Asteria Data Quality Reconciliation

Asteria Data Quality Reconciliation

Start Every Case

1. Read environment_access.md, payloads/case_scope.json, and payloads/answer_template.json.
2. Use only the hub endpoints listed in environment_access.md. Prefer authenticated POST /api/query for complete data; REST endpoints may be page-limited.
3. Query v_source_snapshots, the relevant domain view, v_reference_aliases, v_unit_conversions, and v_fx_rates as needed. Check truncated; if true, page explicitly with limit/offset or narrower queries.
4. Fill exactly the answer template: no extra keys, exact required names, numeric precision, sorted arrays, and JSON only.

Useful query pattern:

curl -sS -X POST "$GDPEVO_ENV_BASE_URL/api/query" \
  -H "Authorization: Bearer <token>" \
  -H "Content-Type: application/json" \
  -d '{"query":"select * from v_source_snapshots where collection_id = '''<collection_id>''' order by snapshot_id"}'

Source Retention

• Identify the authoritative snapshot from v_source_snapshots: use the CERTIFIED snapshot for the scoped collection/cutoff. Treat PROVISIONAL as fill evidence only when a logical record has no certified occurrence. Treat STALE as non-authoritative unless the prompt explicitly asks for stale coverage.
• For overlapping logical records, group by the public logical ID:
  • contacts: row_id is a source row, not the logical person ID; cluster separately.
  • fuel: transaction_id.
  • freight: charge_id.
  • maintenance: event_id.
• Retain the certified occurrence when present. Otherwise retain the latest non-stale occurrence by business update/ingest timestamp. Report duplicate groups from all raw occurrences, with snapshot IDs sorted lexicographically.
• raw_row_count is the in-scope source row count. logical_*_count is the count after grouping by logical ID. duplicate_raw_count is raw rows - logical records unless the contract defines a different duplicate measure.

Alias Recognition

• Use aliases whose business date is within valid_from and valid_to when valid_to is present. Exclude PROVISIONAL aliases from automatic recognition. Closed INACTIVE aliases may still be historically applicable inside their valid interval.
• Match aliases case-insensitively with Unicode normalization and token boundaries. Do not let a shorter alias inside a longer matched alias create false ambiguity: keep the longest non-overlapping spans, then collect distinct canonical values.
• Classification:
  • zero recognized canonical values: unrecognized alias/class.
  • more than one distinct canonical value: ambiguous alias/class.
  • exactly one canonical value: recognized; compare it to the expected fuel type or service class for mismatch reporting.
• Count/report unrecognized and ambiguous records as quarantine/unrecognized where the contract says unresolved category/class cannot enter the ledger.

Normalization

• Use v_unit_conversions by kind and effective date. Apply the factor to reach the requested canonical unit.
• Use only CERTIFIED FX rows from v_fx_rates for the transaction/service date and currency. USD spend is amount * usd_per_unit.
• Exclude quarantined records from normalized totals. Include valid category/class mismatches in normalized totals unless the prompt says otherwise.
• Round only final reported totals to the template precision, usually 2 decimal places for money, volume, weight, distance, and odometer-derived totals.

Fuel Purchases

• Work on retained logical transactions from v_fuel_transactions.
• Quarantine a retained transaction when its recognized category is zero/ambiguous or quantity is missing/nonpositive.
• unrecognized_transaction_ids includes both zero-match and ambiguous-category transactions when the contract describes "no unique recognized category."
• Mismatches are retained transactions with exactly one recognized category different from expected_fuel_type; if the contract says valid mismatches, exclude quarantined transactions from the mismatch ID list.
• Normalized liters use kind='volume'; normalized spend uses certified FX on date(purchased_at).
• Merchant exception ranking counts distinct retained transactions with a mismatch or quarantine condition. Sort by exception count descending, then merchant ID ascending.

Freight Charges

• Work on retained logical charges from v_freight_charges.
• Quarantine a retained charge when service class is zero/ambiguous, billed weight is missing/nonpositive, or distance is missing/nonpositive.
• class_mismatch_charge_ids are valid recognized charges where recognized service class differs from expected_service_class.
• quarantine_reason_counts should partition quarantined logical charges into ambiguous_alias, unrecognized_alias, invalid_weight, and invalid_distance. Prefer alias reasons before physical-measure reasons only when a charge has multiple unresolved conditions; otherwise use the single observed condition.
• Normalize billed_weight with kind='weight', distance with kind='distance', and spend with certified FX on service_date.
• Carrier ranking exposure is normalized USD only from valid service-class mismatches. Sort by mismatch spend descending, then carrier ID ascending. exception_count is the union of valid mismatches and quarantines.

