A comprehension artifact. Built from docs/fulcra-reconstruction-2026-08-05.md, which was itself compiled by reading code, git, gh, in-repo documents and session transcripts.
repo state c29f775 · branch overfetch-24 · origin/main d621cfa · compiled 2026-08-05
← the essay this came out of · Fulcra
Entry point runRetrievalPipeline() at src/mastra/rag/retrieval-pipeline.ts:611, called once per advisory turn, returning relevantContext, sources and rerankStatus. The file is 1,049 lines. Click any stage. The eleven stage names are a typed union in stage-debug-types.ts, so drift between the pipeline, the dev script and the span names becomes a compile error.
Four things in this pipeline have an experiment behind them. Everything else was chosen once, ported from the earlier Convex implementation, or copied from a paper.
| Component | Status | Evidence, or what is missing |
|---|
Dates are commit dates, PR merge timestamps, issue timestamps or artifact mtimes. Reversals are marked and are worth more than the decisions, because each one records a belief that was held confidently and turned out to be wrong.
Thirteen recorded cases where the belief and the finding came apart. Every row here is verified against the artifact named in the last column.
| What was believed | What was found | Recorded |
|---|
Every audit in this project was performed by AI agents. No human reviewer appears anywhere in the record. The human role throughout was direction, acceptance and pre-registration authorship, not review. Where an audit recorded totals but not per-item severity, the individual items carry no severity tag rather than an invented one.
Stated as gaps rather than as a plan, because the reconstruction records several of these as things it could not determine rather than things nobody has got to yet.
Places where two in-repo sources disagree. In each case code and git win, and the disagreement is recorded rather than quietly reconciled.
| Conflict | What the document says | What the code says |
|---|