Skip to content
Business BenchA benchmark for AI agents on business work

Tasks·desk·Bookkeeping

August bank reconciliation for the checking account

bank-reconciliation·time budget 1200 s·source

The request

The task request supplied to the adapter, alongside the workspace below. Runtime instructions, authorized tools and skills also affect the system. Containerized runs withhold host-side references; local mode is not an isolation boundary.

Can you do the August bank reconciliation for our checking account? The bank export, the ledger and Alma's July rec are in the folder, and Dean's email says what he wants. Save it as reconciliation.xlsx.

Workspace

  • bank_rec_2026-07.xlsx
  • email_from_dean.txt
  • gl_1010_checking_2026-08.xlsx
  • harbor_trust_checking_2026-08.csv

Expected deliverables

  • reconciliation.xlsx

This task also has a task-specific grader (check.py).

Recorded acceptance checks

The table uses check names and outcomes from the frozen campaign ledger. Acceptance requires all required predicates to pass; a predicate may itself contain a tolerance or coverage threshold. Its scope is part of the measurement contract.

Recorded checkTypeRequiredFailed / observed
reconciliation.xlsx existsfile_existsyes0 / 6
live formulasxlsx_has_formulasyes0 / 6
no error cellsxlsx_no_errorsyes0 / 6
outstanding checks and deposits in transit totalscustomyes0 / 6
adjusted balancexlsx_value_presentyes1 / 6
both sides tie and items listedcustomyes1 / 6
Current task declaration (6 checks)

From task.yaml. Original runner grading and frozen retrospective scoring are separate. Read the frozen source for the rules behind the reported results.

  • reconciliation.xlsx exists — file_exists, required
  • live formulas — xlsx_has_formulas, required
  • no error cells — xlsx_no_errors, required
  • outstanding checks and deposits in transit totals — custom, required
  • adjusted balance — xlsx_value_present, required
  • both sides tie and items listed — custom, required

Recorded results, desk release

pass pass, frozen scorer only fail timed out
SystemRepetitionsFrozen / 3Original / 3Median minutesMean model estimate (USD)
Proto / DeepSeek V4.1 Flash228.40.1081
Codex / GPT-5.6 sol335.40.6420

These are repeated executions on the released fixture, not independent new business cases. Cost is captured model usage priced under the recorded assumptions, excluding infrastructure and human review. Per-attempt check verdicts, raw and frozen verdicts, usage, receipt and artifact hashes for this task are in results/latest/attempts.jsonl under "task": "bank-reconciliation".

Recorded results, desk comparison, 28 September 2026

SystemRepetitionsFrozen / 3Original / 3Median minutes
Proto / GPT-6 sol224.3
Codex / GPT-6 sol224.9
Proto / DeepSeek V4.1 Flash332.2

A separate campaign under the same frozen scorer, not pooled with the release. Conditions and data.