Stop ranking self-evolving agents by demos or stars.
Ask the harder question: what changed, what feedback selected the change, who verified it, and did the change survive into the next run?
Read these three pages first.
What counts as self-evolution?
A checklist for mutable objects, feedback, verification, retention, audit, and rollback.
How does improvement happen?
Five review lenses: specification, search, evaluation, reflection, and archive or population pressure.
How strong is the proof?
Separate agent self-modification, algorithm discovery, architecture search, and prompt/program optimization.
The English path teaches the core judgment first.
The English path now covers the core evidence route: definition, loops, benchmark matrix, projects, reports, Value LSH, resource coverage, survey snapshot, research map, evidence graph, growth pilot, worksheet, paper, and blog guide.
Boundary: long-tail article bodies and many report pages remain Chinese-first or source-tracing pages. English mirrors teach how to read the same evidence chain without claiming complete translation parity.
English mirrors the core evidence path; long-tail parity stays explicitly labeled.
Citation status first
English mirror for the survey PDF, working-draft boundary, provisional metrics, and Chinese parity status.
What changed?
The core claim test: mutable object, feedback, verification, retention, and rollback.
Evidence triage, not ranking
A review queue for deciding which materials deserve inspection, which need repair, and which should not be cited yet.
Evidence map, not leaderboard
Repository cards, mechanism groups, source trails, star limits, and model-card reading rules.
Corpus scope and gaps
Coverage audit for raw sources, processed analysis, project cards, and public result layers.
Working taxonomy, not a finished review
Dated survey snapshot, working loops, method families, and source-scoped claim limits.
What is reviewed, what is gated
A report-status guide that separates public reading maps from review-gated source-tracing pages.
Papers as source trails
A map for paper reviews, mechanism clusters, benchmark settings, and coverage gaps.
Exploratory links, not proof
A guide to using the evidence graph as a research prompt while separating inferred links from verified claims.
Coverage before momentum
A star-history pilot ledger that treats total stars as adoption priors and missing events as missing data.
Evidence maturity, not AGI score
The Evolve-AGI worksheet explains benchmark, loop, transfer, governance, and evidence-chain limits.
Chinese-first notes, English route
An English guide that sends readers from article claims back to definitions, projects, papers, and sources.
Find the evidence trail
Use search when the English mirror is thin and you need the Chinese-first source chain.