A RAG command-line tool that answers questions about a project's decisions from its specs and decision records. An LLM extracts each spec into typed, cited items in a Neo4j graph. Every answer cites those items, with each citation checked in code against the items the model was sent. Every paid call is checked against a cost ceiling before it is made.

My first design walked links between documents and answered 1 of 5 eval questions, so I measured why, rebuilt it around retrieval plus one checked model call, and froze the prompt before running questions it had never seen. On all nine questions, the five eval ones plus four held out, my earlier tool, decision-memory, abstained on every one; tracepath gave an answer on all nine. That compares two whole systems on different models and different inputs (decision-memory reads only each spec's index.md and rationale.md), so it is not a verdict on graphs versus vector search.

A multi-user job search platform, built from the stack decision onward. OAuth sign-in through Google and GitHub, real-time search against the Adzuna API, and application tracking, backed by a Postgres schema with row-level security enforced on every table. Every account's usage is metered by an atomic gate that holds exactly at its limit under real concurrent load: with 15 accounts against a cap of 5, it let exactly 5 through and none slipped past. GitHub Actions runs lint, type checks, and unit and integration tests against a real local database stack on every pull request, and applies schema migrations on merge.

I tested the failure handling by breaking things on purpose. In development, a forced-failure chain confirmed a Sentry alert reaches a delivered email. A rollback drill, run early on the scaffold homepage, broke the live production site and restored it by promoting the previous build.

It scores real listings against a user's profile through GPT-5.6 Luna in production, with a deterministic filter that drops any skill the model invents but the caller doesn't actually have. I built an eval harness and measured whether the scoring holds up instead of assuming it: of 16 authored ground-truth pairs, 15 landed in their expected band. The ground truth was authored from the same rubric the scorer reads, so this measures consistency with that written definition, not independently validated real-world accuracy. That is the honest scope of what a first-version internal eval can prove. One of the sixteen was corrected after review, a wrong expected answer in the test data itself, not a change to how the model scores. The 15 of 16 is from the run before that correction; I haven't rerun the full set since. The remaining miss traced back to a specific, repeatable gap in the rubric's calibration, and I flagged it for the next revision rather than writing it off as noise.

A locally indexed, cited RAG system that makes software decision history queryable. When the evidence doesn't support an answer it says so instead of inventing one, and the README documents the rate at which it still answers wrong. A per-sub-claim verification layer runs exact containment checks and LLM verification, then drops any sentence its cited evidence doesn't support. I've paused development.

The conformance suite fails any adapter that emits a confident record from malformed, empty, or invalid-UTF-8 files. A live evaluation harness runs the system against a real corpus, checks whether it answers or abstains as expected and cites the required records, and reports its failures openly.