#AI
When Code Bends to Pass Its Tests, the Tests Are the Finding
SWCraft's tests tied the architecture model's contract to source code by line number: 5,843 file.py:line citations, all checked. The code froze around them. What the citations really proved, the switch to symbol citations, and a test rule rewritten to judge what a test asserts rather than what it reads.
Read Post
The Agent Did Not Need a Better Prompt. It Needed the Manual.
Asked about SWCraft itself, an agent on the weakest available LLM got about one question in eight right, and some of its answers were invented. Given the same user guide a person reads, it got 39 or 40 of 40, on three hosts. How the answers were checked without an LLM judge, and what went wrong on the way.
Read Post
Shrinking MCP Answers Bought Nothing. Answering the Question Did.
A Copilot agent on SWCraft’s MCP server made 27 calls for a five-step change, and only one of them was the change. Six measured steps later an architect task costs about three calls, on a showcase and on a bench thirty times its size. What each step taught, including the two that failed.
Read Post
Trace Links Without AI: A Few Labels per Document Did the Work
Linking thousands of requirements to an architecture model looks like a job for an LLM. A pilot on a real automotive project showed that a few hundred decisions by the architect, placed well, do the work, and AI stays on the shelf.
Read Post
How AI Agents Turned a Bug into an Approved Decision
An agent wrote an implementation limitation into an accepted architecture decision. Another preserved it. When a test exposed the damage, the agents changed the test. I caught it in the summary.
Read Post
My MCP Server Had 39 Tools. I Had Designed It for the Wrong Caller.
SWCraft grew to 39 MCP tools by exposing model operations one tool at a time. Separating the software API from the agent interface reduced the surface to 13 without removing capability.
Read Post
An LLM Wrote the Code. ISO 26262 Doesn't Care — and That's the Point.
ISO 26262 tool qualification never asks who wrote the code, because it never trusted the tool in the first place. If the confidence argument does not depend on the AI being right, AI-assisted development and qualification stop being in conflict.
Read Post
Nothing Changed in the API. Everything Changed for the Agent.
An MCP tool schema explains how to call a function, but not when that function should become a candidate. A task-scoped semantic contract can keep a growing tool catalog behaviorally compatible.
Read Post
An Agent Can Pass Validation and Still Make an Unsafe Change
The agent was authorized to modify one component. It helpfully modified two. The model still validated. Why MCP capability is not task authority, and how a task-scoped mutation boundary closes the gap.
Read Post
The Requirements RAG Router Failed. Then the Benchmark Failed Too.
A requirements RAG reliability experiment: heuristic routing failed on independent paraphrases, query grounding exposed a broken benchmark taxonomy, and typed evidence plans produced a narrow but useful result.
Read Post
Everyone Is Building an AI AUTOSAR Toolchain. But How Should We Evaluate One?
Generating plausible AUTOSAR artifacts is an interesting capability. It is not yet evidence of a production-ready engineering system. We need a way to measure the difference.
Read Post
Hybrid Databases Were the Wrong Fix for Requirements RAG
A negative infrastructure result from a requirements RAG lab: Postgres, ParadeDB, and OpenSearch could host useful pieces, but none replaced the current retrieval stack without losing ranking quality or evidence bundles.
Read Post
The RAG Verifier Learned to Say No. It Still Missed a Hard Requirement Failure.
A held-out requirements RAG experiment: claim-level verification reduced harmful published answers from 35.7% selective risk to 4.0%, but strict human audit still found semantic completeness failures.
Read Post
Citations Were Not Enough for Safety-Critical Requirements RAG
A negative result from a requirements RAG lab: even with structured records, cited JSON answers, and deterministic guards, the answerer failed the abstention gate when evidence was insufficient.
Read Post
I Flattened 13,244 Requirements Into Chunks. 97% Lost Their Meaning.
A practical retrieval study over structured automotive requirements: why flatten-and-chunk failed, where ordinary top-k search stopped being the right tool, and how query routing improved both quality and latency.
Read Post
I Tested a Simple RAG Idea: One Summary per Document. It Was Useful, But Not Enough.
A practical retrieval experiment: can an LLM-generated document summary replace chunk-level search? On a 5,000-document corpus, the answer was measurable — useful signal, clear ceiling, and a better baseline.
Read Post
My Two AI Agents Talk MCP to Each Other. There Is No Standalone Tool Server.
The default MCP story is one fat tool server, many clients. When I needed two of my domain agents to talk to each other, that shape was the wrong one — and the alternative shows what MCP is actually good for.
Read Post
I Built an AI Agent for Myself. My Colleagues Wanted to Use It. That Is Where the Problems Started.
When a domain expert builds a working AI agent and people start depending on it, the bus factor problem arrives long before the platform — and a different kind of work begins.
Read Post
The Boring Part of Requirements Review Is Now Automated
How a lightweight AI agent eliminates the context-gathering overhead in safety-critical requirements review — so the architect can focus on judgment, not tool-switching.
Read Post
Building an AI Agent for Embedded Systems Architects
A conversational AI agent that sits in front of complex embedded systems tooling and speaks human — so architects spend less time remembering plugin syntax and more time on actual architecture.
Read Post