Asteria Fleet Data Quality Hub

Asteria Fleet Data Quality Hub — reconciliation & certification

Every task in this family is the same pipeline over a different collection and output contract. Recognize it, run the pipeline, then bend the output to the exact answer_template.json in front of you. Do not invent numbers: every value must be derived from the live hub as of the scoped cutoff.

0. Inputs you are always given

• payloads/case_scope.json — the only source of scope: collection id, the business cutoff / as-of / period, focus items (clusters, assets, events, people, charges, transactions), anchored control cases with pinned evidence rows, ranking limits and tie-breaks, canonical units / base currency, status thresholds and a status→action map.
• payloads/answer_template.json — the output contract (usually JSON Schema, sometimes a field_contract document). It is authoritative for keys, enums, array cardinality (minItems/maxItems), ordering, rounding, and additionalProperties:false. Read it last-mile and satisfy it literally.
• environment_access.md — the only thing that grants network access: the base URL (*_BASE_URL=...) and an AUTHORIZATION: Bearer <token> header, plus the list of allowed endpoints. Send the auth header on every request.

Never hardcode a token, base URL, count, or code mapping from a previous task — they change per task. Read them fresh at runtime.

1. Connect and discover (never assume the API shape)

Use scripts/asteria_client.py (stdlib-only) to read environment_access.md and make authenticated calls, or replicate its logic. Before pulling data:

1. GET /api/catalog/collections — confirm the scoped collection_id exists and learn its domain/metadata.
2. GET /api/catalog/schema — learn the real field names and types for rows, snapshots, and reference tables. Do not guess field names from this skill; the schema is ground truth.
3. GET /api/source-snapshots — snapshot metadata: ids, status (e.g. CERTIFIED / PROVISIONAL / STALE), timestamps, coverage.
4. Inspect one page of the domain endpoint and one POST /api/query response to learn the pagination envelope and the query request/response shape, then adapt get_all() / query() accordingly. Collections span multiple pages — page to exhaustion; never reason from a truncated first page.

Endpoints seen in this family (subset appears per task): /api/catalog/collections, /api/catalog/schema, /api/contacts, /api/transactions/fuel, /api/transactions/freight, /api/maintenance/events, /api/reference/aliases, /api/reference/conversions, /api/reference/fx, /api/source-snapshots, POST /api/query.

2. The reconciliation pipeline (shared spine)

Run these stages in order; every task uses a subset, described in its prompt.

A. Scope & pull. Restrict to the scoped collection_id. Pull all raw rows across snapshots (paginate). raw_row_count = in-scope raw source rows (including cross-snapshot duplicates).

B. Cutoff filter. Keep only rows whose business/effective date is <= the scoped cutoff / as-of / within the period. Use the business-date evidence the prompt names, not ingestion time.

C. Authoritative snapshot & de-duplication. Rows overlap across snapshots. Choose the authoritative snapshot from snapshot metadata (prefer certified/ current status and the latest as-of within cutoff — verify the rule from the schema/metadata, don't assume). Collapse raw rows into logical entities (logical transaction / charge / event / person / contact) by their stable business key. For each duplicate group the retained occurrence comes from the authoritative/retained snapshot. Report duplicate_raw_count = raw − logical and, where required, per-group snapshot_ids (sorted) + retained_snapshot_id.

D. Normalize. Depending on domain:

• Text/contacts: canonical name (Unicode-preserving), email = trimmed NFKC lowercase, phone = digits only (kept as a string), canonical city, canonical region/depot value.
• Fuel/freight: resolve descriptions/aliases to a recognized canonical category/class via /api/reference/aliases; convert quantities to the canonical unit via /api/reference/conversions; convert money to the base currency via /api/reference/fx using the applicable business-date rate.
• Maintenance: convert odometer/labor to declared units; reconstruct per-asset history ordered by reliable timestamp.

E. Classify: valid vs mismatch vs quarantine. Quarantine = a row that cannot enter normalized totals. Reason families observed:

• unrecognized — no recognized category/alias match.
• ambiguous — matches more than one category/alias.
• invalid_quantity / invalid_weight / invalid_distance — non-positive or out-of-range physical measure.
• no usable contact channel — no usable email or phone (contacts/roster).
• maintenance rejects — missing/unparsable timestamp, invalid odometer range, negative or extreme labor. A mismatch (expected category/class ≠ recognized category/class) is still a valid record: it counts in normalized totals and is flagged separately. Quarantined rows are excluded from all normalized totals; valid mismatches are included. An "exception" is a distinct logical record that is a valid mismatch or is quarantined.

F. Canonical resolution & survivors. Group logical entities that are the same real-world entity (identity resolution). Pick a survivor / master_id (a real public row id) and build canonical field values. When sources disagree per field, apply field-level source precedence (discover the precedence order from schema/reference — e.g. CRM vs Compliance Master vs Partner Portal, or HR Directory vs Dispatch vs Identity Registry) and record which source won each field. Entities with genuinely conflicting identity evidence on a watchlist are contested → not auto-merged (CONTESTED_NO_AUTOMERGE). member_row_ids is the deduplicated, lexicographically-sorted set of contributing public row ids.

