BalanceDocket

BalanceDocket is an evidence-bound month-end close reviewer for controllers and senior accountants. Rules calculate the exact treatment, GPT-5.6 performs the citation-bound interpretation in the judged flow, and the accountable human decides—all in one inspectable trail. If model access is unavailable, the same case fails safely to deterministic human review rather than blocking the free judge path.

This is the repository's OpenAI Build Week 2026 · Work & Productivity project. It is built on selected pre-existing Accounting Agent primitives; the Build Week provenance records that boundary.

Public entry: Watch the 172-second demo · View the Devpost submission

Compatibility note: BalanceDocket is the approved public display name. The released product deliberately retains the stable closeproof technical namespace for CLI commands, environment variables, schemas, plugin IDs, source paths, and the public repository URL. Those identifiers remain unchanged so the tested judge path and evidence hashes do not break.

Judge quick start

Prerequisite: Python 3.11+. From a fresh clone, one command generates the synthetic case and serves the checked-in competition bundle on loopback. It does not need Node.js, npm, an OpenAI account, a model, or a rebuild:

git clone https://github.com/Yurii201811/closeproof.git
cd closeproof
./scripts/run_closeproof_prebuilt.sh

To rebuild the reviewer from source first, install Node.js 20.19+ or 22.12+ and npm, then run:

./scripts/run_closeproof_demo.sh

Open http://127.0.0.1:4173. Set CLOSEPROOF_PORT=4187 if the default port is occupied. The script never contacts an accounting, bank, email, tax, payment, or filing system. Alternate ports use a port-scoped ignored output directory by default, so starting a second demo cannot reset the first demo's decisions:

Once open, verify the judge path in four steps:

1. Select Adjustments in the close ledger.
2. Inspect the source evidence, exact allocation, and optional GPT-5.6 advisory.
3. Enter a human rationale and choose a decision.
4. Download the validated, hash-bound workpaper.

CLOSEPROOF_PORT=4187 ./scripts/run_closeproof_prebuilt.sh

The default run stores state in .local/closeproof-demo; a non-default port stores it in .local/closeproof-demo-$CLOSEPROOF_PORT (for example, .local/closeproof-demo-4187). The launcher prints the resolved state directory and advisory case path. CLOSEPROOF_OUTPUT may be used to choose a different dedicated directory explicitly.

Install the optional Codex plugin

The local reviewer works without a plugin or an OpenAI account. To add the BalanceDocket workflow to a plugin-capable Codex installation from this clone:

codex plugin marketplace add .
codex plugin add closeproof@personal

Start a new Codex task so the installed skill is discovered, then invoke $closeproof. The plugin orchestrates the same local scripts and may optionally use the operator's existing Codex sign-in for the bounded advisory step.

Supported and tested platforms

• Supported runtime: macOS or Linux with Bash and Python 3.11+. The source-build route additionally needs a Vite-compatible Node.js release and npm.
• Verified on July 14, 2026: macOS 27.0 arm64, Python 3.11.5, Node.js 25.4.0, npm 11.7.0, Codex CLI 0.144.0, Chromium through the in-app Browser flow, Playwright Firefox/WebKit desktop and 320×800 reduced-motion smoke runs, and native Safari keyboard, VoiceOver, and maximum-zoom checks.
• Native Windows is not supported because the local event log uses POSIX file locking. Linux and WSL remain unverified release surfaces. Safari verification is limited to the macOS version above rather than a broad version-support claim.

What the demo proves

The golden path takes a bundled synthetic GL CSV and a deterministically generated text PDF through duplicate, cutoff, and prepaid controls; shows one ambiguous Adjustments exception; adds a citation-bound GPT-5.6 interpretation through Codex in the competition flow; requires a human approve/request-evidence/reject decision; and exports a hash-bound workpaper. The no-model fallback preserves the deterministic review but does not replace GPT-5.6's material competition role. Neither path performs an ERP write or accounting action.

Run the deterministic demo manually

Prerequisites: Python 3.11+, Node.js 20.19+ or 22.12+, and npm.

python3 -m accounting_agent.cli closeproof-demo
npm --prefix apps/closeproof-web ci
npm --prefix apps/closeproof-web run build
python3 -m accounting_agent.cli closeproof-serve

Open http://127.0.0.1:4173. If that port is occupied, pass another loopback port, for example --port 4187.

The default path is deterministic, model-free, API-free, and network-off. It does not require an OpenAI account. It writes only ignored local artifacts under .local/closeproof-demo:

• case.json — evidence-bound reviewer state;
• invoice_INV-4821.pdf — generated synthetic text PDF;
• manifest.json — inputs, outputs, snapshot, and zero-call counters;
• decision-events.jsonl and its chain head — created only after a human action.

