Import and classify cost lines
Import SAP cost lines, classify each with a reason, and send it for controller sign-off.
Source file
SAP export today, connector in the pilot.
Import this month’s SAP export.
The sample holds PO items, account assignments, goods receipts and invoice status. It is a bilingual synthetic file with the same details a controller needs.
Inspect source fields and SAP mapping
Synthetic joined SAP-style export, not a native SAP extract or live connector. External POSID WBS codes are resolved from the internal account-assignment key. Friendly JSON names are normalised for the demo.
| Friendly field | Technical field | Object |
|---|---|---|
| Purchase order | EBELN | EKKO/EKPO |
| PO item | EBELP | EKPO |
| Item short text (40 characters) | TXZ01 | EKPO |
| PO item long text | PO_LONG_TEXT | PO text extract |
| Vendor (fictional name) | VENDOR_NAME | Vendor joined display |
| Material group (illustrative) | MATKL | EKPO |
| Net value VND | NETWR | EKPO |
| Currency | WAERS | EKKO |
| Plant (fictional) | WERKS | EKPO/PRPS |
| PO document date | BEDAT | EKKO |
| External WBS element code | POSID | PRPS joined to EKKN |
| Account assignment category / WBS | KNTTP | EKPO/EKKN display |
| Goods receipt material document | MBLNR | EKBE/MSEG |
| GR posting date | BUDAT | Material document |
| Responsible cost centre (fictional) | KOSTL | PRPS/ANLZ |
| Joined invoice status | INVOICE_STATUS | Derived extract field |
| Proposed demo asset receiver | PROPOSED_ASSET | Demo preparation field |
| Synthetic line reference | DEMO_LINE_ID | Demo preparation field |
Interactive import accepts the normalised JSON. CSV is a reference export; a pilot adds the approved extract mapping.
Three fictional factory projects
WBS hierarchies · 12 proposed assets · VND
Lotus packaging line
VN.26.PKG · VN-LT
72 cost lines
VND
River wastewater upgrade
VN.26.WWT · VN-RV
72 cost lines
VND
Meadow warehouse extension
VN.26.WH · VN-MD
72 cost lines
VND
Measured, with the misses included.
Exact match on category, capex/expense/split/review treatment, group book and local book. All imported seed lines are in the denominator; missing output is incorrect. Correct review routes count as policy decisions; this is not an approval-automation rate.
216 cost lines · 72 authored scenario families · at most 3 variants per family · 72 lines with long text · 5 predefined site conventions withheld at first pass. Labels were written before model recording and stay fixed.
Fixed seed order: first 72/216 lines (one third, rounded down) supply scripted category/treatment corrections. Corrections propagate by description/vendor/material-group similarity to specific pending lines only. Vague raw source text never teaches another line; vague targets and bundled/split targets never inherit a proposal. A specific purchased item may inherit a supported controller correction even if the model previously abstained; vendor, material group and normalised description or part-code family determine similarity. Explicit manual corrections remain protected. Score only the remaining 144 later lines; their hidden labels are never used for teaching. Compare this held-out score with the full-set first pass cautiously because denominators differ. The heldOutFirstPass field uses the same held-out denominator. Rule fallback outputs count as unavailable model outputs, never correct model predictions. No label-derived tuning.
Policy: sample-2026-11-22-v2 · effective 2026-11-22
Recorded model: gpt-6.1-sol
Measured: 2026-11-22T09:32:13.028Z
Reproduce: pnpm measure
| Source language | Correct / total | Model | Rule only |
|---|---|---|---|
| English | 64 / 77 | 83.1% | 35.1% |
| Vietnamese | 42 / 48 | 87.5% | 41.7% |
| Mixed | 85 / 91 | 93.4% | 40.7% |
| Difficulty tier | Correct / total | Model | Rule only |
|---|---|---|---|
| Clear | 140/141 | 99.3% | 44.7% |
| Ambiguous | 51/60 | 85.0% | 35.0% |
| Site convention | 0/15 | 0.0% | 0.0% |
Person routing: 43/216 (19.9%). Abstentions: 29. A correct review route counts as a policy decision, not an approved asset.
Inspect category accuracy, confusion and first-pass misses (25)
| Ground-truth category | Correct | Model | Rule only |
|---|---|---|---|
| Machinery | 51/57 | 89.5% | 36.8% |
| Buildings | 29/33 | 87.9% | 30.3% |
| Utilities | 29/30 | 96.7% | 40.0% |
| Software | 15/18 | 83.3% | 44.4% |
| Expense | 36/39 | 92.3% | 38.5% |
| Shared costs | 20/24 | 83.3% | 45.8% |
| Unresolved | 11/15 | 73.3% | 46.7% |
Model: Utilities / capex / capex / capex
Model: Utilities / capex / capex / capex
Model: Machinery / capex / capex / capex
Model: Unresolved / review / review / review
Model: Unresolved / review / review / review
Model: Unresolved / review / review / review
Model: Unresolved / review / review / review
Model: Unresolved / review / review / review
Model: Utilities / capex / capex / capex
Model: Utilities / capex / capex / capex
Model: Unresolved / review / review / review
Model: Unresolved / review / review / review
Model: Unresolved / review / review / review
Model: Unresolved / review / review / review
Model: Unresolved / review / review / review
Model: Unresolved / review / review / review
Model: Unresolved / review / review / review
Model: Utilities / capex / capex / capex
Model: Unresolved / review / review / review
Model: Utilities / capex / capex / capex
Model: Unresolved / review / review / review
Model: Unresolved / review / review / review
Model: Unresolved / review / review / review
Model: Unresolved / review / review / review
Model: Unresolved / review / review / review
Category confusion: true category → proposed category (count)