Asteria Fleet Data Quality Reconciliation

Asteria Fleet Data Quality Hub Reconciliation

A family of tasks asks you to audit one Asteria Fleet collection, reconcile its overlapping source records as of a stated business cutoff, and return a single JSON object matching an answer_template.json. The five collection families are:

• contacts (v_contacts) — partner onboarding, dealer, warranty, field-service rosters
• fuel (v_fuel_transactions) — fuel/charging purchases
• freight (v_freight_charges) — carrier invoice charge lines
• maintenance (v_maintenance_events) — work-order events

Every collection is published as 2–3 snapshots (CERTIFIED + PROVISIONAL, sometimes a second CERTIFIED source). Overlapping business records appear in more than one snapshot. Your job is to collapse duplicates to logical records, classify data-quality issues, normalize units/currency, infer opaque control codes, and certify.

This skill gives the procedure and the decoded reconciliation rules. It does NOT contain candidate answers or task-specific final values — apply the rules to the live hub.

1. Inputs you are given

For each task directory train_tasks/<task>/input/:

• prompt.txt — the business framing.
• payloads/case_scope.json — the authoritative specification: collection_id, business cutoff, focus seeds/anchors, decision-panel IDs, thresholds, sort orders. Treat this as the source of truth for parameters.
• payloads/answer_template.json — the JSON-Schema the answer must match exactly (required keys, enums, patterns, item counts, ordering rules). Read it first.
• environment_access.md — hub base URL + bearer token + allowed endpoints.
• judge_access.md — a train-only judge endpoint. Ignore it when solving; it is not part of the procedure and must not be wired into test solving.

2. Hub access pattern

The hub exposes two query styles. Prefer SQL for bulk work:

Read-only SQL (POST /api/query, body {"query": "<SQL>"}):

• Returns {"columns":[...], "rows":[[...]], "truncated":bool}.
• SQLite-flavored. Use single-quoted string literals (e.g. WHERE collection_id='x'). Parameter binding via ?/params is NOT supported.
• Paginate with LIMIT n OFFSET m. truncated:true means more rows exist.
• Whitelisted views only: v_contacts, v_fuel_transactions, v_freight_charges, v_maintenance_events, v_fx_rates, v_reference_aliases, v_unit_conversions, v_source_snapshots. sqlite_master, information_schema, SHOW, DESCRIBE, LIKE, and LOWER() are rejected.
• All * selects work: SELECT * FROM v_freight_charges WHERE collection_id='...'.

REST (GET, header Authorization: Bearer <token>):

• /api/catalog/collections — collection metadata (no filter).
• /api/catalog/schema — the 8 views + field meanings (no filter).
• /api/source-snapshots?collection=<collection_id> — snapshot list. Filter key is collection, not collection_id. (Other keys return {"error":"invalid filter"}.)
• /api/contacts — contacts (paginated).
• /api/transactions/fuel, /api/transactions/freight, /api/maintenance/events — raw rows.
• /api/reference/aliases?domain=<fuel|freight> — alias→canonical map.
• /api/reference/conversions?kind=<volume|distance|weight|odometer> — unit factors.
• /api/reference/fx — FX rates (no filter).

Discovery order that always works: read case_scope.json → GET schema (views+fields) → GET snapshots (?collection=) → SQL-query the relevant view filtering by collection_id → fetch aliases/conversions/fx as needed.

3. Hub data model — field decoding

Snapshots & source systems

Each snapshot carries snapshot_id, snapshot_status (CERTIFIED > PROVISIONAL > STALE), source_system, row_count, business_cutoff, created_at, ingested_at, checksum. Transaction/event/charge views carry snapshot_id but not source_system — join via v_source_snapshots. Contact views DO carry source_system.

Authoritative snapshot = the one with the best snapshot_status (CERTIFIED first). When two snapshots are both CERTIFIED (contacts), prefer by tie-break in the case scope or by snapshot_id ascending. PROVISIONAL is supplementary feed data.

