Office Intelligence.
EN 中文

A RESEARCH PROGRAM FOR KNOWLEDGE-WORK AGENTS

Office Intelligence

Environments, data, and learning for work that continues.

Studying how agents use evidence, produce useful artifacts,
and carry knowledge work forward across tasks.

Environments Evaluation Training Harness

A CONTINUING WORK ENVIRONMENTORGANIZATION & RESEARCH
A cutaway office with research and analysis desks, a review meeting, and a shared-material cabinet. Four annotations explain roles, artifacts, review, and retained context.

EXTRACTED FROM THE WORK

Computer snapshot

Personal filesAuthorized mountsTools & app state

A runnable starting point for a fresh attempt.

01

Roles & collaboration

02

Tasks & artifacts

03

Review & revision

04

Persistent materials

Organization → task-time workspace

01 / RESEARCH USES

From work to research.

Executable task-time workspaces with authorized files, tools, and organizational state.

Roles · materials · permissions

Task-specific criteria for checking submitted artifacts and observable outcomes.

Deliverables · criteria · evidence

Execution demonstrations and interactive environments for supervised and on-policy learning.

Demonstrations · interaction

Tool interfaces and state for document work; principal–assistant workflows for delegation and review.

Tools · memory · delegation
Hover to follow · click to pin
RequestEvidenceArtifactReviewNext task retained context
Work in context. Roles, working materials, and review connect one assignment to the next.

02 / ENVIRONMENTS

OfficeTown

Paper (arXiv)

Sample task

Financial reconciliation

Task

ASSIGNMENT BRIEF

Prepare editable reconciliation workpapers for three exception streams: customer deposits, card clearing, and vendor credits.

  1. Reconcile the supplied records and explain the differences in each stream.
  2. Trace each proposed disposition to supporting records. Keep unresolved differences visible.
  3. Identify the missing evidence needed before an adjustment or period-close conclusion can be supported.

Constraints & deliverables

Preserve the supplied originals. Return an editable workbook with the calculations, record references, and evidence notes needed for review. Do not treat missing support as a resolved exception.

Environment

An NPC’s workspace: personal files, shared resources, tools, and communication.

NPC computerReconciliation Operations ManagerTASK-TIME FILE SNAPSHOT
Source-based summaryAuthored demo
Workspace35 declared initial files

Files

Open a filename to inspect its summary and related app.

Other NPC private homes are not mounted.

/home/npc/computer9 files · personal workspace
  • Documents/
    • Areas/
      • Reconciliation/
        • Role Context.mdMD
          Source-based summary

          Responsibilities, evidence handling, and decision authority for the reconciliation role.

        • Operating Guide.mdMD
          Source-based summary

          Account-level tie-outs, exception classification, aging, and evidence mapping.

    • Projects/
      • README.mdMD
        Source-based summary

        Index of working projects, including the earlier reconciliation-triage folder.

      • Reconciliation triage/
        • README.mdMD
          Source-based summary

          The prior triage memo was prepared on 3 January 2026 from the input brief, JSON packet and workbook. It is advisory and preserves the originals.

        • Triage Memo.docxDOCX
          Source-based summary

          Three streams are ready for evidence-limited workpapers: customer deposits, card clearing and vendor credits. Warehouse, payroll, tax and loan areas lack route-ready evidence.

          Open related app
    • Resources/
      • Templates/
        • Review Checklist.mdMD
          Source-based summary

          State the objective, cite material claims, expose calculations and gaps, identify reviewers, and perform no prohibited external action.

  • Notes/
    • Start Here.mdMD
      Source-based summary

      Projects hold active work; Areas hold responsibilities; Resources hold reusable material. Personal notes persist separately from the read-only company mount.

    • Todo.mdMD
      Source-based summary

      Initial commitments note. Subsequent work requests are available in Notices.

      Open related app
    • Working Style.mdMD
      Source-based summary

      Precise, measured communication; assumptions stated before findings; evidence-led comparison of source records and variance patterns.

R/W mount · preserve supplied originals

/mnt/company16 files · permission-filtered references
  • INDEX.mdMD
    Source-based summary

    Authorized company-file index: project inputs, reusable references, and quality templates.

