Leadership Under Uncertainty · Reference architecture RA-1

The judgment layer on the Microsoft stack

The stack is unusually complete. SharePoint and Fabric hold the documents and the data. Copilot Studio and Foundry build the agents. Word and Teams are where the work happens. An enterprise can put an agent in front of its whole document estate today.

What it has nowhere to put is the judgment: that the organisation believes X at 0.72 confidence, that the belief rests on an assumption nobody verified, that a minority dissents, that it moved from 0.61 last month on two specific filings, and that the last time it reasoned this way it was wrong by two quarters. This is a runnable demonstration of that layer.

The loop

Ten steps, ten artifacts

Every step deposits something structured and addressable rather than a message in a thread. That is the whole difference.

#	Step	Produces
1	Documents ingested from a document library	Claims, evidence, sources, assumptions
2	Views built with confidence and posture	View artifacts
3	Driver questions attach calibrated probabilities	Prediction artifacts
4	An agent requests the current view over MCP	Entity Snapshot
5	The agent asks what changed	Analytic Delta
6	A memo is reviewed against ICD 203 inside Word	Findings: assumptions, alternatives, gaps
7	A decision is recorded, naming the option taken	DecisionRecord
8	The decision is content-hashed into the chain	Integrity record
9	The outcome is returned against the decision	OutcomeObservation
10	The day is sealed; a third party verifies offline	Merkle root + evidence pack

Model-agnosticism

The demonstration issues the same two calls from a Copilot Studio agent and from Claude Desktop. Same endpoint, same tool names, same arguments, same returned artifacts; no server-side branch distinguishes them. If the layer’s value evaporated when the model underneath changed, it would be a prompt-engineering artifact — and those are reset by every model generation. Running it against two model families and showing the artifacts are identical is how that claim gets established rather than asserted.

Demonstrated versus productized

Some of this is working software with tests; some is a scaffold; some does not exist. A reference architecture that blurs the three is worse than none.

Word task-pane add-in (ICD 203 review)	Working
SharePoint document ingestion	Working
MCP server and the tools this loop uses	Working
Decision sealing and offline verification	Working
The loop end to end	Built, not yet run
Copilot Studio agent registration	Documented, unverified
SharePoint SPFx web part	Scaffold
Teams delivery	Orphaned code
Fabric / OneLake ingestion	Not built
Foundry agent registration	Not built

Run it yourself

All eight source documents are fictional — an invented company, counterparty and decision. The corpus is shaped to exercise the layer rather than to read well: it carries a planted untested assumption, a deliberately overconfident consultant memo, a recorded dissent, and a prior analogous decision that failed on exactly the assumption the current recommendation repeats.

npx tsx scripts/demo/seed-ms-reference-demo.ts
npx tsx scripts/demo/ms-reference-loop.ts
npx tsx scripts/attestation-verify-pack.ts <pack.json>

The loop is built and its scripts are committed, but it has not been run end to end: a run writes to a live database and spends model budget. The published transcript says so rather than presenting a rehearsal as a record.

Not a partnership

This is published demonstration content. Vendor product names are used nominatively to describe what the demonstration runs against. Nothing here implies a relationship with, or endorsement by, any vendor.

Want this for your team?

Drop your email and we'll be in touch when access opens for your vertical.