10-Step Ontology Implementation for Fast-Response Clothing Supply Chain

Executive Summary

This case study documents the first full cycle of a "full-chain ontology platform" implementation for a clothing enterprise. The ten-step process runs from strategic positioning to ontology decision upgrade, centered on one core contradiction: growth speed versus inventory capital constraints — more new launches increase unsold risk, and the breakthrough lies in prediction and fast-response capabilities.

1. Strategic Positioning

Output: a structured strategic map with seven categories — vision, mission, strategic position, core contradiction, strategic themes, "do-not" list, and key assumptions. The map is stored as nodes in strategy-context for direct reference by later steps.

Strategic statement: Build growth advantage through rapid new launches, omnichannel operations, and inventory efficiency.

Core contradiction: Growth speed vs. inventory capital — more launches, higher unsold risk. All subsequent goals, rules, and gates are designed around this.

Do-not list: No heavy-asset factory ownership, no offline direct stores, no low-margin inventory fire sales. Recorded in the ontology to prevent building workflows for these.

Three strategic themes: extreme fast-response (in-season replenishment, weekly rolling), omnichannel precision (livestream, platform, private domain synergy), private domain advancement (member segmentation, tag lifecycle, repeat-purchase operations). Four key assumptions logged for later validation: livestream channel share rising, new-item share >70% effective, traffic cost long-term rise, head supplier capacity lockable.

2. Goal Decomposition

Three annual targets: revenue growth (monthly sales ¥35.869M net, +12% YoY), gross margin (financial caliber pending), inventory turnover (GMROI monthly). Each target broken into verifiable sub-indicators: revenue → new-item sell-through rate and livestream channel share; gross margin → new-item full-price rate and discount red line; turnover → turnover days and unsold inventory share.

Translation work: turn management's verbal "wants" into indicators with unique definitions, single owners, traceable data sources. Every displayed number must hang on this indicator tree; if it cannot, it is not shown. Tree currently carries 15 supporting questions, 4 executable.

3. Organizational Alignment

Seven departments each completed a "four-piece" registration: responsibilities (what they do), authority boundaries (what they don't do), aligned goals (which indicator they own), collaboration relationships (handoffs). Four key departments excerpted:

Merchandise Planning (2 people): Annual/wave planning, category structure, price-band distribution. Boundary: does not replace single-item selection decisions, does not intervene in launch execution details. Owns revenue growth and gross margin.

Buyer (1 person): First orders, sample reviews, first-order quantity decisions. Boundary: model suggestion + human override with audit trail, does not replace price confirmation. Owns gross margin.

Fast Supply Chain: AI prediction & capacity collaboration, first-order fast-response & inventory constraints. Boundary: does not replace planning's first-order authority; suggestions must be auditable and rollback-able. Owns inventory turnover.

In-Season Fast-Response Ops (AI-assisted): Replenishment/allocation/clearance suggestions, in-season sell-through and gap monitoring. Boundary: AI only drafts suggestions, does not execute deductions; approval rests with store and team lead. Owns inventory turnover.

Two rules: unconfirmed departments marked "pending rotation" and block downstream SOP generation; AI can draft org suggestions and role drafts but four-piece confirmation must be human.

4. Role-Level SOP

Process hierarchy linked into ontology: value-chain → department → role → human-machine division. Main chain "Product Value Chain · Physical Flow" has eight nodes: season design, planning confirmation, selection review, procurement ordering, factory production, warehousing allocation, store sales, consumer. Each node registers five items: responsible role, inputs, outputs, judgment criteria, manual gates.

Four gates on main chain: cold-start validation (new plans small-scale test, no scale without pass), OTB budget gate (order-time budget check, excess auto-routes to approval), replenishment decision (gap suggestion → fast-response team → approval → actual deduction), clearance decision (unsold candidates → discount red-line check → gate release).

Standard: every node must become an executable workflow node in the ontology, directly referenceable by CQ, rules, and approvals. Granularity reaches "who under what condition must press which gate".

5. CQ Questioning & Semantic Alignment

Competency Questions (CQ) serve as the ontology's acceptance checklist: take the toughest business questions, ask the ontology; if it cannot answer, a gap exists. Example for "stockout" — five questions: AL-01/A: Which stores/SKUs are currently or soon stocked out, and what is the gap quantity? (Current state query — exposes missing inventory_stock semantic node) AL-02/D: Is stockout due to demand surge, insufficient first order, replenishment delay, or uneven store distribution? (Root-cause computable) AL-03/F: Without increasing total inventory, from which stores to which store, how much each? (Plan generatable) AL-04/K: Are reorder point, safety stock, pack precision, and best-seller service level calculated by unified rules? (Rule uniqueness) AL-05/C: Does allocation suggestion become an allocation order and complete outbound, in-transit, receipt, and effect review? (Closed-loop verifiable)

CQ aligns semantics: business terminology, system fields, ontology terms must refer to the same thing. Example "valid order" — business unclear if returns/exchanges/adjustments count; reverse lookup found sales caliber missing this coverage, logged as semantic gap. Resolution: "valid order" is not an independent business object but a judgment state of whether a sales order meets statistical caliber; implemented as SalesOrder.is_valid, single system-wide reference.