  • Projects/
    • Client project/
      • Private/
        • Inputs/
          • Input brief.mdMD
            Source-based summary

            Synthetic input population: 5 invoices, 4 deposits, 3 bills, 2 credits, 6 bank/card rows and 3 clearing differences. No posting, payments or external contact.

          • Operating packet.jsonJSON
            Source-based summary

            Structured tables corresponding to the workbook. Inspect three clearing-difference rows in Spreadsheets.

            Open related app
          • Source workbook.xlsxXLSX
            Source-based summary

            Eleven native worksheets, including customer_deposits, bank_activity and clearing_differences. The original workbook is unchanged; selected input rows are shown in Spreadsheets.

            Open related app
  • Resources/
    • External/
      • Regulatory/
        • Reference 01.pdfPDF
          Source-based summary

          Reference material in the authorized company view.

        • Reference 02.pdfPDF
          Source-based summary

          Reference material in the authorized company view.

        • Reference 03.pdfPDF
          Source-based summary

          Reference material in the authorized company view.

    • Learning/
      • Reconciliation/
        • Worked example.xlsxXLSX
          Source-based summary

          Reference material in the authorized company view.

    • References/
      • Internal control reference.htmlHTML
        Source-based summary

        Reference material in the authorized company view.

      • Chart of accounts.pyPY
        Source-based summary

        Reference material in the authorized company view.

      • Recordkeeping reference.pdfPDF
        Source-based summary

        Reference material in the authorized company view.

      • Reference page A.htmlHTML
        Source-based summary

        Reference material in the authorized company view.

      • Reference page B.htmlHTML
        Source-based summary

        Reference material in the authorized company view.

      • Cash flow calculator.xlsxXLSX
        Source-based summary

        Reference material in the authorized company view.

    • Templates/
      • Quality/
        • Traceability guide.mdMD
          Source-based summary

          Every amount, count and finding must point to a workbook tab, data table or state record. Preserve missing evidence and the pending decision rather than replacing it with an estimate.

        • Workpaper review checklist.mdMD
          Source-based summary

          Check evidence sufficiency, formulas/control totals, source traceability, unsupported claims and authority boundaries. Preserve the useful native format.

Read-only · within this role’s authorization

/mnt/services6 files · frozen historical projections
  • calendar.jsonJSON
    Source-based summary

    Service-state carrier and related application view.

    Open related app
  • chat.jsonJSON
    Source-based summary

    Historical work records with timestamps and source references.

    Open related app
  • git.jsonJSON
    Source-based summary

    Service-state carrier and related application view.

    Open related app
  • mail.jsonJSON
    Source-based summary

    Service-state carrier and related application view.

    Open related app
  • meetings.jsonJSON
    Source-based summary

    Service-state carrier and related application view.

    Open related app
  • notices.jsonJSON
    Source-based summary

    Historical work records with timestamps and source references.

    Open related app

Service-state files · read-only

/mnt/task4 files · request and metadata
  • CONTRACT.en.mdMDEmpty carrier
    Source-based summary

    Empty contract carrier.

  • README.mdMD
    Source-based summary

    Workspace orientation, file locations, and access permissions.

  • TASK.en.mdMD
    Source-based summary

    Prepare an editable internal-use reconciliation workbook for the three scoped streams. Preserve source files, expose evidence limitations, and keep proposed actions advisory.

  • source_inventory.jsonJSON
    Source-based summary

    Manifest of source files and their bindings.

Read-only · evaluator rubric excluded

Terminal & execution

FILES & TOOLS
npc@computer ~/computer
$ pwd
/home/npc/computer
$ ls
Documents/  Notes/
$ ls /mnt
company/  services/  task/

# Tool surface
read   bash   edit   write
Inspect filesExecute commandsEdit working copiesReturn artifacts

Documents

Source-based summary2026-01-03 · Triage Memo.docx
First-wave reconciliation triage
Customer deposits · $850.00

An unassigned receipt needs remittance and deposit/bank support. Invoice candidates provide context, not authority to apply cash.

Card clearing · $42.50

The records flag a duplicate/import issue. Processor batch reports, merchant fee detail and import logs are still needed to establish the cause.

