| name | pho-algorithmic-transport-audit |
| description | Solve Public Health Observatory registered algorithmic-transport audit requests — a prompt plus analysis_request.json and answer_template.json whose evidence comes only from the Observatory Web portal. Use whenever a PHO audit task asks you to resolve portal publications/cohorts, run a declared set of reproducible statistical audit modules (delete-cluster/jackknife fixed effects or GMM/mediation, nested ridge or elastic-net CV, restricted-null wild cluster bootstrap, grouped split conformal, trajectory PCA + deterministic k-means + stability, source/year/group perturbation, partial-R2 sensitivity), and return exactly one JSON object conforming to the template. Carries method only; binds every task-local value from the effective request and never carries solved values across invocations. |
Public Health Observatory Algorithmic-Transport Audit
When to use
Activate when a request consists of:
- a prompt directing you to the Observatory Web portal at
<TASK_ENV_BASE_URL> (resolved from environment_access.md),
- an
analysis_request.json declaring the audit: a protocol_id or request_id, scope (geography, years, reference year), cohorts, an ordered set of audit modules, reproducibility bindings, a reporting block, and a decision_rule,
- an
answer_template.json fixing the output contract,
and asks for exactly one JSON object computed only from portal evidence, with no narrative.
This covers state-level, county-level, and country-level PHO audits (longevity/transport, mediation/transport, robustness/transport, panel/transport, burden-revision). Do not activate for unrelated tasks, and do not activate on family membership or subject-matter similarity alone — a declared protocol id, when present, is the exact activation key.
Core discipline (read first)
This skill is REUSABLE_METHOD_ONLY. It carries method semantics, not instance values.
- Bind everything task-local from the effective request: entities, measures, fields, time coordinates, geography, source filters, random seed/stream/replicate schedule, hyperparameter grids, reporting cutoffs, output names, and business predicates. Never invent these from memory or carry them from any prior solved answer.
- Never carry solved analytical values across invocations. Recompute every coefficient, p-value, count, and label from the effective request and authorized portal evidence. Any solved standard answer present in the workspace is provenance only — it is not solver-visible input, is not expanded or required by the answer template, and is ignored by the evaluator.
- One effective contract. Resolve a single effective request before any data access, fold, random draw, fit, aggregation, or decision, and use it in every module.
- Evidence only from the portal. Reach the network only via
environment_access.md (base URL + allowed GET endpoints). See references/portal_evidence.md.
- No contamination. If the working directory contains material outside the expected request files, stop and report rather than solving.
Procedure
- Inventory and guard. Confirm the working directory holds only the expected request files (the prompt,
analysis_request.json, answer_template.json, and environment_access.md). Stop and write a contamination report if anything unexpected is present.
- Resolve the effective request (
references/effective_request.md). Verify the exact protocol id when declared; fold direct keys and <section>_overrides / module_overrides.<module> / reporting_overrides aliases; deep-merge objects, replace whole arrays, replace explicit scalars/strings/booleans/null only at their paths; inherit absent paths; reject unknown targets, array concatenation/positional patching, inferred aliases, key renaming, and type coercion. Freeze one effective contract.
- Read the contract. Parse
answer_template.json for required top-level keys, per-section required keys, array lengths and orderings, cardinality rules, enum/boolean types, and precision. The template is authoritative for field names and structure.
- Pull evidence (
references/portal_evidence.md). Fetch only the portal endpoints you need; resolve publications and geography; never zero-fill.
- Execute modules in declared order. Run release/cohort resolution first, then each declared audit module, then the controlled decision. Use the shared numerical conventions in
references/numerical_conventions.md and the module catalog in references/module_families.md. Preserve every declared order.
- Decide. Evaluate every gate on unrounded values, count passes, and apply the request's decision rule (count-threshold classification or first-failed-module precedence). See
references/numerical_conventions.md § Decision.
- Emit (
references/reporting_contract.md). Produce exactly one JSON object conforming to the template. Round reported non-integers to the declared precision; keep integers/booleans as natural JSON types; use JSON null only when a statistic is mathematically unavailable (never NaN/Infinity); preserve every declared list order; no narrative.
What is reusable vs. task-local
| Reusable (this skill) | Task-local (bind from the request) |
|---|
| override-resolution algorithm | exact protocol_id string |
| release-resolution priority form | analysis years, reference year, geography scope |
| cohort taxonomy & construction rules | which cohorts are requested and their completeness predicates |
| module-family methods (FE/jackknife, ridge/elastic-net, wild bootstrap, conformal, PCA+kmeans+ARI, perturbation, mediation/GMM, sensitivity) | which modules, their feature/coefficient/instrument order, grids |
| PRNG families (PCG32, xorshift32) & plus-one p | chosen PRNG, seed, stream, replicate count, checkpoint schedule, quantile probabilities |
| numerical conventions (CR1/HC3/WLS, RMSE pooling, rank intervals, ARI) | cutoffs, thresholds, classification vocabulary |
| reporting discipline | numeric precision (4 or 6 dp), exact field names |
References
references/effective_request.md — protocol gating and override mergereferences/portal_evidence.md — portal access, data model, release resolution, cohortsreferences/numerical_conventions.md — shared math across all modules + decisionreferences/module_families.md — the recurring audit modules and their registered variantsreferences/reporting_contract.md — precision, ordering, null, JSON shape
