Document Ingestion in .NET: Throughput, Versions and Retrieval Quality
A bounded plain-text ingestion pipeline preserves versions, cancellation and searchable representation identity, with actual SQLite checks and small-corpus measurements.
THE READING ROOM
Writing & investigations.
One AI Task Contract: Local and Hosted Boundary in a Reference Lab
A real Blazor WASM and .NET ONNX comparison freezes tokens, model, pooling and candidate scope, then tests parity and failure across a loopback HTTP boundary.
Your Data Does Not Become Knowledge by Entering a Prompt
Retrieval supplies material for reasoning. Knowledge requires context, provenance and a defensible connection between evidence and decision.
Indexing JSON: Read Gains, Write Costs and Maintenance
A real SQL Server experiment traces the read benefit, update obligation and storage cost of a computed-column index over a JSON property.
One Million Events, Three Failure Modes
A throughput experiment is incomplete until it demonstrates recovery from duplicates, retries and interrupted processing.
A Recovery Runbook Should Refuse the Wrong Next Step
An executable recovery runbook rejects missing artifacts, wrong identities, stale history and conflicting replay before allowing the next step.
Migrating Agent Behavior, Not Just SDK Calls
A migration can compile successfully, pass a demonstration and still change the behavior that made an agent useful. The new implementation may choose different tools, forget state at a different boundary or retry an operation that the old implementation treated as terminal.
Reservation Confirmed or Work Accepted: Two .NET Architectures
A real HTTP and SQLite fixture shows why a persisted request and a confirmed reservation are different promises. Admission, worker processing and terminal outcomes are verified separately.
Following the Cost of One Task Through Azure
A reproducible C# allocation model separates source records, shared-resource policy and unresolved attribution. Synthetic units make conservation and rounding testable without claiming live cloud costs.
The Agent Says Done. What Does Done Mean?
An agent completion statement should point to observable acceptance, not merely announce effort. The definition of done must survive a handoff.
RAG or Broad Context: Testing Two Paths to the Same Authorized Evidence
Two real context builders share one eligible source catalog and a scripted Microsoft IChatClient boundary, exposing version, size and update contracts without claiming a live-model quality comparison.
An Observable .NET10 Service Before the First Incident
The first incident is a poor time to discover that the most important part of an operation is invisible. The HTTP request looks healthy. The queue accepted a message. The worker emitted a few logs. The customer still did not receive the result.