Vendor credits · $225.00

The credit is unassigned. A vendor statement or credit memo is needed; no external outreach or balance change is authorized.

Source locator · Triage Memo.docx · Executive view / Detailed Open-Items Register

Inspect the input rows

Spreadsheets

Source-based summarySource workbook.xlsx
clearing_differences · three supplied rows (USD)
Input row Book amount Bank amount Listed difference State
1 · Card clearing 2,717.50 2,675.00 42.50 Open
2 · Customer deposits 0.00 850.00 850.00 Open
3 · Vendor credits 225.00 0.00 225.00 Evidence requested internally

Source locator · Operating packet.json · tables.clearing_differences[0–2]; Source workbook.xlsx · clearing_differences, rows 2–4

Summarychart_of_accountscustomersvendorssales_invoicescustomer_depositsvendor_billsvendor_creditsbank_activityclearing_differencesclose_issue_log
Read the memo’s evidence limits

Chat

Source-based summary14 captured records · selected meeting-channel summaries

Earlier intake discussion · before the triage memo.

Engagement lead

Open a bounded first-wave intake review. Identify inputs, workstreams and unresolved dependencies; preserve the private-workspace boundary.

Source locator · chat.json · records[0]

Data lead

The input register is still incomplete. Do not claim completeness or treat reference material as reconciliation evidence.

Source locator · chat.json · records[1]

Quality reviewer

Only conditional internal activation is supported. Log missing controls and scope issues rather than resolving them by assumption.

Source locator · chat.json · records[3]

3 January 2026 · simulation time

Open the meeting summary

Notices

Source-based summary13 captured records · selected work history
  1. Earlier triage memo reviewed

    The earlier triage deliverable passed review.

  2. Triage request incorporated

    The prior triage request is resolved and its memo becomes context for the next assignment.

  3. Current workpapers requested

    Prepare first-wave workpapers from the canonical triage memo. The recorded deadline is 4 January 2026 at 04:04:55 simulation time.

Source locator · notices.json · records[9], [11], [12]

View the dated work timeline

Mail

Authored demo
Drafts · two internal follow-upsNOT SENT
Draft 1 · Request missing evidence

Reconciliation manager → internal data lead

Please locate the remittance/deposit support, processor batch and fee/import logs, and vendor credit support referenced in the triage memo. Record each item’s source, period and authorization status.

Keep the three exceptions open until the requested evidence has been reviewed.

Draft 2 · Confirm review boundaries

Reconciliation manager → internal quality reviewer

The planned workbook will separate deposits, card clearing and vendor credits, with source links and a limitations register. AR/AP aging is context only. Please review the proposed scope before any further action.

No posting, payment, write-off or close conclusion is proposed as an approved action.

Basis: the memo’s missing-evidence list

Calendar

Source-based summaryWork timeline from notices and chat
03January 2026 · recorded context
  1. Intake-routing discussion

    Discussion window recorded in the meeting channel and end notice.

  2. Workpaper assignment arrives

    The current collaboration request follows the incorporated triage memo.

04January 2026 · recorded deadline
  1. Requested delivery deadline

    Delivery deadline stated in the work request.

Source locator · chat.json · records[0]; notices.json · records[8], [12]

Meetings

Source-based summaryMeeting-channel reconstruction
First-wave intake and routing review
2026-01-03 · 02:01–02:4114 captured messages; end notice reports 18 turns

Captured speakers: engagement lead, data lead, close manager, and quality reviewer.

Agenda

Identify available inputs, classify first-wave workstreams, and define the quality boundary before routing.

Recorded constraint

At that meeting, the input register remained incomplete. The quality reviewer allowed only bounded activation, not reconciliation or close-readiness claims.

Next action discussed

Complete the source register with versions, period/entity and source-vs-reference classification before seeking further clearance.

Source locator · chat.json · records[0], [3], [10], [13]; notices.json · records[8]

Inspect selected discussion summaries

Git

Authored demo
demo/reconciliation-workpapersProposed project layout
README.mdScope, input references and restrictions
evidence_requests.csvOpen evidence requests
.gitignoreKeep private inputs and generated artifacts out of Git
Preview the proposed README

Purpose: prepare internal workpapers using read-only source inputs. Track what is known, what remains unsupported, and which reviewer owns the next decision.

