Form filling from a dossier
Fill long PDF or web forms correctly from supporting documents.
20 project-authored proposals · taxonomy v1.1 · 0 new benchmark runs. Verifiers and partial-credit rules below are designs, not executed results.
- Typical applications
- Acrobat · Preview · web forms
- 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.
R04.05.01.02-T01 Fill a vendor tax form from a supplier profile in Acrobat
Original instruction
Fill the AcroForm input/vendor_tax_form.pdf (W-9 style, synthetic) from input/vendor_profile.json: legal name, business name, entity-type checkbox, exemption codes, address, TIN split across the 9 digit boxes, and today's date. Leave the signature field empty. Save output/vendor_tax_form_filled.pdf.
- Platforms
mac· macOSwin· Windows- Applications
- Acrobat
- Interfaces
desk· Desktop GUI- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
CL· CLI-sufficient- Interaction primitives
type· Text & formsclick· Point & click- Work product
WP02· Records/table- Conditions
- None declared
- Proposed difficulty
- 2 / 5
Proposed verifier
pypdf field values equal key (normalized whitespace); only the right entity checkbox is /On; TIN digits one per box; signature field empty; render check that values are visible (NeedAppearances or generated appearances).
Proposed partial credit
50% text fields (pro-rata); 20% checkbox; 20% TIN boxes; 10% visible appearances.
R04.05.01.02-T02 Complete a rental application PDF in Preview from an applicant packet
Original instruction
In Preview, fill input/rental_application.pdf from input/applicant_packet.pdf (ID card scan, pay stubs, prior landlord letter): personal info, 2 previous addresses with dates, employer, monthly income (average of the 3 pay stubs), pets Yes/No checkbox, emergency contact. Save.
- Platforms
mac· macOS- Applications
- Preview
- Interfaces
desk· Desktop GUI- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
CL· CLI-sufficient- Interaction primitives
type· Text & formsclick· Point & clickvis· Visual judgment- Work product
WP02· Records/table- Conditions
- None declared
- Proposed difficulty
- 2 / 5
Proposed verifier
Form field values equal key; income equals the computed average to the cent; checkbox states; saved by Preview (Producer Quartz PDFContext); unfilled optional fields empty.
Proposed partial credit
45% personal and address fields; 25% income computed; 15% checkboxes; 15% other fields.
R04.05.01.02-T03 Fill a multi-page grant application portal from a dossier folder
Original instruction
On the sandbox grants portal (http://grants.local, test login provided), complete application GR-2026-114 from input/dossier/ (PI CV PDF, budget.xlsx, narrative.docx): 6 pages of fields, 3 co-investigator rows, budget categories, uploads of the CV and narrative. Save as draft; do not submit.
- Platforms
web· Any OS (web)- Applications
- web forms · Excel · Word
- Interfaces
brow· Browser- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
type· Text & formsclick· Point & clickmulti· Cross-app- Work product
WP08· App/system state- Conditions
multiapp· Multi-app- Proposed difficulty
- 3 / 5
Proposed verifier
Portal DB: draft record fields equal key (exact for codes/numbers, normalized for text); 3 co-I rows; upload hashes equal input files; status 'draft'. Server rejects direct API writes (CSRF + UI event tokens); request log must show page navigation.
Proposed partial credit
45% fields (pro-rata); 20% co-I rows; 20% budget categories; 15% uploads and draft state.
R04.05.01.02-T04 Place answers on a flat, non-fillable PDF form at the right spots
Original instruction
input/club_membership_form.pdf is a flat scan-like form with no fields. Using Acrobat Fill & Sign, place the 11 answers from input/member.json on the lines beside their labels, tick the 2 correct option boxes with checkmarks, and save output/club_membership_filled.pdf.
