Research log & backlog

The techniques corpus grows through a research pipeline that opens pull requests. Two documents keep it accountable, and both are published as corpus pages rather than summarised here. They are the record, not a report about the record.

The log

→ Research log

Append-only, one row per run, newest first: date, topic, coverage verdict, number of sources, outcome. Written by the research skill at the end of every run, including the runs that produced nothing.

That last part is what makes it worth reading. A run that finds the corpus already covers a topic logs a FRESH row and stops without a pull request; a run that cannot find enough sources logs “insufficient sources” and stops. Both appear in the log next to the runs that authored documents, so the ratio is visible instead of implied.

The outcome column is where the corrections live. Research has, on the record, falsified a cached-prompt TTL figure the corpus had been repeating, corrected a stale delegation multiplier from 15× to 5×, and established by measurement that this project’s own lexical retrieval arm was inert. That became mech-crate-4jw, with a red test against it.

The backlog

→ Research backlog

The queue. Autonomous runs pop the top unchecked entry; anyone, human or agent, may append one. Entries carry a one-line reason and who added it:

- [ ] <topic> — <one-line why> (added YYYY-MM-DD by <who>)

Three things fill it. Humans adding topics they want covered. The techniques skill appending when a rag_context lookup comes back weak: the corpus noticing its own gap mid-task. And the research pipeline itself, which proposes topics back into the queue when the ladder reaches its tech-radar sweep.

Some entries are deliberately parked as watch items rather than research tasks: a wire protocol to re-check when its redesign ships, a Postgres major release to verify pgvector against before upgrading. Those sit unchecked on purpose.

Why they are corpus documents

Both files live in docs/development/ with the same frontmatter as everything else there, which means they are ingested, chunked and retrievable. An agent can ask what the project has already researched, and get an answer, before proposing to research it again. Publishing them here is the human-facing half of the same fact.