Do not commit private inputs or completed workbooks. Keep source hashes in a local provenance register; share artifacts only through an authorized process.

Personal home + authorized mounts + tool access + service stateFiles · communication · work history

Evaluation

Check the delivered work, not how convincing the final message sounds.

Evaluation dimensions
Check What must be inspectable
Numerical consistency The three streams reconcile to the supported totals.
Record-level evidence Proposed dispositions cite the records that support them.
Unresolved differences Missing evidence and unsupported conclusions remain explicit.
Usable artifact The workbook is editable and the work can be reviewed.
Source preservation Supplied originals remain unchanged.

03 / DOCUMENT HARNESS

DocAtlas

NeurIPS 2026 · Accepted Project page

Read the source. Keep the finding. Let it inform the next search.

Inside the harness

Recorded interfaces

Web harness

WEB HARNESSOpen larger
01 Inspect source02 Navigate structure03 Update notes

Terminal harness

TERMINAL HARNESSOpen larger

04 / LONG-DOCUMENT EVALUATION

XL-DocBench

NeurIPS 2026 · Accepted Project page

Follow a regulatory reference across documents, then locate the supporting provisions.

Sample question

Legal & regulation · Reference chain

Question

BENCHMARK QUESTION · adubench_cross_000095

In the Basel Framework referenced by the 2017 final reforms integration note, which chapter contains the capital-instrument eligibility criteria?

Expected answer: a chapter identifier and title.

01Follow the reference

Find the reform document’s integration notice.

02Resolve the destination

Navigate the consolidated framework.

03Check the provisions

Locate instrument eligibility, not just a matching phrase.

Documents

Eight related regulatory PDFs form the supplied series context. Two carry the released evidence annotations for this question.

  1. 01 77 pp.
  2. 02 180 pp.
  3. 03
    Basel III: Finalising post-crisis reforms December 2017 Annotated evidence
    162 pp.
  4. 04 96 pp.
  5. 05 82 pp.
  6. 06
    Pillar 3: updated framework Updated disclosure requirements
    62 pp.
  7. 07
    Basel III monitoring instructions January 2023 · Version 3.9
    142 pp.
  8. 08
    The Basel Framework Consolidated framework Annotated evidence
    1,982 pp.

Evidence & answer

  1. 01 / REFORM DOCUMENTAnnotated p. 1 ↗
    “This standard has been integrated into the consolidated Basel Framework”

    The integration notice identifies the document to follow.

  2. 02 / CONSOLIDATED FRAMEWORKAnnotated p. 102
    “Throughout CAP10 the term ‘bank’ is used to mean bank, banking group or other entity”

    Locate the chapter and check that it covers regulatory capital categories.

  3. 03 / ELIGIBILITY PROVISIONSAnnotated pp. 104, 110
    “For an instrument to be included in Common Equity Tier 1 capital it must meet all of the criteria that follow.”

    Section 10.8 specifies CET1 eligibility. The annotation at p. 110 also points to AT1 instrument criteria in §10.11.

REFERENCE ANSWERCAP10 Definition of eligible capitalString answer · case-insensitive exact-match rule

Source: XL-DocBench dataset ↗ Question ID: adubench_cross_000095

05 / CURRENT COVERAGE

Eight professional domains.

The current OfficeTown task collection spans the settings below, with domain-specific roles, materials, and requirements.

Assurance

Evidence & data lineage

Clinical

Operations & documentation

Design

Production & accessibility

Entertainment

Events & production

Finance

Analysis & reconciliation

Industrial

Engineering & manufacturing

Legal

Review & compliance

Software

APIs, analytics & delivery

Research projects

Evaluation, document interaction, and environments for knowledge work.

BENCHMARK

XL-DocBench

Evidence-grounded reasoning over long documents.

Public project AGENT SYSTEM

DocAtlas

Stateful interaction with document evidence.

Public project ORGANIZATIONS & ENVIRONMENTS

OfficeTown

Continuing professional work as a source of tasks, environments, and learning resources.

Paper on arXiv