Imported from Prism-Shadow/GDPevo (experiments/claude_code_glm_5_2_max/reports/skills/task_group_021/fewshot/fewshot_attempt_02/SKILL.md). Install upstream with npx skills add Prism-Shadow/GDPevo --skill fewshot_attempt_02. Copyright stays with the author.

Asteria Fleet Data Quality Hub — Reconciliation & Certification Audit

This skill solves the recurring task family: "audit collection X as of cutoff Y against the Asteria Fleet Data Quality Hub and return one certified JSON answer." Every task in the family has the same shape — only the collection, cutoff, focus IDs, thresholds, and output contract change. The procedure below is parameterized by those inputs; it does not depend on any specific answer. Never copy values from another task's answer — re-derive everything for the current task's data.

Inputs (always present in the task directory)

• prompt.txt — narrative of what to audit and report.
• payloads/case_scope.json — the parameters: collection_id, cutoff timestamp, focus/decision IDs, ranking limits and ordering rules, certification thresholds, status→action map, control-case anchors.
• payloads/answer_template.json — the output contract: a JSON Schema (or field contract) with required keys, enums, patterns, minItems/maxItems, ordering rules, and numeric precision. Emit EXACTLY this shape.
• environment_access.md — the only source of network access: base URL, bearer token, and the allow-list of endpoints.

Golden rules (read first)

1. One JSON object, exactly the template. No extra keys, no missing keys, no commentary, no Markdown fences. Templates use additionalProperties: false — every key you emit must be in the schema.
2. Discover, don't assume. Fetch /api/catalog/schema and /api/source-snapshots at runtime to learn field names and snapshot statuses. Do not hardcode today's field names or row counts.
3. The CERTIFIED snapshot wins. Overlapping records across snapshots are reconciled by retaining the occurrence from the authoritative (CERTIFIED) snapshot.
4. Quarantined records are excluded from normalized totals; valid mismatches are INCLUDED. This is the most common accounting error.
5. Control codes are inferred per-ID from evidence, never hardcoded. The allowed code values come from the template; which code applies to which ID is derived from that ID's audit outcome.
6. Respect ordering and precision exactly. Lexicographic ascending, by-ID ascending, rank ascending; round money/quantity to the decimals declared in the template; counts are exact integers.
7. Paginate everything. Collections are explicitly larger than one page.

Procedure

1. Connect

Read environment_access.md. Set BASE_URL and Authorization: Bearer <token>. Use ONLY the endpoints listed there. All reads are GET; the only write-shaped call is POST /api/query, which is a read-only query interface that still requires the bearer credential. See references/endpoints.md.

2. Load parameters and contract

Parse case_scope.json (parameters) and answer_template.json (contract). Note every required key, enum, pattern, minItems/maxItems, ordering rule, and precision declaration. Plan the output object against the contract BEFORE computing.

3. Resolve the source of truth

• GET /api/catalog/collections → find the collection whose stable ID equals case_scope.collection_id.
• GET /api/source-snapshots → list that collection's snapshots. The authoritative snapshot has status CERTIFIED (fall back to the newest non-STALE snapshot if none is CERTIFIED). authoritative_snapshot_id = its stable ID; authoritative_row_count = its row count.
• scoped_raw_row_count = in-scope raw rows (all snapshots within the cutoff; see the per-task contract for the exact denominator).

4. Fetch and scope the raw records

Fetch the relevant record endpoint (/api/transactions/fuel, /api/transactions/freight, /api/maintenance/events, or /api/contacts) and page through every page. Use POST /api/query to filter by collection and cutoff when the raw endpoint is large. Keep only rows whose business timestamp is <= case_scope.cutoff_at (maintenance: within business_period.start..end). raw_row_count = total in-scope raw rows across snapshots.

5. Deduplicate and pick survivors

Group raw rows by their stable logical key (logical transaction / charge / event / contact-cluster ID). Rows sharing a logical key across snapshots are one logical entity. For each group:

• retained occurrence = the one from the authoritative snapshot; others are duplicates.
• duplicate_raw_count = raw_row_count − logical_count. Emit duplicate groups per the contract (charge_id/logical_event_id, snapshot_ids sorted + unique, retained_snapshot_id).

6. Normalize measures

• Units: apply /api/reference/conversions to bring every measure to the canonical unit declared in case_scope (volume→L, weight→KG, distance→KM).
• Currency: apply /api/reference/fx to convert each amount to the base currency (USD) at the rate valid on the record's business date. Keep full precision through the pipeline; round only at emission.

7. Resolve canonical category / class

Map each record's raw description/alias/merchant to a canonical category via /api/reference/aliases. Outcomes per record:

• exactly one recognized → recognized class.
• zero recognized → unrecognized.
• more than one recognized → ambiguous. Unrecognized and ambiguous both mean "cannot be assigned to exactly one recognized category."

