| name | autoresearch-db |
| description | Use when building the evidence base for a chosen autoresearch topic — maps the insight areas needed to propose new ideas, downloads references to cover them via download-ref, distills each area into research/INSIGHTS.md for user selection, builds a structured domain database and pinned reference implementations, and catalogs all algorithms/software in research/CATALOG.md. Owns the survey gate. Stage 2 of the autoresearch pipeline. |
Autoresearch DB
Input: a topic with metrics in topics.md. Output: a knowledge base, an
insight file, a domain database, pinned reference code, a catalog — and the
survey gate passed. The point is not "papers exist on disk" but "the agent
holds the insights needed to propose new ideas".
Procedure
- Map insight areas. From the topic and its metrics, list the insight
areas an idea-proposer needs: algorithmic techniques, proof/analysis
methods, data structures, benchmark practices for this problem class.
Show the list to the user before downloading anything.
- Download for coverage. Acquire references with the
download-ref
skill into <project>/.knowledge/. Coverage check: every insight area is
covered by ≥1 downloaded reference; areas with no coverage trigger
further search/downloads. SOTA-results-only coverage is insufficient.
- Distill. For each insight area, write one entry in
research/INSIGHTS.md following
skills/autoresearch-db/references/insights-template.md (Technique /
Applies when / Limits / Sources). Distill from the rendered papers in
.knowledge/, not from memory.
- User selects. Present the distilled areas via
AskUserQuestion
(multi-select): which are needed for idea generation on this topic?
Selected entries go under ## Selected, the rest under ## Shelved.
The run loop draws only on Selected.
- Domain database. Build the structured dataset the topic needs (e.g.
QEC codes as JSON) under
research/database/, with a README.md
documenting schema and provenance of every record.
- Pin reference implementations. For each key algorithm with public
code: record URL + commit hash, build it, and run its own smoke
test/example. Implement small reference versions only where no code
exists and the algorithm is load-bearing.
- Catalog. Write
research/CATALOG.md: every algorithm/software
relevant to the topic, one row each — name, source (paper/repo),
status reproduced (we ran it) / pinned (source-locked, not yet run) /
paper-only (no code exists), and notes.
Survey gate (owned by this skill)
Flip survey_gate: passed <date> and set stage: validator in
research/STATE.md only when all of:
research/CATALOG.md exists and every key algorithm from the insight map
appears in it;.knowledge/INDEX.md lists every reference cited in INSIGHTS.md and
CATALOG.md;research/INSIGHTS.md has a non-empty ## Selected section chosen by the
user;- every
reproduced-status entry actually ran (its output is recorded in
CATALOG.md notes).
