Computer-Use Agent Taxonomy中文
R04.07.01.01 · TASK CLASS

Data entry from documents & scans

Transcribe fields from invoices, forms and scans into a system of record without errors.

20 project-authored proposals · taxonomy v1.1 · 0 new benchmark runs. Verifiers and partial-credit rules below are designs, not executed results.

Typical applications
Excel · web forms · OCR tools
Class GUI-need label
HY · Hybrid

Task proposals · 20

Expand to inspect the complete original wording and proposed checks. Original proposal text remains English in both interfaces.

R04.07.01.01-T01 Key header fields from twenty scanned invoices into a register

Original instruction

input/invoices_scan/ holds 20 scanned vendor invoices (PDF images, synthetic). Fill input/invoice_register.xlsx sheet 'Register', one row per invoice: Invoice No, Invoice Date (real date), Vendor (as in input/vendors.csv), Currency, Net, Tax, Total. Totals must equal Net + Tax.

Platforms
mac · macOS win · Windows
Applications
Excel · Preview
Interfaces
desk · Desktop GUI
GUI need · task setting
HY · Hybrid
GUI need · ordinary access
CL · CLI-sufficient
Interaction primitives
vis · Visual judgment type · Text & forms multi · Cross-app
Work product
WP02 · Records/table
Conditions
None declared
Proposed difficulty
2 / 5

Proposed verifier

Row-level match by Invoice No vs key: dates exact, vendor exact canonical name, amounts exact to the cent; Net+Tax=Total check; 20 rows, no extras.

Proposed partial credit

15% invoice numbers; 30% dates and vendors; 45% amounts (three columns, equal shares); 10% row count.

R04.07.01.01-T02 Enter fifteen scanned timesheets into an HR web form

Original instruction

Using the sandbox HR portal (test supervisor login), enter the 15 scanned weekly timesheets in input/timesheets/ (JPG): employee ID, week ending date, daily hours Mon–Sun, overtime flag if total > 40. Save each as submitted-for-approval.

Platforms
web · Any OS (web)
Applications
web forms
Interfaces
brow · Browser
GUI need · task setting
HY · Hybrid
GUI need · ordinary access
G2 · GUI-essential
Interaction primitives
vis · Visual judgment type · Text & forms click · Point & click
Work product
WP08 · App/system state
Conditions
None declared
Proposed difficulty
2 / 5

Proposed verifier

Portal DB: 15 timesheet records; per-day hours exact; week-ending dates exact; overtime flag consistent; status 'pending_approval'; entries created via UI sessions (audit log user agent + CSRF tokens).

Proposed partial credit

15% records present; 60% daily hours (pro-rata); 15% dates; 10% overtime flags.

R04.07.01.01-T03 Enter multi-line vendor bills into Odoo so line totals reconcile

Original instruction

In the sandbox Odoo (test accountant), create draft vendor bills for the 8 invoices in input/bills/ (3–12 lines each): vendor, bill reference, date, each line's product, quantity, unit price and tax, so each bill's total equals the invoice total. Do not post them.

Platforms
web · Any OS (web)
Applications
Odoo · web forms
Interfaces
brow · Browser
GUI need · task setting
HY · Hybrid
GUI need · ordinary access
G2 · GUI-essential
Interaction primitives
type · Text & forms click · Point & click vis · Visual judgment
Work product
WP08 · App/system state
Conditions
None declared
Proposed difficulty
3 / 5

Proposed verifier

Odoo DB (account.move/lines): 8 draft bills; header fields exact; line count and per-line qty/price/tax equal key; amount_total equals invoice total to the cent; state 'draft'.

Proposed partial credit

15% headers; 55% lines (pro-rata); 20% totals reconcile; 10% left in draft.

R04.07.01.01-T04 Code handwritten survey forms into a spreadsheet using a codebook

Original instruction

input/survey_scans/ has 40 handwritten paper surveys (synthetic). Code them into input/survey_data.xlsx per input/codebook.pdf: checkbox items as codes, written ages as integers, 'Other' text verbatim, and code 99 where a mark is ambiguous per the codebook's rule (two boxes or erasure).

