Cold War: Decided

The project began as an interactive Cold War decision simulator and evolved into a production test bed for auditable multi-agent AI. The goal became more than generating an answer — the system needed to:

• isolate evidence between specialists;
• preserve uncertainty and competing hypotheses;
• challenge conclusions through feedback and red-team review;
• produce one decision-ready recommendation;
• enforce high-consequence action boundaries;
• require human authorization;
• preserve every stage for later inspection; and
• evaluate decision correctness separately from writing quality.

1. Interactive command-center prototype and deterministic demo mode.
2. Live AI pipeline with four specialist agents.
3. Chief feedback, controlled revisions, and Counterintelligence review.
4. Protected serverless APIs, structured outputs, retry handling, and secret management.
5. Neon Postgres mission ledger and protected Run History.
6. PDF scenario ingestion, evidence routing, page citations, and document fingerprints.
7. Scenario adjudication, independent Decision Auditor, deterministic policy matrix, regression testing, and production validation.

1. A built-in brief or uploaded PDF enters the system.
2. Evidence is routed into four separate silos: Submarine Intelligence, ELINT, Air Intelligence, and HUMINT.
3. Four specialists produce independent initial reports.
4. The Chief gives each specialist targeted feedback.
5. Each specialist completes one controlled revision.
6. Counterintelligence red-teams the combined evidence.
7. The Chief creates a structured proposal.
8. A deterministic software policy matrix calculates eligible actions.
9. An independent Decision Auditor challenges the proposal and its evidence.
10. The Chief produces a final synthesis.
11. The software recomputes the policy result and builds a coherent Action / Why / Avoid / Reconsider decision block.
12. A human remains responsible for authorization.
13. All 16 stage outputs, evaluations, telemetry, and provenance records are preserved in the mission ledger.

Production testing showed that the Chief repeatedly selected the same response across materially different scenarios. Additional prompt instructions changed the preferred default but did not reliably eliminate the bias.

The persisted run history made it possible to inspect the exact evidence, structured factors, Auditor feedback, and selected actions for every run. That diagnosis showed prompt-only control was insufficient — the fix couldn't just be a better instruction.

The final solution used a hybrid AI and deterministic policy architecture:

• The AI interprets evidence, identifies uncertainty, and populates structured factors.
• Deterministic software calculates which actions are eligible.
• Explicit action contracts protect consequential choices.
• The Decision Auditor challenges unsupported factors.
• The human retains final authority.
• The complete decision block is generated from the policy result so the label and explanation cannot contradict each other.

• Both completed all 16 expected reports.
• Both received 100% scenario-specific Decision Accuracy in the final production validation.
• Two materially different evidence patterns produced two different actions.
• The final action explanations matched the enforced decisions.
• Build checks and production runtime logs were clean.

These results apply to the supplied benchmark scenarios — not a claim of universal accuracy, defense-grade reliability, or statistically comprehensive validation.

The Cold War scenario is fictional, but the workflow is directly transferable to accounting, tax, finance, compliance, and professional-services operations. In those settings, the evidence changes from intelligence reports to client documents, transaction records, tax notices, workpapers, policies, and financial data. The same architecture can route information to specialized reviewers, apply deterministic firm rules, require human approval, and preserve a defensible audit trail — the kind of foundation a project like FirmOS would build on.

Example reuse: a tax-notice response workflow

All product names below are possible future integrations, not part of this project's actual technology stack.

1. A tax notice enters through a client portal, SharePoint, OneDrive, SmartVault, or email intake.
2. An OCR/document product (e.g. Azure AI Document Intelligence, Google Document AI, Amazon Textract) extracts dates, agencies, balances, periods, and notice codes.
3. The system retrieves relevant client information from QuickBooks Online, Xero, NetSuite, Sage Intacct, a tax platform, or the FirmOS database.
4. Specialist agents separately review notice classification and deadline, client account history, tax-law or procedural requirements, financial reconciliation, and missing documents and risk.
5. A manager agent synthesizes the findings.
6. Deterministic rules enforce deadlines, approval authority, materiality, required documentation, and actions the AI may not take.
7. A technical reviewer or partner approves the response.
8. The workflow creates or updates a task in a practice-management system (e.g. Karbon, Financial Cents, Canopy, TaxDome).
9. A draft response is created in Microsoft Word or Google Docs.
10. Final documents may flow through DocuSign or Adobe Acrobat Sign.
11. Communications may be delivered through Outlook, Gmail, Microsoft Teams, or a secure client portal.
12. Every source, revision, approval, and final action remains in the engagement audit trail.