What is live, what is built, and what is being built.

Live on our hosted reference deployment today

1. Spend by person, team, product, project, site, provider and gateway, including coding agents and direct contracts; provider, gateway and device readings kept apart; unpriced tokens shown as a floor; unseen sources named as gaps.
2. The AI spend with no business case, listed first as cost with no value claimed.
3. Business cases with baselines locked at approval, signed by the named finance approver.
4. Forecast, provisional and confirmed never added together; confirmed means an allocation proposed by an owner or administrator (never the case's own finance approver), checked against an imported ledger variance bridge and approved by each case’s named finance approver. This is reviewed attribution, not independent causal proof.
5. The month close: proposed by an owner or administrator, decided by a different person with finance authority, receipted; corrections go to the next period.
6. The retire ledger: overdue parallel runs priced in dollars a month, systems kept on purpose.
7. Approved policy plans prepared for supported gateways and coding-agent hooks, and compared with the configuration each target reports back. Kong OSS 3.9.3 has been exercised in a local lab for policy apply, readback and drift recovery. That run did not establish successful model inference. Azure API Management and the coding-agent hooks are tested against each vendor’s documented interface. A policy is published after a second approver, and drift is shown. On managed coding-agent machines, sensitive-data and MCP-server rules can run in observe, then warn, then enforce, and the baseline enrollment, identity and model rules can be set to observe, warn or enforce for the organization or a group. At Kong and Azure API Management a rule goes in as enforce: it can be replayed against the last 24 hours of recorded decisions with Preview impact, and a second person approves it before it is published. Team and person layers can only tighten, in enforce mode.
8. Why last month changed, in dollars by driver; next month with an 80% interval; the pre-invoice alert; commitments and capacity; the right-sizing ledger ranked in dollars, with quality on the cheaper model marked as unmeasured until there is evidence.
9. Calibration from the organization's own closes, shown as its own reading and kept within that organization.
10. Ed25519-signed, hash-chained receipts for every registration, approval, policy change, sign-off, close and export, verifiable offline with ace audit verify.
11. The allocation journal and a cost export in the FinOps FOCUS 1.4 format, priced at published list rates and labeled so: files finance takes away.

Built, with the first run in a design partnership

1. Ask ARC1: the model picks evidence and the server writes every sentence and figure. It runs once your deployment connects a model provider.
2. The Copilot license and credit ledger: Microsoft 365 Copilot, GitHub Copilot and Copilot Studio, from collectors tested against the vendor’s documented interface.

Being built, in priority order

Our next proof target is 6 November 2026: one workflow on Azure, Kong and Azure API Management, with named evidence for the decisions it supports. These are acceptance milestones, not completed customer integrations.

1. Install ARC1 in a real Azure subscription, with Microsoft Entra sign-in.
2. Read Azure API Management token logs, plan an approved policy change and read back the target configuration.
3. Run Claude Code and Codex hooks on our own managed machines and retain their real usage evidence.
4. Show MCP server registration and approval end to end.
5. Demonstrate routing, token optimization, data sanitization, governance and token-spend management on one consistent dataset. Evaluate quality and cost for one workflow before claiming an improvement.

Finance evidence follows this stack proof: reconciled versus attributed value, the signed close pack and its offline verifier. Additional connectors and model-training work stay outside this proof scope.