- Platforms
mac· macOSwin· Windows- Applications
- Acrobat
- Interfaces
desk· Desktop GUI- GUI need · task setting
G2· GUI-essential- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
click· Point & clicktype· Text & formsgeom· Precise geometry- Work product
WP02· Records/table- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Each placed text annotation/flattened text: content equals key and bbox lies within the target answer region (IoU ≥0.5, baseline within 3 pt of the line); checkmarks inside the right boxes; no marks elsewhere. Acrobat's JavaScript console is disabled and the VM has no PDF libraries; Producer must be Acrobat.
Proposed partial credit
60% answers placed correctly (pro-rata); 25% checkmarks; 15% no stray marks.
R04.05.01.02-T05 File an insurance claim with conditional sections from an incident report
Original instruction
On the sandbox insurer portal, open a new property claim and fill it from input/incident_email.eml and input/photos/: date/time, location, description, answer 'Yes' to injury so the injury section appears and fill it, list 4 damaged items with values, upload the 6 photos. Save the draft claim.
- Platforms
web· Any OS (web)- Applications
- web forms · Outlook
- Interfaces
brow· Browser- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
type· Text & formsclick· Point & clickmulti· Cross-app- Work product
WP08· App/system state- Conditions
multiapp· Multi-app- Proposed difficulty
- 3 / 5
Proposed verifier
Portal DB draft: fields equal key; injury subsection filled; 4 item rows with exact values; 6 attachments with matching hashes; status 'draft'. UI-only submission enforced via CSRF/UI event tokens.
Proposed partial credit
35% core fields; 25% injury section; 25% item rows; 15% photo uploads.
R04.05.01.02-T06 Prepare a synthetic tax return form from wage and interest statements
Original instruction
Fill input/form_1040_synthetic.pdf for the taxpayer in input/docs/ (2 W-2 PDFs, 3 1099-INT PDFs, prior-year return): income lines, totals, standard deduction for the filing status, taxable income and tax from input/tax_table.pdf. Leave signature blank. Save output/return_filled.pdf.
- Platforms
mac· macOSwin· Windows- Applications
- Acrobat
- Interfaces
desk· Desktop GUI- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
CL· CLI-sufficient- Interaction primitives
type· Text & formsvis· Visual judgmentmulti· Cross-app- Work product
WP02· Records/table- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Field values equal key; arithmetic lines consistent (grader recomputes totals from filled inputs); tax amount equals table lookup; filing status checkbox correct; signature blank.
Proposed partial credit
35% source amounts transcribed; 30% totals/deduction; 25% tax lookup; 10% status and blanks.
R04.05.01.02-T07 Complete an ATS job application from a resume and submit to sandbox
Original instruction
On the sandbox ATS (OpenCATS-style, test candidate account), apply to job REQ-0412 using input/resume.pdf: 5 work-history entries with month/year pickers, 2 education entries, skills tags, salary expectation 95,000, upload resume and input/cover_letter.pdf, then submit and save the confirmation number to output/confirmation.txt.
- 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
type· Text & formsclick· Point & click- Work product
WP08· App/system state- Conditions
risk· Irreversible/risky- Proposed difficulty
- 3 / 5
Proposed verifier
ATS DB: application row for REQ-0412 submitted once; work/education rows equal key (dates month-exact); tags set; uploads hashed; confirmation.txt equals DB confirmation id. Duplicate submissions penalized.
Proposed partial credit
40% work history; 15% education; 30% other fields and uploads; 15% single submission + confirmation.
R04.05.01.02-T08 Fill a character-box customs form on an image-only scan
Original instruction
input/customs_declaration.jpg (converted to input/customs_declaration.pdf, image only) has one-character boxes for name, passport number and flight. Place one character per box from input/traveller.json using text annotations, tick the 'Nothing to declare' box, and save output/customs_filled.pdf.
- Platforms
mac· macOSwin· Windows- Applications
- Acrobat · Preview
- Interfaces
desk· Desktop GUI- GUI need · task setting
G2· GUI-essential- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
geom· Precise geometrytype· Text & formsclick· Point & click- Work product
WP02· Records/table- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Each character annotation centered inside its ground-truth box (center within the box, ≥1 pt from edges); characters in order equal key; the checkbox mark inside the right box; no marks outside target regions. No PDF libraries or PyObjC in the VM; Acrobat JavaScript and Automator PDF actions disabled; Producer must be Acrobat or macOS Quartz.
