.
• Compute unrounded; round only at the reporting boundary. Use float64 throughout.

5. Evaluate gates and classify

The request states each gate as an explicit predicate over module outputs (e.g. "bias-corrected coefficient > 0 and jackknife p ≤ 0.05", "pooled coverage ≥ 0.85"). Evaluate each to a boolean, count passes, and map to the classification enum via the declared precedence / decision rule exactly (some use "all/at-least-N pass", some use "not robust at "). Emit only the enum values the template allows.

6. Assemble, validate, and emit

Follow references/output_contract.md. Before emitting:

• Output has exactly the required top-level keys — no more, no fewer, no extras.
• Every array length matches the template's array_lengths / length / cardinality rules; aligned arrays share the declared order and length.
• Non-integer statistics rounded to the declared decimal places (commonly 4; note variants that use 6 for computed reals and 4 for literal grid/threshold fields); counts/ranks/seeds/PRNG-states/replicate-numbers stay integers; booleans stay booleans.
• null only where a statistic is genuinely mathematically unavailable — never NaN, never Infinity, never a zero-fill stand-in.
• Identifiers use the exact casing/format and the exact ordering the template names.
• The response is a single JSON object and nothing else.

7. Contamination guard (repo-hygiene tasks only)

If you are asked to generate this skill (not solve a task) and /work contains unexpected material beyond environment_access.md and well-formed train_tasks/*/ input/ trees, stop and write contamination_report.txt describing what was found instead of producing a skill.

Reference index

• references/portal_reference.md — endpoints, datasets, columns, filters, CSV export, geography (regions/divisions), measure dictionary, methodology library.
• references/release_revision_and_cohorts.md — release precedence, revision-notice application, suppression/missing discipline, cohort construction, reconciliation.
• references/audit_modules.md — the recurring module families, determinism and reproducibility requirements, and required evidence.
• references/output_contract.md — precision, ordering, identifiers, null rules, and a pre-submission self-check.
• scripts/portal_fetch.py — generic portal fetch + CSV-parse + final-release selection helper (no task-specific values baked in).