Maintenance Events

• Work on retained logical events from v_maintenance_events; group duplicate occurrences by event_id.
• Reject events with missing timestamp, unparsable timestamp, negative/missing odometer, negative labor, or extreme labor. In the Asteria maintenance sets, extreme labor is the synthetic high outlier value around 120.0 hours; use a practical daily threshold such as > 24 hours unless the prompt provides one.
• issue_counts count retained logical events by reason; a single event may increment multiple issue counts. invalid_event_ids is the sorted unique union of events rejected for field validity. Do not include sequence-only odometer regressions in invalid_event_ids.
• Convert odometers with kind='odometer'. For regression checks, sort valid events by asset_id, parsed event time, then event ID. A regression is an event whose odometer is lower than the previous reliable odometer for that asset; flag it and do not update the previous reliable odometer with the regressed reading.
• Corrected distance is the sum over assets of last reliable odometer_km - first reliable odometer_km in the reconstructed period. Exclude rejected field-invalid events and sequence regressions from reliable-distance endpoints.
• Asset risk rankings usually sort by rejected event count descending, then regression event count descending, then asset ID ascending, unless case scope overrides.

Contact Survivorship And Readiness

• Normalize emails with Unicode NFKC, trim, and lowercase. Treat blank, N/A, NA, none, and NULL as unusable.
• Normalize phones by removing non-digits; treat blank placeholders as unusable. For North American numbers with a leading 1 country code, compare on the last 10 digits when identifying duplicates, but report the canonical phone as digits only from the selected source value.
• Build person clusters from strong identifiers:
  • Merge rows sharing a normalized usable email.
  • Merge rows sharing a usable phone only when the phone is not a shared/helpdesk-style identifier across multiple distinct names/emails.
  • Do not merge on name alone, noisy/shared master_hint, or a phone shared by many different people. Report those as contested/no-automerge when requested.
• Quarantine source rows with no usable email and no usable phone.
• Choose canonical fields by field-level precedence: prefer verified/certified evidence, majority-agreeing values, authoritative identity/compliance sources, then latest business update. Preserve Unicode; normalize display casing only when needed. Report the source system that supplied each canonical field.
• A contact/person is dispatchable or readiness-eligible only when active and retaining at least one usable email or phone. A channel is ready only when consent is GRANTED.
• Readiness partitions:
  • both: active, consent granted, usable email and phone.
  • email_only: active, consent granted, usable email only.
  • phone_only: active, consent granted, usable phone only.
  • not_ready: active with usable contact but non-granted consent.
  • Inactive people are not readiness-eligible; report them in inactive exclusions or blocked inactive counts when required.

Compact Code Semantics

Use these reusable mappings unless the current hub evidence clearly contradicts them:

• Reference policy: RB-83 active/effective authoritative alias; RB-42 inactive, superseded, future, or otherwise not effective for the business date; RB-17 provisional/untrusted reference row.
• Source basis/retention: SB-79 certified singleton retained; SB-61 duplicate overlap retained from certified over lower-trust occurrence; SB-24 provisional-only fill retained.
• Ledger disposition: LD-14 valid match/accrue; LD-31 valid mismatch/review; LD-53 unrecognized alias/class quarantine; LD-72 ambiguous alias/class quarantine; LD-88 invalid physical measure or quantity quarantine.
• Maintenance source: MS-86 certified singleton retained; MS-47 duplicate overlap retained from certified; MS-12 provisional-only retained.
• Maintenance history route: HR-74 accepted reliable history event; HR-33 sequence-only odometer regression; HR-19 rejected for invalid field values.
• Identity controls: IC-90 high-confidence automerged identity cluster; IC-70 unique usable single-source identity; IC-40 weak/no-identifier or name-only evidence not automerged; IC-25 contested shared identifier/no automerge.
• Outreach controls: OR-80 active, consent granted, both email and phone ready; OR-60 active, consent granted, exactly one channel ready; OR-35 usable channel but blocked by non-granted consent; OR-15 no usable outreach path or inactive exclusion.
• Field provenance controls: FP-75 multi-source field-level precedence applied; FP-55 accepted single-source field provenance; FP-20 poor/unusable/quarantined field provenance.

Status Decisions

• Apply explicit thresholds, action maps, or certification gates from case_scope.json first.
• If no explicit map exists: PASS only when there are no exceptions; PASS_WITH_EXCEPTIONS when only reviewable valid mismatches remain; HOLD when quarantines, rejected invalid records, unresolved identifiers, or explicit gate failures remain.
• Use the exact action/routing enum paired with the chosen status: release for pass, review for pass-with-exceptions, block/remediate for hold.

Final Checks

• Recompute summary counts from the same retained population used for detail arrays.
• Confirm partition counts sum where the template says they must.
• Sort every ID list and ranked array exactly as specified by the answer contract.
• Validate JSON parseability and, when practical, validate against the provided JSON Schema or field contract before returning.