| name | atlas-commerce-ops-review |
| description | Solve Atlas Commerce Operations cutoff-based analytical review tasks (fulfillment, refund, carrier, warehouse, support). Read request + template contracts literally, replay event history at a cutoff, dedup imported rows, and produce a JSON object that conforms exactly to the answer template. |
Atlas Commerce Operations Review
These tasks ask for a cutoff-based operational review: take a cohort of production business records, freeze their state at an as-of cutoff, apply stated business definitions/SLAs/status policies, and emit one JSON object conforming exactly to an answer template. The plumbing (query service, schema, data dictionary, auth) is provided in the environment; this skill is the analytical methodology.
0. Read the contract before writing a line of SQL
Every task ships two JSON files in input/payloads/:
- a request (
*_request.json): scope, windows, cutoff, business definitions, thresholds, status rules, ordering/tie-breaks.
- an answer template (
answer_template.json): the exact output schema.
Treat both as load-bearing contracts:
- The template is the output law.
required, additionalProperties:false, minItems/maxItems, pattern, multipleOf, and each field's description / x-precision define what to compute and how to format it. Emit only the schema's fields, in its shape, no commentary, no extra keys. Mismatched precision or an extra key fails the whole object.
- The request definitions are literal. Every clause carries weight:
inclusive/exclusive, at or before vs strictly before, more than 24 hours, below (strict <) vs at least (>=), which count is the denominator. Decide each threshold operator from the wording; > vs >= flips borderline counts and everything downstream (rates, status, IDs).
If a definition and the template field description disagree, reconcile both before computing; the template field description usually restates the precise formula.
1. Two timestamps you must never conflate: window vs cutoff
- Eligibility window (e.g.
created_at/opened_at in [start, end]): selects the cohort. Boundary inclusivity comes from the request.
- As-of cutoff: where you freeze state to measure elapsed time and terminal status. Events after the cutoff do not exist for this review.
A record is eligible by its window; its state and metrics are measured at the cutoff. Do not pull state from after the cutoff.
2. Production/test/internal scoping
- Always exclude test rows (
is_test = 0).
- "Production accounts" scope: also exclude internal accounts (
is_internal = 0) unless the request explicitly includes them. This is the most common population error — decide it deliberately from the request's population field.
- Segment/tier/region/warehouse filters come from the request's
account_scope/cohort. Note whether region is the account's region or the warehouse's region — regional rollups use the assigned facility's region, not the customer's.
3. Imported rows are append-only and duplicated — dedup first
Raw event/scan/attempt tables receive import retries. The same logical event arrives multiple times under the same source_system + external_event_id with different ingested_at. Before any analysis, dedup by (source_system, external_event_id) keeping the earliest ingested_at. This applies to lifecycle event tables (support case events, warehouse task events) and to carrier scans.
Logical-business-row dedup is different: e.g. a refund_id is one logical refund that may have multiple settled rows — keep one row per logical id (min row id) to avoid double counting. Know which identity is the logical id vs the physical row id for each table; the data dictionary states this.
4. Derive state from event replay, never the status snapshot
Header tables carry a current_status convenience column the dictionary warns "may lag append-only event history." Never trust it. Replay the deduped, cutoff-filtered events in (event_at, row_id) order to derive the true state at the cutoff. Examples:
- support case open/resolved/reopened — terminal state is whatever the last relevant event leaves it as.
- warehouse task completed/rework — completeness comes from
COMPLETED/REWORK events ≤ cutoff, not current_status.
- carrier shipment status — the latest canonical scan ≤ cutoff by
(canonical_event_at, scan_row_id).
5. Recurring / terminal events: first vs last
A status event can recur (e.g. RESOLVED followed by another RESOLVED). When measuring "time to resolution" you almost always want the final occurrence of the terminal event, not the first — the work was actually completed at the last one. Decide per definition; for state-at-cutoff questions, replay gives the final state automatically. Getting first-vs-last wrong shifts medians and breach tallies silently.
6. Breach populations and elapsed-time clocks
A "breach" metric is measured on the whole eligible population, not only cases that reached a terminal state:
- A resolved/closed case uses elapsed time from its start to its terminal event.
- A still-open case uses elapsed time from its start to the cutoff (the definition says "an active case uses active elapsed time at the cutoff").
Measure each case to its own endpoint. Do not restrict the breach count to resolved cases only.
Support active-time clock
"Support active time" excludes periods the case sat waiting on the customer. Implement a clock that:
- starts at the case
opened_at;
- pauses on
WAITING_CUSTOMER;
- resumes on
CUSTOMER_REPLIED, REOPENED, or OPENED;
- ignores other events for the waiting state.
Walk deduped events ≤ endpoint in (event_at, case_event_id) order accumulating only non-waiting seconds. First-response time = active time to the first AGENT_RESPONDED; if none, active time to the cutoff. Resolution time = active time to the (final) RESOLVED, or active time to the cutoff if still open. Thresholds are per priority and compared strictly (exceeds = >).
Other elapsed clocks
Warehouse/productivity tasks use productive minutes attached to completion events (units / productive_minutes × 60), not wall-clock time. Use the unit each metric specifies.