Overlap rule (the core reconciliation)

The same business record appears in multiple snapshots. The only field that differs between duplicate occurrences is typically ingested_at (and for maintenance, event_status/ingested_at; for contacts, the whole row is a fresh source-system view). Collapse by the business key:

• fuel: transaction_id
• freight: charge_id
• maintenance: event_id
• contacts: by identity key (see §6), not by row_id (each snapshot invents fresh row_ids)

raw_row_count − logical_count = duplicate_raw_count. The authoritative occurrence is retained; the others are dropped. For contacts, every source-system row is retained as a member of the cluster.

Aliases (category resolution)

v_reference_aliases (domain, alias_id, alias_text, canonical_value, valid_from, valid_to, reference_status, published_at). Map a free-text description (purchased_description / description) to a recognized canonical category by matching alias_text substrings (case-insensitive) in the description.

• An alias is effective as-of the business cutoff only if valid_from <= cutoff_date AND (valid_to is null OR valid_to >= cutoff_date) AND reference_status != INACTIVE.
• The deliberate trap is a single alias_text that has more than one reference row (e.g. an INACTIVE/expired one plus an ACTIVE one with later valid_from, or a short generic token like priority used as a catch-all). A naive reference_status=ACTIVE filter picks the wrong row when the cutoff falls before the new row's valid_from. The ONLY safe selection is the row effective as-of the cutoff per the rule above. Always GROUP alias rows by alias_text and check every row's validity window before deciding.
• PROVISIONAL alias rows are candidate mappings not yet adopted — treat a description whose only effective match is a PROVISIONAL alias as not-yet-recognized (do not promote it to a canonical category for reconciliation). Scan reference_status for PROVISIONAL rows in the live alias set rather than assuming any named text.
• A description that matches alias_texts leading to two distinct canonical values = ambiguous → unrecognized. A description with no effective alias match = unrecognized.
• Resolution detail: if all matched alias_texts agree on one canonical value (even several alias_texts), that is a unique match, not ambiguous (e.g. "ULSD road diesel" matches both "ulsd" and "road diesel" but both → DIESEL → unique).

Unit conversions (v_unit_conversions, kind/from_unit/to_unit/factor/precision)

value_canonical = value_raw * factor. Canonical units: L (volume), KM (distance & odometer), KG (weight).

• volume: US_GAL→L 3.785411784, IMP_GAL→L 4.54609, L→L 1.0
• distance: MI→KM 1.609344, KM→KM 1.0
• odometer: MI→KM 1.609344 (precision 1), KM→KM 1.0
• weight: LB→KG 0.45359237, KG→KG 1.0 Round volume/weight/distance totals to 2 decimals; the precision field is the per-value display precision, not the aggregate rounding.

FX (v_fx_rates, rate_date/currency/usd_per_unit/rate_status/published_at)

usd_amount = amount * usd_per_unit. USD's own rate is 1.0.

• Match the rate by currency + rate_date == the transaction's business date (purchased_at/service_date date portion). If the exact date is missing, use the latest prior available date.
• Prefer the CERTIFIED rate over PROVISIONAL for the same currency+date — when two rows share a currency+date, keep the one with the better rate_status (CERTIFIED > PROVISIONAL) and, on a tie, the later published_at. Verify the live v_fx_rates rows rather than assuming any fixed publish time.

4. Universal certification gate

Status/action mapping (always present in case_scope as status_action_map or status_thresholds/certification_gate):

• PASS → RELEASE
• PASS_WITH_EXCEPTIONS → REVIEW_EXCEPTIONS
• HOLD → BLOCK_AND_REMEDIATE

Thresholds (contact tasks) use quarantine_rate = quarantined_rows / canonical_entity_count rounded to 4 decimals: PASS if <= pass_max_quarantine_rate (usually 0.0); PASS_WITH_EXCEPTIONS if <= pass_with_exceptions_max_quarantine_rate (usually 0.04); else HOLD.

Transaction/maintenance gates: the case_scope states the exact gate (e.g. maintenance's certification_gate directly says HOLD/BLOCK_AND_REMEDIATE; freight/fuel infer HOLD when invalid physical measures exist, else PASS_WITH_EXCEPTIONS if any mismatch/quarantine, else PASS). When the scope states the gate explicitly, use it verbatim.