8. Classify each logical record

• Quarantine if ANY of: unresolved class (unrecognized OR ambiguous), nonpositive/invalid quantity, nonpositive/invalid weight, nonpositive/invalid distance, or (contacts) no usable contact channel. Track per-reason sub-counts when the contract asks.
• Mismatch (valid) if recognized class ≠ expected class (expected class comes from the record's own declared/expected field). Mismatches are VALID.
• Valid otherwise. invalid_quantity_count, unrecognized_count, ambiguous_count are reported separately when the contract includes them.

9. Exception accounting

An exception = a logical record that is a valid mismatch OR a quarantine. exception_count = mismatch_count + quarantine_count over distinct logical records (no overlap — mismatches are valid, quarantines are not). This drives the rankings.

10. Normalized totals

Sum over valid records only (exclude quarantined; include valid mismatches):

• volume / weight / distance in canonical units, spend in base currency.
• Group by canonical category (fuel_type / service_class) → one row per category, sorted by category ascending.
• total_* = sum of the group totals. Round each emitted number to the decimals declared in the template (typically 2). valid_transaction_count/valid_charge_count = count of valid logical records. See references/pipeline.md for the accounting identities that must hold.

11. Rankings

Apply the ranking policy from case_scope exactly — the limit AND every tie-break:

• Merchants (fuel): exception_count DESC, then merchant_id ASC, limit merchant_ranking_limit.
• Carriers (freight): mismatch_spend_usd DESC (exposure = normalized USD on valid class mismatches only; quarantined charges excluded), then carrier_id ASC, limit carrier_ranking_limit.
• Assets (maintenance): sort keys from case_scope (e.g., rejected_event_count DESC, then regression_event_count DESC, then asset_id ASC), limit from case_scope. Assign rank 1..N ascending. Reversing a tie-break changes the answer.

12. Control / decision code panels

For every public ID listed in a case_scope decision panel (reference alias IDs, transaction/charge IDs, event IDs, focus-cluster IDs, control-case IDs, focus-people anchors), infer the applicable internal code:

• The allowed code values and code field names come from answer_template.json. The human-readable expansions are intentionally NOT supplied — infer from evidence.
• Look up each ID's evidence via POST /api/query and the reference/contacts data, then map its audit outcome (which snapshot/source system it came from, validity, mismatch vs. quarantine reason, channel readiness, consent, record status, field provenance) onto the code enum for that panel.
• Code families you will meet (exact allowed values are in the template): reference-policy RB-*, source-basis/retention SB-*, ledger-disposition/routing LD-*, maintenance-source MS-*, history-route HR-*, identity IC-*, outreach OR-*, field-provenance FP-*.
• Emit one row per scoped ID, sorted by that ID ascending. See references/codes_and_status.md.

13. Certification / close status

Apply the certification rules declared in case_scope:

• Threshold + action-map style (e.g., partner onboarding): compute the gate metric (e.g., quarantine_rate = quarantined_rows / canonical_entities, rounded to the declared decimals), classify PASS / PASS_WITH_EXCEPTIONS / HOLD against status_thresholds, map status → action via status_action_map.
• Gate style (e.g., maintenance odometer regression): a certification_gate declares a condition and its status/action; a triggered gate ⇒ that status/action directly (e.g., any regression ⇒ HOLD / BLOCK_AND_REMEDIATE).
• Default (no explicit rules): infer conservatively — any quarantine / invalid / regression ⇒ HOLD / BLOCK_AND_REMEDIATE; a fully clean audit ⇒ PASS / RELEASE. The status/action pair must be one of the template's allowed enum pairs.

14. Emit and validate

• Build ONE JSON object matching answer_template.json exactly: all required keys, no extras, correct enums/patterns, correct minItems/maxItems, declared ordering, declared precision.
• Run python3 skill/scripts/validate_answer.py <answer.json> <answer_template.json> as a pre-flight sanity check.
• Output only the JSON. No prose, no fences. See references/output_contract.md.

Per-task focus areas

The contract varies; let it drive what you compute. Recurring panels:

• Source/audit summary — counts (raw, logical, duplicate, valid, mismatch, quarantine) + the authoritative snapshot ID.
• Mismatch / quarantine ID lists — complete, deduplicated, lexicographically ascending.
• Normalized totals — per category + grand totals, valid records only.
• Focus rollups — one row per focus asset/person/cluster with survivor + canonical fields + source-system provenance.
• Rankings — top-N with the declared sort and tie-breaks.
• Decision/code panels — one coded row per scoped public ID.
• Status — status + action from the certification rules.

Common pitfalls

• Undercounting because you didn't paginate.
• Including quarantined records in normalized totals (or excluding valid mismatches).
• Retaining the wrong snapshot's occurrence (must be the CERTIFIED/authoritative one).
• Rounding too early, or to the wrong number of decimals.
• Reversed tie-break order in rankings.
• Hardcoding a control code instead of re-deriving it for the scoped ID.
• Emitting extra keys, missing keys, or any text outside the JSON.

Reference files

• references/endpoints.md — endpoint catalog and runtime discovery checklist.
• references/pipeline.md — detailed phases 3–11 with the accounting identities.
• references/codes_and_status.md — control-code inference and certification status.
• references/output_contract.md — ordering, precision, and JSON emission rules.
• scripts/validate_answer.py — pre-flight output sanity check.