Peopleops Console Reconciliation

PeopleOps Console Reconciliation

You are an HR ops controller working a "PeopleOps Console" task. The portal holds deliberately conflicting records for the same fact (several leave records, draft vs submitted payroll, several audit events per case). Your job is to pick the authoritative record, list the excluded losers, apply the business rules, and emit a JSON answer that matches the task's answer_template.json exactly.

The scoring is on the answer JSON, so two things matter equally: getting the right record/decision, and reporting it with the template's normalized enum labels (not free-text). Lists are sets of IDs.

Standard operating procedure

1. Read the task assets. Read the prompt, then input/payloads/answer_template.json (or the path the prompt gives). The template is the contract: every key, the type of each field, and for enum fields the allowed_values you are permitted to emit. Plan to fill exactly those keys.
2. Reach the API. Read the task's environment_access.md for the base URL, then GET /health to confirm it is up. The web-UI login is not needed — call the JSON API directly with curl -s.
3. Identify the task shape from the prompt and follow the matching playbook in references/playbooks.md:
   • A — Onboarding closeout verification (leave + payroll for an employee)
   • B — Case folder + formal-notice review
   • C — Leave source-precedence validation for an employee
   • D — Payroll assignment + accrual readiness for an employee
   • E — Recruitment outcome reconciliation for an opening A task may combine shapes; run each relevant playbook and merge the fields.
4. Pull only the named entity's records. Look entities up by their specific IDs (?q=<ID>, the /<id> detail route, or /api/audit?case_id=<CASE-ID>). See the endpoint map and field shapes in references/api.md.
5. Apply the business-judgment rules in references/rules.md to select the authoritative record, compute exclusions, test folder readiness, detect notice defects, choose the controlling/supporting audit event, and apply the approval-vs-closeout gate.
6. Assemble the JSON, map every decision to the template's enum label, double- check every required key is present and typed correctly, then return only the JSON (no markdown, no commentary).

The core mental model: precedence and exclusion

Almost every field is one of: "which record is authoritative?", "which records are excluded?", or "what does that imply?". The precedence ladder (full detail in references/rules.md):

• Approved / Submitted record of the correct type for the period → authoritative.
• Employee profile summary (/api/employees row) → convenience summary, can be stale, loses to an approved/submitted assignment.
• Case summary text → weakest, never authoritative alone.
• Draft / Superseded / voided / obsolete → always excluded, even if a profile summary agrees with them. Excluded-ID list fields collect every loser as a set.

For leave, the controlling row is the record_type: "Leave assignment" that is Approved/Submitted — not an HRMS leave ledger or People Ops adjustment worksheet row. For payroll, it is the Submitted Salary assignment; the Draft is excluded.

Reading the authoritative answer from records (not guessing)

• Audit detail text is authoritative for result labels and for which record won. It often literally says "QA result: . controls/matches ...". Read the label and map it to the template enum; don't infer a different one.
• Notice quality lives in the notice/message packet (quality + defects[]), not in the case summary. Copy the defect codes verbatim.
• Folder readiness is a set comparison of required_files/required_tags against files/tags; the folder-checklist attachment and folder.tag_missing audit just restate the gaps.
• Escalation owner / SLA must be read verbatim from the named audit package's detail (or its linked policy). Never invent a number or reuse one from another task — each package has its own.

Output conventions

• The JSON must match answer_template.json exactly: same keys, correct types (string/integer/number/boolean/list[...]/enum).
• For enum and list[enum] fields, emit only a value from that field's allowed_values. These normalized labels (defect codes, gate labels, status, scope, source, result, action labels) are the output contract — prefer them over any free-text reasoning.
• List fields are sets: include every member, no duplicates; order is not meaningful but include each ID once. Array fields like candidate lists contain IDs only.
• Numbers are raw numbers (e.g. a 2026-04 period implies an effective_date of 2026-04-01; a cost total is the arithmetic sum of every cost-ledger line).
• Return only the JSON object — no markdown fences, no explanation — when the prompt says so.

Common pitfalls

• Using a draft or superseded record because it is newest. Recency does not beat approval/submission; drafts are excluded and go in the excluded list.
• Treating a worksheet/adjustment leave row as the entitlement. Only the Leave assignment record type is the controlling leave source.
• Trusting the profile summary. It can carry a stale policy even when its balance looks right; the approved assignment still controls and profile_policy_ignored is true.
• Deciding notice quality or candidate outcome from the case summary. Use the notice packet and the committee decision + offer respectively.
• Including off-scope audit events in a scoped decision. A folder.*/notice.* event is excluded from a leave/payroll-precedence decision (and vice versa); put it in excluded_audit_event_ids.
• Letting an approval imply closeout. Approval is sufficient only when records are clean; a defective folder or notice forces a hold.
• Free-text enum values. If your label is not in allowed_values, it is wrong — map it to the closest allowed normalized label.
• Enumerating collections to "explore". Query only the named entity's records; unfiltered list browsing is out of scope.

References

• references/api.md — endpoint map, scope discipline, and field shapes.
• references/rules.md — the full source-precedence, folder, notice, audit, gate, recruitment, and escalation rules with the normalized vocabularies.
• references/playbooks.md — step-by-step query sequences per task shape.