Platforms
mac · macOS win · Windows
Applications
Excel · Preview
Interfaces
desk · Desktop GUI
GUI need · task setting
G2 · GUI-essential
GUI need · ordinary access
G2 · GUI-essential
Interaction primitives
vis · Visual judgment type · Text & forms
Work product
WP02 · Records/table
Conditions
None declared
Proposed difficulty
3 / 5

Proposed verifier

Cell-level accuracy vs key ≥98% for full credit on checkbox items; ages exact; 'Other' text normalized match ≥0.9; ambiguous cells coded 99 exactly where the key says so. Scans are handwritten, so OCR alone fails the ambiguous cases. VBA, COM and AppleScript automation disabled, no Python in the VM; file must be saved by Excel (app.xml).

Proposed partial credit

45% checkbox items; 15% ages; 15% other text; 25% ambiguity coding.

R04.07.01.01-T05 Create contacts from business-card photos without duplicates

Original instruction

Add the 18 people in input/cards/ (business-card photos, synthetic) to macOS Contacts in group 'Expo 2026': name, company, title, emails, phones (E.164), website. Six are already in Contacts; update those cards instead of creating duplicates.

Platforms
mac · macOS
Applications
Contacts · Preview
Interfaces
desk · Desktop GUI
GUI need · task setting
HY · Hybrid
GUI need · ordinary access
CL · CLI-sufficient
Interaction primitives
vis · Visual judgment type · Text & forms click · Point & click
Work product
WP08 · App/system state
Conditions
None declared
Proposed difficulty
3 / 5

Proposed verifier

Contacts DB (AddressBook sqlite): group with 18 members; fields equal key (phones E.164); 6 pre-existing record ids updated, no duplicates by email; 12 new records.

Proposed partial credit

50% field accuracy; 20% group membership; 20% no duplicates; 10% phone format.

R04.07.01.01-T06 File multi-currency receipts into an expense report per policy

Original instruction

In the sandbox expense app (Expensify-style, test employee), create report 'Berlin trip' with the 14 receipts in input/receipts/ (photos): merchant, date, amount in original currency, category per input/policy.pdf, and attendees for meals. Attach each image. Do not submit.

Platforms
web · Any OS (web)
Applications
web forms
Interfaces
brow · Browser
GUI need · task setting
HY · Hybrid
GUI need · ordinary access
G2 · GUI-essential
Interaction primitives
vis · Visual judgment type · Text & forms click · Point & click
Work product
WP08 · App/system state
Conditions
None declared
Proposed difficulty
3 / 5

Proposed verifier

App DB: 14 expenses in the report; amount/currency/date exact; category equals policy mapping; meal attendees set; each receipt image attached (hash); report status 'open'.

Proposed partial credit

45% amounts/currency/dates; 25% categories; 15% attendees; 15% attachments and open state.

R04.07.01.01-T07 Transcribe scanned bank statements into a transaction sheet that balances

Original instruction

Transcribe the 6 scanned monthly statements in input/statements/ into input/transactions.xlsx: Date, Description, Debit, Credit, Balance per line. Each page's opening balance plus transactions must equal its printed closing balance; flag in column G any line where your entries do not reconcile.

Platforms
mac · macOS win · Windows
Applications
Excel · OCR tools
Interfaces
desk · Desktop GUI
GUI need · task setting
HY · Hybrid
GUI need · ordinary access
CL · CLI-sufficient
Interaction primitives
vis · Visual judgment type · Text & forms multi · Cross-app
Work product
WP02 · Records/table
Conditions
None declared
Proposed difficulty
3 / 5

Proposed verifier

Row-level comparison with key (date, amounts exact; description normalized ≥0.9); per-statement running balances equal printed closings; column G flags only where the key has a true discrepancy (1 seeded printing error).

Proposed partial credit

55% transactions exact (pro-rata); 25% statements reconcile; 20% correct flagging.

R04.07.01.01-T08 Enter paper claims into a 3270 green-screen claims system

Original instruction

Using the 3270 terminal emulator connected to the sandbox mainframe claims app (test operator ID), enter the 10 paper claim forms in input/claims/ via transaction CLM1: policy number, claimant, loss date, cause code, amount. Clear any field-validation error messages and record each assigned claim number in output/claims.csv.