Ledger status: 142 business terms, 26 linked to CQ usage, 58 with implementation evidence, hotspot coverage 93% .

6. Rule Reasoning

One-way channel: fact inputs (inventory events, sales flows, CQ judgments) → rule library (L4 semantic layer, rules collected with status and owner) → conclusions (gaps, risks, candidate lists) → action suggestions (replenish, allocate, clear, escalate). Four representative rules deployed: OTB-01: First-order budget check. Order amount exceeding remaining OTB auto-routes to approval; approval triggers real-time deduction, balance visible instantly. ALLOC-02: Stockout auto-generates allocation suggestions. When store i stocked out and store j has allocatable inventory, suggestion list generated by gap and in-transit; constraint: no total inventory increase, per-store calculation. RISK-03: Unsold candidates auto-identified. New item 14-day sell-through below threshold enters clearance candidate pool; candidate must carry evidence card (sell-through source, allocation structure, price band) to pass gate. EVID-04: Conclusions must carry evidence chain. Any reasoning conclusion forced to include indicator-card source, object list, event chain; conclusions lacking evidence chain barred from workflow and approval.

Bottom line: machine can produce conclusions, but each must withstand human "why"; rules that cannot explain themselves are not loaded.

7. Causal Analysis

GMROI monthly deviation -8% . Analysis drills down via attribution tree: first layer splits into gross profit contribution (-5.2pt, 91%) and average inventory cost (-2.8pt, 9%); second layer further splits into sales (-18%), gross margin (-5.6pt, main factor, further drilled), inventory quantity (-3.1%), unit cost (-1.1%).

Three disciplines for attribution tree: (1) layer-by-layer decomposition without changing caliber, each child node uniquely defined by parent; (2) factor contributions must sum to total gap, reconcile accounts before attributing; (3) every conclusion carries evidence card, drillable to object — gross margin factor → responsible store/SKU → handling suggestion → escalation.

8. Workflow & Approval

CQ AL-03 and rule ALLOC-02 connect here: allocation suggestions enter single workflow channel, seven steps: stockout suggestion → fast-response team approval (gate) → allocation order generation → outbound → in-transit → receipt → effect review.

Three disciplines: full audit trail; gate unpressed → process suspended, pressed → auto-advance; every order traceable — who approved, based on which number, when effective. Replenishment, clearance, OTB excess approvals reuse this channel. AI drafts suggestions, but gates pressed by role holders. Approval is not ceremony; it gives every machine conclusion a clear accountable person before becoming real action.

9. Conclusion Feedback to Strategy

Execution results must return to strategic layer, otherwise the loop is only half done. Three feedback paths: (1) strategic assumptions validated or falsified monthly — e.g., "livestream share rising" checked against channel structure data, status moves from pending → validating → established; (2) strategic themes adjust resource weights per execution results; (3) strategic priorities refresh with gaps, linking to next month's targets.

Real example in monthly review: "gross margin financial caliber unconfirmed" flagged red and suspended. Unverifiable assumptions cannot stay idle on map; must generate to-do and assign owner. Feedback essence is forced reconciliation.

10. Ontology Decision Upgrade

Completing the loop yields not only operational conclusions but revisions to the ontology itself. Every CQ acceptance, attribution, and feedback patches the ontology. Revisions land in five layers:

Terms (L3 ontology layer): Unified semantics. "Stockout" upgraded from colloquial word to structured object, adding inventory semantic node.

Dictionary (L4 semantic layer): Unified indicator definitions. "Valid order" implemented as SalesOrder.is_valid, return/exchange caliber now single.

Rules (L4 semantic layer): Unified rule library. AL-03 question materialized as ALLOC-02 rule, bound to approval gate and review feedback.

Chain (L2 integration layer): Single data path. Source document → processing → indicator → page, no more manual copying or second caliber.

Accounts (L5 application layer): Automated reconciliation. After caliber unification, reconciliation shifts from human work to system action.

Foundation: SQLite main DB + DuckDB analytical layer — transactions and semantics in main DB, reasoning and attribution in analytical layer. Ontology changes also follow process: semantics traceable, versions rollback-able, caliber changes require approval.

Key Takeaways

After one full cycle, three core lessons remain: (1) Ontology is not a document but the daily operating mode — roles, rules, gates all live in the ontology, business and operations share the same semantics. (2) Machine produces conclusions, human presses gates — reasoning, attribution, suggestions delegated to system, but every action that reaches reality has a role owner and full audit trail. (3) The loop's endpoint is not a report but an upgrade — execution results rewrite strategy, revise semantics, and the next cycle starts from the new strategic position.