One operational vocabulary joins the line to the ledger

A contractual KPI dictionary is only as reliable as the atomic records feeding its base elements.[S08][S38] When factories digitise logs or connect sensors, they often model data in one of two fragile ways: as bare time-series tuples (timestamp, tag_name, numeric_value), or as mutable status rows where a cell such as order_status is overwritten whenever state changes.[S06][S08] Both strip away shop-floor context.

A bare sensor tuple such as ("2026-10-30T09:14:22Z", "SEW_HEAD_14_SPEED_RPM", 3200) records that a shaft rotated at 3,200 revolutions per minute at an instant, an illustrative reading.[S06] Alone, that tuple cannot tell an analyst whether the machine was sewing a customer order (order_id), running test scrap, re-stitching a rejected collar (defect_id), or consuming a specific cotton lot (lot_id).[S05][S08] Conversely, a mutable row for bundle LOT-2026-0842 whose status cell is overwritten from CUT to SEWN, DEFECT_HOLD, REPAIRED and DISPATCHED erases the intermediate defect and repair, hiding first-pass failures and introducing target leakage into ML training sets.[S04][S08]

To preserve context without mutating history, the data spine models shop-floor occurrences as immutable visibility events under the GS1 Electronic Product Code Information Services (EPCIS) Release 2.0 standard across four dimensions:[S08]

Under this discipline, defect_observed, repair_completed, lot_split and dispatch_confirmed are distinct, append-only event records linked by shared identifiers and provenance under W3C PROV-O.[S08][S09] If a bundle of 50 panels is split into 45 conformant panels and 5 stained panels, a transformation event records the parent lot_id and two child lot IDs (45 pieces and 5 pieces, illustrative quantities).[S08] Current work-in-progress state is derived as a view over the event stream, while the immutable log remains intact for KPIs and model training.[S08][S09]

To make this structure concrete, Exhibit 11 specifies the minimum ten-field envelope required for every visibility event before it enters the curated store.[S07][S08][S09]

Every field in Exhibit 11 blocks a documented integration failure.[S07][S08][S09] Carrying both event_time and ingestion_time ensures that when an edge gateway buffers counts during a two-hour network outage (illustrative duration), the delayed upload is credited to the shift in which the garments were sewn (event_time) while the lag remains visible (ingestion_time - event_time).[S08] Requiring source_method and quality_state prevents manual entries, optical counters and imputed estimates from being silently merged.[S09] Wire serialisation and cryptographic signing belong to the transport and security layers examined next and in Chapter 07.

A frequent procurement error in industrial digitalisation is assuming that adopting a communication protocol automatically creates a shared vocabulary.[S06][S07][S10] Vendors often claim that connecting machines to an MQTT broker or an OPC UA server unifies the factory data model. Moving packets across a wire and agreeing on the business meaning of the payload inside those packets are separate engineering tasks.[S06][S07] Confusing transport with semantics produces an automated pipeline of incompatible numbers.[S10]

The contrast between OPC Unified Architecture (OPC UA, IEC 62541 / OPC 10000-1) and MQTT Version 5.0 defines the boundary between transport and meaning:[S06][S07]

When an engineer publishes {"val": 42, "ts": 1793336400} to cameroon/plant1/weaving/loom07/picks, MQTT 5.0 governs delivery according to Quality of Service (QoS 0, 1 or 2).[S07] It does not verify whether 42 is picks per minute, hundreds of picks since reset, or a fault code; whether ts is UTC seconds or local milliseconds; or which lot_id and order_id are running on loom07.[S07][S08] A topic path is a routing address, not a governed data contract.[S05][S07]

Separating transport from payload semantics lets the same governed event contract (Exhibit 11) travel over three transport mechanisms as the plant matures, without altering a downstream KPI formula or ML feature definition:[S06][S07][S0302]

Because consumers read the governed event contract rather than binding to CSV columns, MQTT topics or PLC registers, upgrading a line from manual sheets to sensors is a non-breaking change in source_method.[S05][S10][S38]

Even when events arrive in a ten-field envelope, categorical payloads (event_type, defect_id, downtime cause, waste category and unit) fragment unless governed by canonical taxonomies.[S01][S04] Different consumers need different granularity from the same observation:

Logging only coarse categories (Major Defect) at entry destroys diagnostic signal: once 40 defect types are collapsed into Major, no query or ML pipeline can recover oil stains versus skipped stitches.[S04][S09] Letting each team maintain private mapping tables is equally destructive.[S01][S39]

GFT enforces a two-part rule: capture at the atomic leaf code, and project external views through versioned crosswalk tables:[S08][S09][S0301]

As shown in Exhibit 13, one governed defect observation event feeds three views without three collection routines.[S08][S38]

In Exhibit 13, Node 1 records leaf code DEF-FAB-019 (such as an illustrative weft slub above 4 millimetres).[S08] Node 2 validates the ten-field envelope (Exhibit 11), binding asset_id, lot_id, order_id and event_time.[S08] Node 3 applies the versioned crosswalk (tax-defect-v1.3) without altering the stored event.[S09][S0301] Node 4a feeds the shift's ISO 22400-2 First-Pass Yield calculation (wp02);[S38] Node 4b supplies a morphological label with lot context for vision evaluation (wp03);[S04] and Node 4c projects buyer AQL severity and rejected mass for traceability and carbon accounting (wp05 and wp16).[S08] Verifying consistent human coding across shifts requires the quality gates in Chapter 04 and wp03.[S04]