Proposed partial credit
70% characters in correct boxes (pro-rata); 20% checkbox; 10% no stray marks.
R04.05.01.02-T09 Enter a household into a benefits portal wizard and save the draft
Original instruction
On the sandbox benefits portal (test account), start a SNAP-style application and fill it from input/household_dossier.pdf: 4 household members with relationships and birthdates, 3 income sources with frequency, rent and utilities. Save and exit at the review page without submitting.
- 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
type· Text & formsclick· Point & clicknav· Pan/zoom/scroll- Work product
WP08· App/system state- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Portal DB draft: 4 member rows (names, relationships, DOB exact), 3 income rows (amount, frequency), expense fields; application status 'in_progress' (not submitted); UI-only writes enforced via session tokens.
Proposed partial credit
40% members; 30% incomes; 15% expenses; 15% saved, not submitted.
R04.05.01.02-T10 Fill a loan form from a dossier with conflicting addresses and flag them
Original instruction
Fill input/loan_application.pdf (4 pages) from input/borrower/ (ID scan, two utility bills, bank letter, employment letter). The documents disagree on address and employer start date: use the most recent document for each, fill the form, and add a sticky note on page 1 listing each conflict and the source you chose.
- Platforms
mac· macOSwin· Windows- Applications
- Acrobat
- Interfaces
desk· Desktop GUI- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
CL· CLI-sufficient- Interaction primitives
type· Text & formsvis· Visual judgmentmulti· Cross-app- Work product
WP02· Records/table- Conditions
- None declared
- Proposed difficulty
- 4 / 5
Proposed verifier
Field values equal key (rule-resolved values); note text mentions both seeded conflicts with chosen source file names (regex); no other annotations; other fields exact.
Proposed partial credit
45% fields; 25% conflicts resolved per rule; 30% conflict note content.
R04.05.01.02-T11 Fill a speaker application with strict length limits and taxonomy fields
Original instruction
On the sandbox conference CFP site, submit talk details from input/talk_dossier.md: title ≤80 chars, abstract ≤1,200 chars, bio ≤500 chars (condense the 1,400-char bio), choose track and 3 topic tags mapping the dossier's keywords to the site's fixed taxonomy, and save the draft.
- 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
type· Text & formsclick· Point & click- Work product
WP08· App/system state- Conditions
- None declared
- Proposed difficulty
- 4 / 5
Proposed verifier
Draft row: lengths within limits; title exact; abstract equals dossier abstract; bio keeps the 5 required facts (regex checks for employer, role, 2 projects, award) — no LLM judge; track and tags equal key mapping.
Proposed partial credit
15% title; 20% abstract; 30% bio facts within limit; 35% track and tags.
R04.05.01.02-T12 Submit a prior-authorization request in a payer portal from clinical notes
Original instruction
In the sandbox payer portal (test provider login), create a prior-authorization request for the synthetic patient in input/clinic_note.pdf: member ID, diagnosis ICD-10 codes and procedure CPT code chosen via the portal's search fields, dates of service, ordering provider NPI, and attach the note. Save without submitting.
- Platforms
web· Any OS (web)- Applications
- web forms · Acrobat
- Interfaces
brow· Browser- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
type· Text & formsclick· Point & clickmulti· Cross-app- Work product
WP08· App/system state- Conditions
- None declared
- Proposed difficulty
- 4 / 5
Proposed verifier
Portal DB draft: codes equal key (exact), dates, NPI and member ID exact; attachment hash; status 'draft'. Code fields accept only values picked from the search widget (stored with picker event id).
Proposed partial credit
20% member/provider; 40% ICD-10 and CPT codes; 20% dates; 20% attachment and draft state.