Platforms
win · Windows mac · macOS
Applications
3270 emulator
Interfaces
desk · Desktop GUI term · Terminal
GUI need · task setting
G3 · GUI-only
GUI need · ordinary access
G3 · GUI-only
Interaction primitives
type · Text & forms key · Shortcuts vis · Visual judgment
Work product
WP08 · App/system state
Conditions
None declared
Proposed difficulty
3 / 5

Proposed verifier

Sandbox host DB: 10 claim records with fields equal key; claims.csv numbers equal host-assigned ids; no duplicate claims. The host exposes only the 3270 screen (no API); the emulator's scripting/macro interface is disabled and the agent has no TN3270 library.

Proposed partial credit

60% claim fields (pro-rata); 20% claim numbers recorded; 20% no duplicates/errors.

R04.07.01.01-T09 Accession lab requisition forms into a LIMS

Original instruction

In the sandbox SENAITE LIMS (test lab clerk), register samples from the 12 scanned requisition forms in input/requisitions/: client, patient ID, sample type, collection date/time, and the ordered analyses (checkboxes on the form). Receive each sample after creation.

Platforms
web · Any OS (web)
Applications
web forms · SENAITE
Interfaces
brow · Browser
GUI need · task setting
HY · Hybrid
GUI need · ordinary access
G2 · GUI-essential
Interaction primitives
vis · Visual judgment type · Text & forms click · Point & click
Work product
WP08 · App/system state
Conditions
None declared
Proposed difficulty
3 / 5

Proposed verifier

LIMS DB: 12 samples with fields equal key; analyses set equal ticked tests; state 'sample_received'; no extra samples.

Proposed partial credit

35% sample fields; 40% analyses; 15% received state; 10% no extras.

R04.07.01.01-T10 Enter purchase orders into a remote-desktop-only legacy app

Original instruction

The legacy purchasing app is only reachable through a remote desktop session (pixels only, no clipboard). From the 12 PO PDFs in input/pos/, create each PO in the app: supplier code, delivery date, and line items (part no, qty, price), then save and note each PO's system number in output/po_numbers.csv.

Platforms
win · Windows
Applications
Remote Desktop · legacy app
Interfaces
desk · Desktop GUI
GUI need · task setting
G3 · GUI-only
GUI need · ordinary access
G3 · GUI-only
Interaction primitives
vis · Visual judgment type · Text & forms click · Point & click
Work product
WP08 · App/system state
Conditions
multiapp · Multi-app
Proposed difficulty
4 / 5

Proposed verifier

App DB inside the remote VM: 12 POs with headers and lines equal key; po_numbers.csv equals DB ids; clipboard redirection disabled, so values must be typed; no API or DB access from the agent's machine.

Proposed partial credit

20% headers; 50% line items (pro-rata); 20% PO numbers recorded; 10% none duplicated.

R04.07.01.01-T11 Normalize multilingual invoice fields into one register

Original instruction

input/eu_invoices/ has 16 invoices in German, French, Japanese and English with local date and number formats ('1.234,56 €', '2026年3月5日'). Enter them into input/eu_register.xlsx with ISO dates, numeric amounts, ISO currency codes and VAT IDs without spaces.

Platforms
mac · macOS win · Windows
Applications
Excel · Preview
Interfaces
desk · Desktop GUI
GUI need · task setting
HY · Hybrid
GUI need · ordinary access
CL · CLI-sufficient
Interaction primitives
vis · Visual judgment type · Text & forms
Work product
WP02 · Records/table
Conditions
None declared
Proposed difficulty
3 / 5

Proposed verifier

Per-row exact match vs key: date cells are real dates equal key; amounts numeric exact; currency ISO codes; VAT IDs normalized; 16 rows.

Proposed partial credit

30% dates; 40% amounts; 15% currency; 15% VAT IDs.

R04.07.01.01-T12 Key synthetic CMS-1500 claims into a billing system

Original instruction

In the sandbox billing system (OpenEMR-style, test biller), create claims from the 9 scanned CMS-1500 forms in input/cms1500/: patient, insured ID, diagnosis codes (box 21) with pointers, and each service line (date, CPT, modifiers, units, charge). Leave claims unbilled.

Platforms
web · Any OS (web)
Applications
web forms · OpenEMR
Interfaces
brow · Browser
GUI need · task setting
HY · Hybrid
GUI need · ordinary access
G2 · GUI-essential
Interaction primitives
vis · Visual judgment type · Text & forms click · Point & click
Work product
WP08 · App/system state
Conditions
None declared
Proposed difficulty
4 / 5

Proposed verifier

Billing DB: 9 claims; ICD-10 codes and order exact; service lines exact (CPT, modifiers, units, charge, diagnosis pointers); totals equal form box 28; status unbilled.

