01. Define
Build the input, questions, and every allowed option. Include a way to abstain.
PayItemQuestions.ForBID REVIEW QUEUE / 15-MINUTE WALKTHROUGH
Decisions,
not chat.
Let a model classify. Let your .NET application decide when to trust, escalate, or ask a person.
THE BUSINESS PROBLEM
35 lines. Not 35 easy decisions.
An estimator needs to map bid items to utility systems and work types before pricing. Some descriptions are clear; others need judgment. The goal is a reviewable classification, not awarding a contract, pricing a bid, or sending it.
02
THE RESPONSIBILITY BOUNDARY
The model answers. Your code owns the rule.
02. Classify
Microsoft-Decision-1 returns a choice and per-option probabilities, not a written rationale.
state + questions03. Validate & route
Check the response contract. Apply the threshold. Ask a second model only for eligible uncertain lines.
ReviewPolicy.Evaluate04. Resolve
Review unresolved cases. Changing an answer invalidates its previous approval.
Review, correct, approveThese calls are orchestrated by the sample, not by an automatic Foundry workflow. The lab uses Blazor Server and a Minimal API with shared services; GitHub Pages hosts only this static guide.
Why no agent framework?
Each item needs bounded classification, with no autonomous tools, memory, or conversation. A model cascade with human review is enough. Interfaces separate live calls from replay; an agent framework would add orchestration this task does not need.
03
ONE CASE BEFORE THE WHOLE DATASET
Would you accept item 13?
PAY ITEM 13 / SYNTHETIC
Sanitary Sewer Bypass Pumping, Complete
Unit: LS (lump sum). The model receives the number, description, and unit. Quantities and prices stay in the .NET application.
Pause here. Is this gravity sewer work or force main work? The application offers those categories; the model cannot introduce a new accepted category.
FIRST MODEL / RECORDED
Microsoft-Decision-1
SystemForce main
0.7078selected-option probability
not measured accuracy
not measured accuracy
Work typeSupport work
Work-type probability: 0.9954. The weakest selected probability drives the threshold rule.
SECOND OPINION / RECORDED
gpt-6-luna
SystemSanitary gravity
Different system. Same work type.
Work typeSupport work
The second opinion supplies choices, not a replacement probability. Disagreement remains evidence even if you lower the threshold.
CHANGE THE POLICY, NOT THE EVIDENCE
Where would you draw the line?
The initial threshold is 0.90. Explore only this recorded case; this is not the full app's review queue.
APPLICATION ROUTE AT 0.90
A person must review
The system probability is below 0.90 and the recorded models disagree. Neither a valid response nor a second opinion proves correctness.
Inspect the original choices and probabilities
A reduced view of the bundled recordings, not a new response. The separate service confidence field is not used by the threshold rule.
Enable JavaScript to inspect the recorded JSON, or open the recording in the repository below.04
FAMILIAR .NET, A PROBABILISTIC DEPENDENCY
Follow one decision through C#.
Start with the existing endpoint. Blazor calls the same service and policies directly; the UI does not need an HTTP round trip to this endpoint.
Program.cs / POST /api/decideC#
var outcome = await decisions.DecideAsync(request.Item, request.Mode, request.ReverseOptions, ct);
var result = outcome.Result;
var review = ReviewPolicy.Evaluate(request.Item, result, outcome.Questions, request.Threshold);Define a closed contract
PayItemQuestions.For creates the questions and options. unclear permits abstention. A span copied from text can still be the wrong span.
Treat the model as an external dependency
IDecisionProvider separates Foundry from replay. The service uses explicit mode selection, cancellation, timeouts, and visible errors. A failed live request never becomes a silent replay.
Validate before applying policy
DecisionWire.ParseResponse checks each question, offered choice, and finite probability. Every offered option must be scored, with a sum of 1 within the application's 0.02 tolerance.
Keep business rules in ordinary C#
The rule accepts only when every selected probability meets the threshold, every choice is most probable, and none is unclear. Reviewer edits require approval again.
Add the second opinion after the basic flow is understood
Only eligible lines below the threshold reach the second model. Abstentions and non-top choices go to a person instead. All choices must agree for acceptance by agreement; a failed second opinion leaves the line with a person.
Program.cs / selective escalationC#
if (request.SecondOpinion && EscalationPolicy.Eligible(result, outcome.Questions, request.Threshold))
{
try { opinion = await decisions.SecondOpinionAsync(request.Item, outcome.Questions, request.Mode, ct); }
catch (DecisionFailure failure) { secondOpinionError = failure.Message; }
}A model cascade with human review describes this sample's design, not a built-in Foundry approval feature. The JavaScript experiment above illustrates only item 13; the .NET policies remain the application implementation.
05
YOUR LIVE PRESENTATION
One story. Five stops.
- 00:00 / 2 MIN
Start with the problem
"We are classifying bid lines, not approving contracts." Introduce the application, model, and person as separate responsibilities.
- 02:00 / 4 MIN
Follow item 13
Ask the audience to predict. Reveal the recorded choices, their disagreement, and the review route. Name the recorded mode before showing the results.
- 06:00 / 4 MIN
Open the C#
Show the endpoint, the parser, and the policy. "The model is a dependency. Validation and business rules are still our job."
- 10:00 / 3 MIN
Change policy; then demonstrate review in the app
Move the threshold without a model call. In the separate .NET app, approve a line, edit an answer, and show approval clear. Export originals separately from edits.
- 13:00 / 2 MIN
Close with limits, not a success rate
"Two models can be wrong together. Choose a threshold on representative labeled data, not because 0.90 sounds reassuring."
Launch the separate .NET app for the review and export steps
From a repository checkout in PowerShell 7, with the exact SDK from global.json installed:
.\scripts\Test-Prerequisites.ps1
.\Start-Demo.ps1 -Lab 2 -OfflineOpen http://localhost:5190. At 0.90 with recorded second opinions enabled, the bundled demo expects 35 decided: 24 by rule, 8 by agreement, 3 awaiting review. Confirm these during rehearsal; new live calls need not reproduce them.
Restore and compile before presenting. -Offline disables live configuration, not network access for package restore. Stop with Ctrl+C and remove exported synthetic JSON when no longer needed.
06
WHAT THIS DEMO DOES NOT PROVE
Keep the evidence in view.
Probability is not accuracy
A 0.90 prediction does not establish 90% correctness. Calibration is measured across comparable labeled cases, not inferred from one answer.
Agreement is a policy
Two models can share an error. Agreement does not create a new calibrated probability or replace meaningful human review.
Replay is not availability
These October 14, 2026 recordings do not establish current Foundry availability, latency, pricing, or model quality on your workload.
This is not production approval
The app keeps review state in its Blazor session. It is not a durable audit trail. This guide stores no approvals and sends no bid data.
Sources & version boundary
The sample targets .NET 11 RC1, Microsoft-Decision-1 version 1, and the Microsoft provider's v1/systemone API. The sample treats the decision model as preview: its recorded catalog observation and documentation signals differ. See the source register rather than assuming general availability.
- Microsoft Learn: deploy and use Microsoft-Decision-1Typed decisions, probabilities, and workload-specific evaluation. Accessed October 15, 2026.
- Lab source registerOfficial references, API versions, preview uncertainty, and measured observations.
- Verification recordDated runs and limitations. Synthetic evaluation is not an industry benchmark.
- Full code walkthroughFollow the request, authentication, validation, policy, and second opinion.
Build it. Understand it.
Make it reproducible.