R04.05.01.02-T13 Fill and sign a permission slip PDF on iPad with Markup
Original instruction
On iPad, open 'Field Trip Permission.pdf' in Files, fill its 6 form fields from the note 'Trip details' in Apple Notes, draw a signature on the signature line with Markup, and save. The signature must sit on the line and not overlap printed text.
- Platforms
ios· iOS/iPadOS- Applications
- Files · Apple Notes
- Interfaces
mob· Mobile touch- GUI need · task setting
G2· GUI-essential- GUI need · ordinary access
G3· GUI-only- Interaction primitives
touch· Touch gesturestype· Text & formspath· Continuous path- Work product
WP02· Records/table- Conditions
multiapp· Multi-app- Proposed difficulty
- 3 / 5
Proposed verifier
PDF synced back: field values equal key; one ink annotation whose bbox bottom is within 4 pt of the signature line and does not intersect printed text bboxes; no other annotations. Shortcuts is disabled on the agent's iPad and it has no Mac or iCloud web access; fields and ink are added on-device.
Proposed partial credit
55% fields; 30% signature placement; 15% no stray marks.
R04.05.01.02-T14 Complete and submit a nine-page visa application in a sandbox portal
Original instruction
On the sandbox visa portal (test account), complete the tourist visa form from input/passport_scan.jpg, input/itinerary.pdf and input/employment_letter.pdf: 9 pages incl. a 12-row travel-history table, upload the photo meeting the portal's size rules (crop/resize input/photo.jpg as needed), submit and record the confirmation number.
- Platforms
web· Any OS (web)- Applications
- web forms · Preview
- Interfaces
brow· Browser- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
type· Text & formsclick· Point & clickvis· Visual judgment- Work product
WP08· App/system state- Conditions
long· Long-horizonrisk· Irreversible/riskymultiapp· Multi-app- Proposed difficulty
- 5 / 5
Proposed verifier
Portal DB: submitted application with fields equal key; 12 travel rows exact; photo passes portal validator (600×600 px, face-centered by detector); single submission; confirmation recorded in output/confirmation.txt.
Proposed partial credit
35% personal/passport fields; 25% travel table; 15% photo accepted; 25% other pages, single submit and confirmation.
R04.05.01.02-T15 Generate 25 filled permit forms from a spreadsheet of applicants
Original instruction
Create one filled copy of input/parking_permit.pdf for each of the 25 rows in input/applicants.xlsx, named output/permits/permit_<PlateNo>.pdf, mapping 9 columns to form fields (dates in MM/DD/YYYY, plate uppercase), using Acrobat data-merge features or field import rather than typing each.
- Platforms
mac· macOSwin· Windows- Applications
- Acrobat · Excel
- Interfaces
desk· Desktop GUI- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
CL· CLI-sufficient- Interaction primitives
click· Point & clickmulti· Cross-apptype· Text & forms- Work product
WP09· Files & folders- Conditions
long· Long-horizonmultiapp· Multi-app- Proposed difficulty
- 5 / 5
Proposed verifier
25 files with exact names; each file's field values equal its row (formatted); fields remain fillable (not flattened); no extra files.
Proposed partial credit
20% file count/names; 65% field values (pro-rata); 15% formats and fillable.
R04.05.01.02-T16 Enter a grant budget with dynamic lines that must reconcile
Original instruction
On the sandbox grants portal, fill the budget section of draft GR-2026-207 from input/budget.xlsx: 3 years × 7 categories, 11 personnel lines (add rows dynamically), indirect cost at 52% of modified total direct costs. Resolve every portal validation error until it shows 'Budget complete', then save.
- Platforms
web· Any OS (web)- Applications
- web forms · Excel
- Interfaces
brow· Browser- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
type· Text & formsclick· Point & clickmulti· Cross-app- Work product
WP08· App/system state- Conditions
long· Long-horizonmultiapp· Multi-app- Proposed difficulty
- 5 / 5
Proposed verifier
Portal DB budget lines equal workbook (to the dollar); 11 personnel rows; indirect = 52% MTDC per year; portal validation state 'complete'; totals reconcile with workbook grand total.
