FINANCE AGENT LAB — OPEN-SOURCE REUSE AND COMPLETED BUILD RECORD October 2, 2026 CURRENT RELEASE 32 working prototypes, 96 editable synthetic scenarios, 7 working finance domains, and 8 proposed agents across 9 catalog domains. The five-agent finance expansion is complete. Each working agent has three synthetic cases. The Reporting & FP&A spotlight offers three selectable visual stories and a four-agent toolkit, including Billing Completeness Reviewer. SHIPPED HELPER ADAPTATIONS 1. opensyndicate/dispute-agent — MIT Pinned b231843853633d27015661a644ac128334f73c02. https://github.com/opensyndicate/dispute-agent/tree/b231843853633d27015661a644ac128334f73c02 A small requirement/alternative coverage helper is adapted from reasons.py and caseengine.py. The lab adds original synthetic cases, scope/conflict gates, refund-history checks and visual stories. Upstream fight/accept scoring, narrative assertions and integrations are not used. Full copyright and MIT notice are shipped. 2. FoundrySoftHQ/agent-for-accounting — MIT Pinned 7a9e8f9a81eb314b70418051b75213cf6230a01f. https://github.com/FoundrySoftHQ/agent-for-accounting/tree/7a9e8f9a81eb314b70418051b75213cf6230a01f A modified amount/date candidate helper supports comparisons. Accepted matching still requires the lab's original obligation-identity and scope rules. No greedy matching or upstream integration is used. Full MIT notice is shipped. 3. esploro-group/closegate — Apache-2.0 Pinned 443520dea2f417b2e81fd163c6997762281a103c. https://github.com/esploro-group/closegate/tree/443520dea2f417b2e81fd163c6997762281a103c A modified supplied-identifier separation predicate supports the journal review sample. It does not authenticate identities, authorize posting or integrate the upstream platform. Complete license and attribution are shipped; no upstream NOTICE was present in the pinned tree. 4. Nixtla/utilsforecast — Apache-2.0 Pinned 9c88409d2e2efcf8808fad64754160bbcfab47c2. https://github.com/Nixtla/utilsforecast/tree/9c88409d2e2efcf8808fad64754160bbcfab47c2 Modified MAE and forecast-minus-actual bias helpers operate only over accepted observed flow pairs. The lab supplies exact identity, direction, date and source-coverage checks. Complete license and attribution are shipped; no upstream NOTICE was present. The metrics are not accuracy percentages or confidence estimates, and no forecast model is integrated. 5. JeffBrines/openfpa — MIT Pinned ddac4e90edc7b93392d94f33c8f579100fe3f7e7. https://github.com/JeffBrines/openfpa/tree/ddac4e90edc7b93392d94f33c8f579100fe3f7e7 A small ratio-of-sums helper is adapted from extract_lines / aggregate_periods in pyfpa/backtest/score.py and ships in dist/reporting-tools.js. It preserves numerator and denominator totals before division. The dependency-free JavaScript adaptation adds integer-input and safe-total validation; empty inputs and zero denominators remain unavailable. The lab supplies its own metric definition, unit normalization, component identity, scope, lineage, period and coverage checks. Copyright (c) 2026 Guiderail. The complete MIT license and attribution are included in THIRD_PARTY_NOTICES.md and the shipped notices; the full license is also distributed as openfpa-LICENSE.txt. OpenFPA's treasury/forecast integration, weekly forecast/report layouts, Python runtime and report-writing behavior are not used. The earlier Treasury evaluation did not adopt them. Reporting now uses only the ratio-of-sums helper described here. COMPLETED WAVES Fraud & Payments: Payment Anomaly Triage, Payee Change Risk Reviewer, Chargeback Evidence Assembler. Payables & Close: Supplier Statement Reconciler, Purchase Order Commitment Analyst, Intercompany Balance Reconciler, Journal Entry Support Reviewer. Treasury & Receivables: Cash Position Consolidator, Forecast Reliability Monitor, Short Payment Investigator, Collections Priority Planner. Reporting & FP&A: KPI Consistency Reviewer, Management Reporting Composer. These earlier waves added 13 agents to the 14-prototype suite, reaching 27 prototypes and 81 cases at that stage. The five-agent finance expansion below brings the current suite to 32 prototypes and 96 cases. Every added agent has three editable scenarios, a distinct visual story and a reviewer packet. Seven working-domain spotlights support quick browsing. ORIGINAL LAB WORK Synthetic fixtures, typed evidence and scope gates, source versioning, numeric validation, allocation and overlap rules, finance calculations beyond the small adapted helpers, visual stories and reviewer boundaries are original. Management-reporting narratives are original deterministic templates. Arithmetic contributors do not establish business causes; unsupported commentary remains a question. No LLM or upstream report generator is used by the standalone samples. EVALUATED VISUAL AND DATA CANDIDATES — NOT INTEGRATED D3 (ISC) remains a candidate for more involved plots and scales; it is not installed or copied. The current release uses original lightweight SVG stories and the adapted Canvas context mesh. https://github.com/d3/d3 IBM AMLSim (Apache-2.0) remains a possible synthetic transaction-pattern reference; its runtime is not part of the site. https://github.com/IBM/AMLSim These candidate projects are separate from the five shipped helper adaptations above. REUSE STANDARD Verify source and license at a pinned revision before copying. Preserve copyright, license and any applicable NOTICE text. Distinguish code adaptations, original fixtures and design references. Port the smallest useful component and retain existing architecture unless a capability requires more. Test decision-changing evidence cases and verify desktop, mobile, keyboard and reduced-motion behavior. PROJECT BOUNDARIES All data and policies are synthetic and illustrative. Standalone samples use deterministic calculations and original template explanations in the browser. They do not query live banks or processors, authenticate evidence, submit disputes, post entries, update suppliers, contact customers or publish reports. The deeper connected invoice workbench remains available separately. Its optional provider integration does not turn standalone narratives into model-generated output. Human decisions and financial actions remain outside these demos. FINANCE EXPANSION / FIVE WORKING AGENTS + EIGHT PROPOSED The fraud network, fee audit, bank/GL, multi-invoice application and billing coverage engines use original dependency-free JavaScript. No NetworkX, DuckDB, GX Core or River library is imported by these five. Future evaluated foundations: NetworkX for offline graph validation; DuckDB-Wasm for browser query execution over finance record populations; GX Core for declarative feed validation patterns; River for a later statistical behavior model. These are selected capabilities to evaluate at the relevant build wave, not finished agents or deployed integrations. Pin and verify the selected revision, retain notices and validate decision-changing cases before importing code. Read next-agent-roadmap.md for the three planned waves and acceptance cases.