The exact control result is SEK 5,260.27 June expense and SEK 114,739.73 prepaid asset from 16 of 365 inclusive service days. The generated case starts with the advisory not_requested; it contains no fixture or simulated model response.

Add GPT-5.6 advice with your existing Codex sign-in

The primary competition path uses Codex signed in through ChatGPT and the GPT-5.6 access included with the operator's eligible plan allowance. It does not need a separate API key or API billing account. Plan availability, rate limits, and usage terms still apply; subscription-backed use is not universally free.

The advisory commands are:

python3 -m accounting_agent.cli closeproof-advisory status
python3 -m accounting_agent.cli closeproof-advisory prepare
python3 -m accounting_agent.cli closeproof-advisory import
python3 -m accounting_agent.cli closeproof-advisory codex \
  --confirm-use-codex-allowance
python3 -m accounting_agent.cli closeproof-advisory api \
  --enable-network-advisory

Those defaults target the port-4173 state in .local/closeproof-demo. For a demo started on another port, pass its port-scoped case and keep the prepared request/manual response in that same state directory. For example, after starting with CLOSEPROOF_PORT=4187:

state_dir=.local/closeproof-demo-4187
case_path="$state_dir/case.json"

python3 -m accounting_agent.cli closeproof-advisory status \
  --case "$case_path"
python3 -m accounting_agent.cli closeproof-advisory prepare \
  --case "$case_path" \
  --output "$state_dir/advisory-request.json"
python3 -m accounting_agent.cli closeproof-advisory import \
  --case "$case_path" \
  --input "$state_dir/advisory-output.json"
python3 -m accounting_agent.cli closeproof-advisory codex \
  --case "$case_path" \
  --confirm-use-codex-allowance
python3 -m accounting_agent.cli closeproof-advisory api \
  --case "$case_path" \
  --enable-network-advisory

• status reports the current provider/provenance advisory envelope without making an advisory request.
• prepare writes the citation-bounded prompt, evidence packet, and response schema to .local/closeproof-demo/advisory-request.json by default.
• codex is the judged path: use the existing Codex sign-in, request the concrete Codex catalog model gpt-5.6-sol, validate the structured result, and retain its provenance. The optional Responses API route uses the gpt-5.6 API alias, which OpenAI maps to GPT-5.6 Sol; BalanceDocket does not silently substitute between provider identifiers.
• import validates a structured result produced from the prepared prompt. For the manual fallback, paste the prompt into ChatGPT, choose GPT-5.6, save only the requested JSON to .local/closeproof-demo/advisory-output.json, and import it. A ChatGPT subscription is an interactive product entitlement, not an API credential.
• api is an optional Responses API route. It must require explicit opt-in and OPENAI_API_KEY supplied through an approved secret mechanism. BalanceDocket must never print or persist the key.

All advisory routes may send or expose only the bundled synthetic source excerpts and deterministic amounts. Validation rejects unknown citations, changed amounts, malformed output, or a false cannot_approve invariant. Provider-authored prose is never persisted or rendered: BalanceDocket retains only the validated citations, uncertainty, exact amounts, and whether evidence was flagged, then generates controlled display language locally. Codex and API results must also pass the available model/provenance checks. A manual ChatGPT import remains labeled Unverified model identity unless its identity can be independently established. If advisory generation is unavailable, the local case remains reviewable and exportable with the advisory marked not requested or unavailable.

The Build Week recording must show a real Codex run requesting gpt-5.6-sol whose structured output passes BalanceDocket validation. Codex CLI 0.144.0 did not report the returned model identity in the retained run, so the repository does not claim that identity was independently verified. A fixture advisory, manual text pasted without validated provenance, or API-shaped mock response is not acceptable as competition proof. Run /feedback in the primary Codex task and retain the returned session ID for the submission.

Verify BalanceDocket

./scripts/verify_closeproof.sh

Or run the layers separately:

python3 -m unittest tests.test_closeproof
npm --prefix apps/closeproof-web test
npm --prefix apps/closeproof-web run build
python3 -m unittest discover

See PRODUCT.md, DESIGN.md, the Build Week plan, the problem validation, the provenance record, the demo script, and the submission checklist. The submitted Devpost record lives in docs/devpost_submission.md.

How Codex and GPT-5.6 contributed

Codex accelerated the Build Week work end to end: competition and practitioner research, repository mapping, product architecture, Python and React implementation, UI mockup iteration, browser checks, accessibility tests, security review, and plugin/release packaging. The human entrant selected the Work & Productivity track and controller persona, chose the evidence-first exception-review wedge, required the local/API-free default, kept calculations deterministic, prohibited ERP writes and model approval authority, and accepted the final product direction.

