ev-scout — breadth-first pre-curation context
Run before Phase 1 of /gaia-full-pipeline, including an already-submitted
gaia push batch (Strategy F). May also run standalone, before L4 mapping.
This is a lightweight read/search pass, not the later /ev-discovery Phase 0
nor a substitute for /ev-pipeline. Scouting is deliberately permissive;
downstream verification and /ev-adversarial-audit are deliberately skeptical.
Inputs
Use the declared skill URL(s), upstream SKILL.md or repository, intake issue,
and proposed generic concepts (if available). Keep multiple skills in a batch
distinct, especially for named-only evidence. Do not require an L4-approved
generic mapping: report tentative alternatives when topology is unresolved.
Search widely, without grading
Explore several search angles and independent source families:
- Generic capability: peer-reviewed papers, preprints, evaluations,
general benchmark methodologies, foundational approaches, rival systems,
and criticism or failure modes.
- Named implementation: source repo, specific
SKILL.md blobs,
releases, tests, CI/reproduction claims, independently authored reviews,
reported use, security findings, adoption or reproducible demonstrations.
- Comparison: alternative architectures and negative findings that could
falsify claimed benefits. Prefer disconfirming evidence when the first
results are uniformly promotional.
- Evaluation opportunities: candidate benchmark IDs, public methodology,
datasets, harnesses and missing reproducibility details. A vendor claim is
never a verified
benchmark-result.
Try broad and narrow queries; sample independent sources rather than repeatedly
searching the same host. Favor diversity of useful leads over an illusion of
completeness. Do not discard a relevant lead just because it may later fail.
Do not turn mere topic relevance into implementation-specific proof.
Use available search and repository-read tools. If Firecrawl is available,
follow /ev-discovery's authentication/preflight and quoting conventions.
If not, use other available read/search tools, or report search limitations.
Absence of search is not a curation blocker: return a clearly incomplete packet.
Output: scout packet
Produce a Markdown packet in the handoff or at
generated-output/ev-scout/<intake-or-slug>.md when a persistent file is
needed. Include:
- target skill(s), upstream URL(s), date, searched queries/source families,
tools used or unavailable, and mapping status (tentative or ratified);
- a table of leads, with URL, title/publisher, publication date if known,
possible canonical Gaia evidence type, generic/named/benchmark-catalog scope,
what it might substantiate, and what a verifier must still confirm;
- a separate counterevidence and competing approaches section;
- likely evaluation benchmarks as benchmark-source candidates only,
including proposed generic applicability and missing reproducibility fields;
- a short mapping-relevant synthesis that distinguishes architecture from
demonstrated performance and marks uncertainty explicitly;
- suggested verification priorities and handoff to /ev-discovery, /ev-pipeline,
and /gaia-ingest-batch after their normal human gates.
Use "candidate", "possible", or "unverified" until evidence is checked.
Never fabricate publication dates, citations, percentile values, peer-review
status, scores, or external adoption. Mark unavailable facts unknown.
Non-authority and handoff
No registry mutations, evidence-lake edits, GitHub approval labels, L4
ratification, star grades, Trust Magnitude, or automatic evidence ingestion.
This packet informs /gaia-draft-curate, /gaia-curate[-chain/-dynamic],
and the human L4 review; it does not override their workflows.
Later /ev-discovery can investigate selected Stage-2 leads; /ev-collection
normalizes typed candidate rows; /ev-benchmark-verification and
/ev-adversarial-audit challenge them; /ev-link-validation checks links;
human review and canonical Gaia CLI ingestion remain authoritative.
If a recent scout packet for this exact source exists, reuse and refresh it
rather than rescouting blindly. Record any new material or limitations.