7. Money, FX, and reversals
- Monetary amounts are minor units (smallest unit of the row's currency). Convert the minor amount by
(amount_minor / 100) * usd_per_unit.
fx_rates.usd_per_unit is USD per one unit of the named currency, keyed by (rate_date, currency). Use the rate for the transaction's service date and the row's currency — not the order's date — unless the request names a different basis.- Effective = settled minus reversals. A reversal links to its parent via
linked_refund_id; count a reversal only if its parent is an in-scope effective settled logical refund. Net = Σ settled USD − Σ effective reversal USD.
- Comparison FX basis: when comparing a refund value to an order's gross, value the order gross in its own currency at the refund's service-date rate (per the request's
order_gross_comparison clause).
8. Corrections (carrier/inventory) — minimal canonical, never raw
When a task approves a canonical correction:
- Correct one canonical field on one business row with a guarded
UPDATE. Never touch raw source values, source-identity fields (source_system, external_event_id, row ids), or unrelated rows.
- Append one
correction_audit row carrying every audit column (audit_id, correction_key, entity_type, entity_id, source_row_id, field_name, old_value, new_value, reason_code, corrected_at, actor) using the approved values from the request.
- Run the update + audit insert in one transaction with the stated
expected_total_changes.
- Report
APPLIED only if exactly one business row and one audit row commit and a post-change read query confirms the new canonical value. Any other outcome → NOT_APPLIED with the results actually observed. Do not claim success without the verifying read.
- Backlog/quality counts are taken after the correction (post state), and only over the in-scope cohort (e.g. shipments with an effective scan in the named batch at/before the cutoff).
9. Rankings, rollups, and ordering
- Honor the request's order list exactly, including tie-breaks. "Rate descending then id ascending," "first two by rate ascending then region ascending," "top three by units/hour then employee_id ascending" — each key and its direction is significant; ties resolved by the next key.
- Regional/segment rollups: pick the entity the request names (assigned warehouse region for fulfillment; account tier/segment for refunds).
- Re-derive each ranking from the same deduped, cutoff-frozen population as the counts — a different scoping for the rollup vs the totals is a common silent error.
10. Status classification
Apply the status rules in the order written (e.g. HEALTHY → WATCH → CRITICAL; STABLE → PRESSURED → AT_RISK; LOW → MODERATE → HIGH; CONTROLLED → ELEVATED → SEVERE). The last rule is the catch-all ("otherwise" / "all other outcomes"). Rates use the specified denominator (often eligible count — incomplete/active cases stay in the denominator). Evaluate conditions at full precision.
11. Rounding and precision
- Round only final reported values to the precision the template/request states (e.g. rates to 4 decimals, money to 2, medians to 2). Keep full float precision through every intermediate step; rounding mid-pipeline drifts rates and status boundaries.
- Median: over the resolved-at-cutoff population; for an even count, average the two central values, then round.
x-precision: 0 means integer.
12. Self-verification before submitting
- Schema-check: validate the emitted JSON against the template keys, types, precision (
multipleOf), array lengths/uniqueness, and that no extra fields exist. A single schema slip fails the object.
- Population consistency: does the sum of mutually-exclusive states (open + resolved, complete + incomplete) equal the eligible count? Do rate numerators/denominators use the populations the request named?
- Variant cross-check: when a definition is genuinely ambiguous (first vs last terminal event; breach over resolved-only vs all cases; include internal or not), compute every defensible variant and pick the one matching the literal wording. If feedback indicates an error, the wrong field is almost always downstream of one interpretation choice — isolate which metric moved rather than re-rolling the whole answer.
- Sanity bounds: breach rates, completion rates, medians should sit in plausible ranges; a median at exactly a threshold or a 0%/100% rate usually means an off-by-one or a population bug.
Quick task-type map
- Fulfillment scorecard: order cohort by campaign + creation window; complete = ≥1 physical shipment AND all physical shipments effectively DELIVERED by cutoff; on-time = every shipment delivered ≤ its promised_delivery_at; severe-exception by the 24h-over-promise rule; worst regions by unrounded regional rate asc then region asc.
- Refund reconciliation: dedup logical refunds, settle−reversals net in USD via service-date FX, top reasons by net USD desc then reason asc, leakage candidates (refund > gross in USD, OR ≥2 same-recode settled refunds), cohort risk by candidate rate + net USD thresholds.
- Carrier quality: one raw/canonical contradiction → minimal canonical correction + audit + post-verify; backlog = cohort whose final carrier status ≠ DELIVERED.
- Warehouse productivity: completion/rework from events ≤ cutoff; units/hour per employee from completed-event units & productive minutes; top employees and lowest team by stated orders; delayed high-priority = HIGH/URGENT, due before cutoff, not completed by cutoff.
- Support health: eligible case population; active-case state at cutoff; first-response and resolution active-time breaches over all cases (open cases measured to cutoff); severe-active = open/reopened at cutoff AND URGENT/HIGH AND beyond resolution threshold; worst 3 accounts by severe desc, breach desc, account asc; resolved-at-cutoff active-time median; risk by severe + first-response rates.