G. Rollups, rankings, readiness.

• Rollups (region/depot, fuel_type, service_class): one row per represented value from the contract's enum, canonical-entity/charge counts, sorted ascending by the key.
• Rankings (merchant / carrier / asset): sort by the primary metric descending then the declared tie-breaks (usually id ascending); truncate to the scope's limit; assign rank from 1.
• Readiness (contacts/roster): an entity is readiness-eligible only when it is active AND retains at least one usable email or phone. A channel is ready only when consent is granted. Partition eligible entities into mutually exclusive buckets (both / email_only / phone_only / not_ready) that sum to the eligible count. "Dispatchable" = active + usable channel + consent granted; blocked reasons (consent / no-contact / inactive) partition the depot total.

H. Control codes (opaque; expansions withheld on purpose). See §3.

I. Certification / status. Compute the gate metric — usually quarantine_rate = quarantined_rows / canonical_entities, rounded to 4 dp — and apply the scope's thresholds: at-or-below pass_max → PASS; else at-or-below pass_with_exceptions_max → PASS_WITH_EXCEPTIONS; else HOLD. Honor any hard gate the scope names (e.g. any odometer regression forces HOLD / BLOCK_AND_REMEDIATE) — a hard gate overrides the rate. Map status → action with the scope's status_action_map (typically PASS→RELEASE, PASS_WITH_EXCEPTIONS→REVIEW_EXCEPTIONS, HOLD→BLOCK_AND_REMEDIATE).

J. Assemble & self-check. Build the object to the contract, then run §4.

3. Assigning the opaque control codes

Codes like IC-25/40/70/90, OR-15/35/60/80, FP-20/55/75, RB-17/42/83, SB-24/61/79, LD-14/31/53/72/88, MS-12/47/86, HR-19/33/74 are labels for categorical reconciliation outcomes within a family (identity, outreach/ readiness, field-provenance, reference-basis, source-basis, ledger-disposition, maintenance-source, history-route). Their plain-language expansions are deliberately not provided — you must infer the mapping from the data, not memorize it. Method:

1. Count the outcomes. The enum size = number of distinct outcomes to distinguish (e.g. FP/RB/SB/MS/HR have 3; IC/OR have 4; LD has 5). Derive that many categorical outcomes from your reconciliation for that family.
2. Use built-in calibration. When the contract pairs a code family with named categories, that pairing reveals the family's semantics — e.g. a readiness_partition object that asks for an OR code per both/email_only/phone_only/not_ready bucket tells you OR codes label readiness states; solve the 1:1 assignment from that. Anchored control cases (pinned evidence rows in the scope) exist precisely to let you tie a known reconciliation outcome to its code.
3. Cross-check with the hub. Look for a signal in the records, reference tables, or /api/query results that separates the outcomes; assign codes so the observable ordering/partition is consistent.
4. Be internally consistent. The same outcome always maps to the same code, everywhere in the answer. Do not assume the numeric suffix implies severity or order unless the evidence shows it.

4. Output contract compliance (do this before returning)

The contract — not this skill — is authoritative. Verify each:

• Shape: every required key present; no extra keys (additionalProperties: false / additional_top_level_keys_allowed:false). Objects/arrays nested exactly as specified.
• Cardinality: arrays with minItems/maxItems (or a length) have exactly that many items — one row per requested focus item / rollup enum value / scope id, no more, no fewer.
• Ordering: apply each list's stated rule (usually lexicographic ascending by id; rankings by metric desc then tie-break). Set membership is deduplicated.
• Enums: every enum-typed value is one of the allowed literals (statuses, actions, source systems, categories, control codes).
• Rounding / precision: money/volume/weight/distance to 2 dp; rates to 4 dp; honor multipleOf. Counts are exact integers. Phone stays a string of digits.
• Ids: use only stable public ids present in the data or supplied in the scope; match id patterns.
• Partitions add up: mutually-exclusive count groups sum to their stated total (readiness buckets, depot dispositions).
• Consistency: e.g. dispatchable_person_count == length of dispatchable_master_ids; duplicate_raw_count == raw − logical; quarantine_rate == quarantined / canonical (rounded).
• Return one JSON object only — no Markdown, no commentary, no code fences.

5. Cross-cutting reminders

• Discover, don't assume: field names, pagination, snapshot-selection rule, and precedence order all come from the live catalog/schema/metadata.
• Reconcile as of the cutoff; ignore rows and rate versions outside it.
• Quarantined rows never enter normalized totals; valid mismatches always do.
• Recompute derived counts from your own reconciled set — never copy a count the API reports for a different scope.
• Keep every derivation deterministic so the same inputs always yield the same answer.

See references/task-family-map.md for how each observed task variant maps onto this pipeline, and scripts/asteria_client.py for the authenticated client.