Proposed partial credit
40% category amounts; 25% personnel rows; 20% indirect correct; 15% validation complete.
R04.05.01.02-T17 Control pair of T01: fill the vendor tax form with pypdf
Original instruction
Control pair of T01. With pypdf from the terminal, fill input/vendor_tax_form.pdf from input/vendor_profile.json with the same fields, checkbox and TIN boxes, generating appearances so values are visible, and save output/vendor_tax_form_cli.pdf.
- Platforms
mac· macOSwin· Windowslin· Linux- Applications
- pypdf
- Interfaces
term· Terminal- GUI need · task setting
AP· API-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 with the GUI variant to measure modality gap.
Proposed partial credit
50% text fields; 20% checkbox; 20% TIN boxes; 10% appearances.
R04.05.01.02-T18 Fill a web survey that uses canvas pickers, sliders and drag-ranking
Original instruction
On the sandbox site survey http://survey.local/site-audit, answer from input/site_audit.json: pick the date in a canvas calendar, set 3 percentage sliders to exact values, drag-rank 6 priorities, and drop a map pin within 20 m of the given coordinates. Save the response.
- Platforms
web· Any OS (web)- Applications
- web forms
- Interfaces
brow· Browser- GUI need · task setting
G2· GUI-essential- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
drag· Drag & dropgeom· Precise geometryclick· Point & click- Work product
WP08· App/system state- Conditions
- None declared
- Proposed difficulty
- 4 / 5
Proposed verifier
Response row: date exact; slider values exact (widgets snap to 1%); ranking order exact; pin within 20 m (haversine). The survey sandbox issues no API token and accepts responses only from browser sessions carrying widget interaction tokens; widgets expose no ARIA values.
Proposed partial credit
15% date; 30% sliders (equal thirds); 35% ranking; 20% map pin.
R04.05.01.02-T19 Complete a dynamic XFA form that adds rows as you type
Original instruction
Fill the dynamic XFA form input/equipment_request_xfa.pdf in Acrobat from input/request.csv: requester section, 8 equipment rows (use the form's 'Add item' button), cost center from the dropdown, and justification text. The form recalculates totals; save output/equipment_request_filled.pdf.
- Platforms
mac· macOSwin· Windows- Applications
- Acrobat
- Interfaces
desk· Desktop GUI- GUI need · task setting
G2· GUI-essential- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
click· Point & clicktype· Text & forms- Work product
WP02· Records/table- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Extract XFA datasets XML: requester fields, 8 item rows with qty/price equal CSV, dropdown value, computed total equals key; still XFA (not flattened). Acrobat's JavaScript console is disabled and the VM has no PDF/XFA libraries; Producer must be Acrobat.
Proposed partial credit
20% requester; 50% eight item rows; 15% dropdown; 15% total recalculated.
R04.05.01.02-T20 Renew a vehicle registration in a sandbox portal up to payment
Original instruction
On the sandbox DMV portal (test account), start renewal for the vehicle whose VIN must be read from input/door_sticker.jpg, enter insurance details from input/insurance_card.pdf, upload the card, confirm address, and stop at the payment page without entering payment data. Save output/stop_page.txt with the fee shown.
- 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 judgmenttype· Text & formsclick· Point & click- Work product
WP08· App/system state- Conditions
risk· Irreversible/risky- Proposed difficulty
- 4 / 5
Proposed verifier
Portal DB: renewal record with VIN exact, insurer/policy/expiry exact, upload hash, address confirmed, state 'awaiting_payment' with no payment attempt logged; stop_page.txt equals the displayed fee.
Proposed partial credit
30% VIN read correctly; 30% insurance fields; 15% upload; 25% stopped at payment boundary with fee.
Mapped research references
Existing benchmark-to-class links from the research catalog. They are not record-level annotations, executions or product endorsements.