GPT-5.6 has a deliberately material but bounded role: interpret the ambiguous synthetic service-period exception, cite only the supplied evidence, preserve the exact deterministic amounts, state uncertainty, and flag missing evidence. BalanceDocket discards its prose and renders those validated selections through local controlled-language templates for human review. A retained Codex CLI run requested gpt-5.6-sol and its output passed BalanceDocket validation; Codex CLI 0.144.0 did not report the returned model identity, so this repository does not claim independent model attestation. The Build Week provenance records the detailed timeline, inherited-code boundary, acceleration, and human/model decision split.

Inherited Accounting Agent foundation

The broader repository is a Sweden-first, provider-neutral accounting automation foundation for safe local proposals, exception review, reconciliation, and evidence packaging. BalanceDocket deliberately uses only the bounded primitives documented in the provenance record.

The functional runtime still uses synthetic local fixtures. It extracts invoice fields, isolates duplicates by client, checks supplier and VAT risk, proposes BAS/VAT treatment, prepares bank-reconciliation proposals, writes human review packets, and records local evidence. It does not process real client data or post, approve, send, pay, delete, change ERP settings, file tax, or call a live provider.

Version 1 adds explicit legal-entity journals, evidence-bound approvals, dependency-aware close readiness, resumable preparation agents, guarded ERP read contracts, provider-neutral advisory-model routing, and a bilingual Guided/Expert cockpit. International support is a schema and jurisdiction-pack foundation, not a blanket compliance claim. Live provider writes remain forbidden.

Inspect that machine-readable boundary:

python3 -m accounting_agent.cli platform-status
python3 -m accounting_agent.cli platform-status --json
python3 -m accounting_agent.cli v1-system-check --json

Run the complete local demo:

python3 -m accounting_agent.cli demo-supplier-invoice-autopilot

The demo writes normalized intake cases, extracted invoice JSON, accounting proposals, risk findings, policy decisions, execution permits where applicable, approval packets, Fortnox dry-run payloads, gnubok shadow output, and an audit log to .local/demo_supplier_invoice_autopilot.

Run the full fake-client month dry run:

python3 -m accounting_agent.cli fake-client-dry-run

The fake-client run generates a synthetic Swedish consulting/service business month, runs supplier invoice autopilot and bank reconciliation locally, writes outputs to .local/fake_client_dry_run, and writes docs/fake_client_dry_run_report.md. It uses no real client data and no live external APIs.

Build the read-only local operations cockpit:

python3 -m accounting_agent.cli build-operations-cockpit

The cockpit writes reports/operations_cockpit/index.html. It is a local, read-only Workbench generated from repo artifacts. Guided mode gives accountants one clear next decision; Expert mode exposes policy, provider, specialist-agent, and evidence detail. The main interface, review reasons, filters, workflow, provider matrix, and dates switch between English and Swedish. The selected guidance perspective changes explanatory emphasis only; risk priority and safety policy remain fixed. It has no CDN or live-provider dependency and remains observe/review only.

Build the separate deployable preview from built-in synthetic examples only:

python3 -m accounting_agent.cli build-public-preview \
  --output .local/accounting-agent-v1-preview

The public builder never reads local workflow artifacts and refuses unexpected files in its target directory.

Private course references, if present, stay under .local/course_reference/ and are excluded from Git. They can inform topic coverage and plain-language explanations, but current official law, BFN, Skatteverket, BAS, and SIE sources remain authoritative.

Run the lower-level sample pipeline:

python3 -m accounting_agent.cli process-fixtures

Run tests:

python3 -m unittest

Fixture commands use an explicit 2026-05-16 evaluation date by default so risk outcomes do not change with the wall clock. Pass --as-of YYYY-MM-DD to the supplier fixture or demo command when testing a different review date.

License

BalanceDocket is available under the MIT License. Copyright © 2026 Yurii Bakurov. Runtime open-source components and their license texts are listed in THIRD_PARTY_NOTICES.md.

Pre-existing and open-source components

BalanceDocket meaningfully extends selected primitives from a pre-existing Accounting Agent foundation. Only the dedicated BalanceDocket case, controls, snapshot and citation contract, bounded GPT-5.6 evaluator, human decision chain, React reviewer, Codex plugin, workpaper export, tests, and release material added during the Build Week submission period are presented as the competition contribution. The dated boundary and file-level evidence are in the Build Week provenance record.

The reviewer uses React, React DOM, Scheduler, and Lucide React under their respective open-source licenses. Exact runtime versions, upstream projects, copyright notices, and license texts are recorded in THIRD_PARTY_NOTICES.md; the npm lockfile retains the complete dependency graph.

Start with docs/accounting_agent_v1.md for the product, jurisdiction, ERP, model, autonomy, computer-use, and release boundary. See docs/local_model_evaluation.md for the local Ollama comparison, docs/v1_research_decisions.md for the problem-to-product decisions, docs/local_demo_supplier_invoice_autopilot.md for the one-command local demo, docs/mvp_supplier_invoice_autopilot.md for the packet schema, and docs/operations_cockpit.md for the interface workflow.