Proposed partial credit

15% patient/insured; 25% diagnoses; 45% service lines; 15% totals and status.

R04.07.01.01-T13 Receive scanned shipping manifests into a WMS and flag discrepancies

Original instruction

In the sandbox WMS (Odoo Inventory), receive the 5 inbound shipments using the scanned manifests in input/manifests/ (some pages rotated): for each receipt, enter received quantities per SKU from the handwritten 'Received' column, set lot numbers, and add a note listing SKUs that differ from the printed 'Shipped' quantity.

Platforms
web · Any OS (web)
Applications
Odoo · web forms
Interfaces
brow · Browser
GUI need · task setting
HY · Hybrid
GUI need · ordinary access
G2 · GUI-essential
Interaction primitives
vis · Visual judgment type · Text & forms click · Point & click
Work product
WP08 · App/system state
Conditions
None declared
Proposed difficulty
4 / 5

Proposed verifier

stock.move.line quantities and lots equal key; receipt notes list exactly the discrepant SKUs; receipts validated (state done) and backorders created only where quantities short.

Proposed partial credit

45% received quantities; 15% lots; 20% discrepancy notes; 20% validation/backorders.

R04.07.01.01-T14 Double-key 120 handwritten census forms to 99.5% accuracy

Original instruction

Enter the 120 handwritten household forms in input/census_forms/ into the sandbox data-entry app, which enforces double entry: key each form, then run the app's verify pass and resolve every mismatch it reports. Target ≥99.5% field accuracy across 14 fields per form.

Platforms
web · Any OS (web)
Applications
web forms
Interfaces
brow · Browser
GUI need · task setting
HY · Hybrid
GUI need · ordinary access
G2 · GUI-essential
Interaction primitives
vis · Visual judgment type · Text & forms
Work product
WP02 · Records/table
Conditions
long · Long-horizon
Proposed difficulty
5 / 5

Proposed verifier

App DB: 120 verified records; field-level accuracy vs key (≥99.5% full credit, linear down to 95%); verification log shows both passes and mismatch resolution for every record.

Proposed partial credit

70% field accuracy (scaled); 20% all records verified; 10% mismatches resolved.

R04.07.01.01-T15 Post ten vendor invoices through an SAP GUI transaction from PDFs

Original instruction

In the SAP GUI sandbox client (test user), enter the 10 vendor invoices in input/sap_invoices/ with the logistics invoice verification transaction against their purchase orders: invoice date, reference, amount, tax code, and PO line matching, resolving any quantity variance messages per input/ap_rules.md. Park, do not post.

Platforms
win · Windows
Applications
SAP GUI
Interfaces
desk · Desktop GUI
GUI need · task setting
G3 · GUI-only
GUI need · ordinary access
G3 · GUI-only
Interaction primitives
type · Text & forms key · Shortcuts vis · Visual judgment
Work product
WP08 · App/system state
Conditions
licensed · Licensed pro app long · Long-horizon
Proposed difficulty
5 / 5

Proposed verifier

Sandbox tables (RBKP/RSEG) read by the grader: 10 parked documents with header fields and PO item assignments equal key; variances handled per rules; no posted documents. SAP GUI scripting is disabled on the client.

Proposed partial credit

25% headers; 40% PO line matching; 20% variance handling; 15% parked not posted.

R04.07.01.01-T16 Transcribe a cursive 19th-century ledger into a structured sheet

Original instruction

Transcribe the 12 page images of input/ledger_1872/ (cursive, synthetic reproduction) into input/ledger.xlsx: Date, Account, Particulars, £, s, d per entry, with column H 'uncertain' = TRUE where a reading is doubtful. Page totals must reconcile to the written totals in £/s/d.

Platforms
mac · macOS win · Windows
Applications
Excel · Preview
Interfaces
desk · Desktop GUI
GUI need · task setting
HY · Hybrid
GUI need · ordinary access
G2 · GUI-essential
Interaction primitives
vis · Visual judgment type · Text & forms nav · Pan/zoom/scroll
Work product
WP02 · Records/table
Conditions
None declared
Proposed difficulty
5 / 5

Proposed verifier

Entry-level match vs expert key: amounts exact (pounds/shillings/pence), dates exact, particulars CER ≤10%; per-page totals reconcile; uncertainty flags overlap key's doubtful set (F1 ≥0.6).