5. Reconciliation per family

Fuel & freight (parallel pipeline)

1. Collapse snapshots → logical records (retain CERTIFIED occurrence per business key).
2. Resolve each description → recognized canonical category via effective aliases.
3. Classify per logical record:
   • mismatch: recognized category (unique) ≠ expected_fuel_type/expected_service_class.
   • unrecognized: no or ambiguous alias match → also quarantined.
   • invalid quantity/measure: quantity <= 0 (fuel), billed_weight <= 0 or distance <= 0 (freight) → quarantined.
   • quarantine = unrecognized OR invalid measure. Quarantined records are EXCLUDED from normalized totals; valid mismatches ARE included in totals.
4. Normalize: volume→L, weight→KG, distance→KM, amount→USD (certified FX).
5. Aggregates: per-category totals (sorted by category), total volume/weight/distance/spend, focus-asset rollups, merchant/carrier exception ranking (exception_count = distinct logical records with a mismatch OR quarantine; sort by exception_count desc then id asc for merchants; for carriers by mismatch_spend_usd desc then carrier_id asc — mismatch spend = USD on valid class-mismatch records only).
6. Duplicate groups (freight: one object per charge_id with >1 raw occurrence, retained snapshot = authoritative).

Maintenance

1. Collapse snapshots → 1 retained row per event_id (retain CERTIFIED).
2. Issue counts (row-level unless scope says logical): missing event_time_raw (null/blank), invalid timestamp (unparseable, e.g. 2026-99-45 25:61), invalid odometer (<=0), negative labor (<0), extreme labor (a sentinel cap stated in the case_scope; flag any labor value at/above it), odometer regression (later event's odometer < earlier event's max for the same asset, chronological by time). Pull the concrete labor cap from the case_scope rather than assuming a fixed value.
3. invalid_event_ids = events rejected for missing/unparseable time or invalid odometer/labor range. Sequence-only regressions go in corrected_metrics, not invalid_event_ids.
4. corrected_metrics.total_distance_km = Σ over assets of (last reliable odometer − first reliable odometer) in the reconstructed Q1 history, odometer converted to KM, kept monotonically non-decreasing (drop regressed readings), 2 decimals.
5. asset_risk_ranking: sort by rejected_event_count DESC, regression_event_count DESC, asset_id ASC; top 5.
6. duplicate_groups: one per event_id in >1 snapshot — logical_event_id, all snapshot_ids (sorted), retained_event_id, retained_snapshot_id (authoritative).
7. event_decision_panel: one coded decision per scoped event_id (sorted asc).

Contacts (partner/dealer/warranty/field-service)

1. Build clusters of the same logical person across source systems using the identity key: the numeric token in the clean @example-fleet.com email local-part (e.g. sofia.smith.0 → token 0; token is stable across all 3 source-system rows of one person). Rows whose email is a placeholder (null/none/n/a/blank) or a .part.NNN/ .fiel.NNN @mail.example decoy do NOT share a token — treat each as its own single-source identity (these are deliberately-distinct decoys, NOT the 3rd snapshot of a clean person).
2. Canonical fields per cluster: pick by survivorship precedence — prefer authoritative snapshot, then verified_flag=1, then source priority (contacts: Compliance Master > Partner Portal/HR Directory > CRM/Dispatch; field-service: HR Directory > Identity Registry > Dispatch). Email = NFKC-trim-lowercase. Phone = digits only, stripping spurious country codes (+1/+44/+21 leading) to the canonical 10-digit form. City/name = trimmed NFKC; name title-cased.
3. Quarantine = cluster has no usable email AND no usable phone.
4. Channel readiness (over readiness-eligible = ACTIVE + ≥1 usable email/phone): bucket by which channels are usable AND consent is GRANTED: both / email_only / phone_only / not_ready (consent not GRANTED).
5. Contested / no-automerge: a shared identifier (any master_hint value shared by

   1 distinct identity, or a phone number shared across distinct identities, or any field the case flags as a shared business line) is contested → CONTESTED_NO_AUTOMERGE. Same name+phone across sources that genuinely disagree = do not merge. Detect these by scanning the live data: any identifier value used by rows belonging to >1 identity key is contested — do not hardcode specific hint strings.