Proposed partial credit

45% amounts; 15% dates/accounts; 15% particulars; 25% page reconciliation and uncertainty flags.

R04.07.01.01-T17 Control pair of T01: extract invoice headers with an OCR pipeline

Original instruction

Control pair of T01. From the terminal, run OCR (tesseract or ocrmypdf) on input/invoices_scan/ and write a parser that fills output/invoice_register_cli.xlsx with the same 7 columns for the 20 invoices.

Platforms
mac · macOS win · Windows lin · Linux
Applications
tesseract · Python
Interfaces
term · Terminal
GUI need · task setting
CL · CLI-sufficient
GUI need · ordinary access
CL · CLI-sufficient
Interaction primitives
code · Code & commands
Work product
WP02 · Records/table
Conditions
None declared
Proposed difficulty
2 / 5

Proposed verifier

Same checker as T01; compared against the GUI/manual variant.

Proposed partial credit

15% numbers; 30% dates and vendors; 45% amounts; 10% row count.

R04.07.01.01-T18 Log ten photographed receipts in an Android expense app

Original instruction

On the Android emulator, open the expense app (Pro Expense-style) and add the 10 receipts whose photos are in the gallery album 'Receipts': amount, date, merchant, category per input/categories.txt, and attach the photo to each entry.

Platforms
and · Android
Applications
expense app · Gallery
Interfaces
mob · Mobile touch
GUI need · task setting
HY · Hybrid
GUI need · ordinary access
G3 · GUI-only
Interaction primitives
touch · Touch gestures vis · Visual judgment type · Text & forms
Work product
WP08 · App/system state
Conditions
None declared
Proposed difficulty
3 / 5

Proposed verifier

App SQLite (adb pull by grader): 10 entries; amounts/dates/merchants exact; categories per mapping; attachment URIs reference the right photos.

Proposed partial credit

50% amounts/dates; 20% merchants; 15% categories; 15% attachments.

R04.07.01.01-T19 Read analog meter dials from photos and submit readings

Original instruction

On the sandbox utility portal (test account), submit readings for the 8 meters in input/meter_photos/ (analog 5-dial gas and electric meters). Read each dial by the standard rule (when between digits take the lower; check the dial to the right), and enter readings with the photo upload.

Platforms
web · Any OS (web)
Applications
web forms
Interfaces
brow · Browser
GUI need · task setting
HY · Hybrid
GUI need · ordinary access
G2 · GUI-essential
Interaction primitives
vis · Visual judgment type · Text & forms click · Point & click
Work product
WP08 · App/system state
Conditions
None declared
Proposed difficulty
4 / 5

Proposed verifier

Portal DB: 8 readings equal key exactly (5 digits); meter IDs matched; photo hashes attached; portal's plausibility check passed (no override).

Proposed partial credit

70% readings exact (pro-rata over 8); 15% meter mapping; 15% uploads.

R04.07.01.01-T20 Enter scanned wage and interest statements into tax software

Original instruction

In OpenTaxSolver (GUI) on the test machine, enter the synthetic taxpayer's 2 W-2s and 4 1099-INT/DIV scans from input/tax_docs/ into the federal form, run the computation, and save output/return.txt. Amounts must match the scans; do not e-file.

Platforms
win · Windows mac · macOS
Applications
OpenTaxSolver · Preview
Interfaces
desk · Desktop GUI
GUI need · task setting
HY · Hybrid
GUI need · ordinary access
CL · CLI-sufficient
Interaction primitives
vis · Visual judgment type · Text & forms click · Point & click
Work product
WP02 · Records/table
Conditions
None declared
Proposed difficulty
4 / 5

Proposed verifier

Saved return file: input lines equal key amounts; computed total income, AGI and tax equal reference computation; no e-file artifacts.

Proposed partial credit

60% source amounts (pro-rata); 30% computed lines; 10% saved correctly.

Mapped research references

Existing benchmark-to-class links from the research catalog. They are not record-level annotations, executions or product endorsements.

In this workflow family