6. Opaque control codes — inference

Codes are not documented in the task materials; infer them from record characteristics. The allowed enum values come from each task's answer_template.json. Use these mappings as the starting interpretation and confirm against the enum set:

Contacts (IC/ OR/ FP):

• identity_code (IC-25/40/70/90): single-source identity → IC-25; 2 merged sources → IC-40; 3+ merged sources → IC-70; contested/conflicting identity (shared identifier across distinct people) → IC-90.
• outreach_code (OR-15/35/60/80): consent status GRANTED→OR-80, PENDING→OR-60, DENIED→OR-35, UNKNOWN→OR-15.
• field_provenance_code (FP-20/55/75): authoritative source verified_flag=1 → FP-75; ≥2 contributing source systems → FP-55; single unverified source → FP-20.

Fuel/freight (RB / SB / LD):

• reference_policy_code / reference-row decision (RB-17/42/83): ACTIVE alias effective as-of cutoff → RB-83; ACTIVE but not-yet-effective-as-of-cutoff OR PROVISIONAL → RB-42; INACTIVE/expired → RB-17. (The reference set is seeded so that across the scoped alias_ids each of RB-17/42/83 appears at least once — apply the as-of rule per row to distribute them, do not hardcode.)
• source_basis_code / source-retention (SB-24/61/79): retained occurrence from the CERTIFIED/authoritative snapshot → SB-79; from PROVISIONAL → SB-24.
• ledger_disposition_code / ledger-routing (LD-14/31/53/72/88): quarantined → LD-14; valid class mismatch → LD-72; REVIEW record_status → LD-53; non-USD currency → LD-31; clean POSTED/BILLED USD → LD-88.

Maintenance (MS / HR):

• maintenance_source_code (MS-12/47/86): retained from authoritative snapshot → MS-86; from provisional → MS-47; (MS-12 reserved for the third/legacy source when present).
• history_route_code (HR-19/33/74): rejected/invalid event → HR-19; regression event → HR-33; clean sequenced event → HR-74.

When a record could match two rules, prefer the most specific / most severe applicable code, and always pick a value from that field's enum.

7. Output discipline

• Produce one JSON object matching answer_template.json exactly: every required key present, no extra top-level keys beyond the template, every array at its minItems/maxItems, every value from its enum, every string matching its pattern.
• Sort/Ordering is graded — follow each field's description/x-ordering_rules: IDs lexicographically ascending; cluster/asset arrays by their stated sort key; ranking arrays by rank ascending. Member-row-id lists are deduped + lexicographically sorted.
• Numeric precision: counts are exact integers; currency/volume/distance/odometer rounded to the stated decimals (usually 2; quarantine_rate to 4). No floats where the contract says integer.
• Stable IDs only: use IDs present in the live hub or supplied in the case_scope. Never invent canonical classes, snapshot IDs, or codes outside the enums.
• No commentary, no Markdown wrapper — the answer is bare JSON.

8. Transferable procedure (run this for every task)

1. Read case_scope.json + answer_template.json fully. Note item counts, enums, sort rules, the cutoff date, and the certification gate/thresholds.
2. GET /api/catalog/schema; identify the family's view and its fields.
3. GET /api/source-snapshots?collection=<collection_id>; pick the authoritative snapshot; record authoritative_snapshot_id.
4. SQL-dump the family view WHERE collection_id='<id>' (paginate). Fetch aliases (?domain=), conversions (?kind=), fx as the family needs.
5. Collapse snapshots → logical records by business key; compute duplicate_raw_count.
6. Apply family-specific classification (mismatch/unrecognized/invalid/quarantine; issue_counts; regressions).
7. Normalize units + FX; aggregate totals, per-category rows, focus rollups, rankings.
8. Infer control codes per the §6 mappings for every scoped decision-panel ID.
9. Compute the certification status/action from the scope's gate/thresholds.
10. Assemble the JSON to the template, enforce ordering & precision, and emit bare JSON.

Supporting reference tables live next to this file in reference/.