Finding
Repo-To-Skill (arXiv:2609.02749, submitted 2026-09-02) argues that repositories contain operational knowledge that agents repeatedly rediscover. Its AREX Skill Library contains 5,000+ verified skills distilled from 1,000 ML repositories across 20 areas and 178 capability families.
With a GPT-5.5 backbone, research harness, and downstream execution budget held fixed, the originating team reports gains of 134.3% on MLE-bench, 34.4% on PaperBench, 9.2% on FrontierCS, and 14.0% on PassNet versus the same agent without skills. Treat these as originating-team results. The arXiv page does not provide enough evidence yet to assume the full library, distillation pipeline, licenses, and benchmark setup are independently reproducible.
RuV opportunity
Do not create another skill store. Use Dream Machine to compile verified operational skills into the existing RuV substrate:
- repository commit and license provenance
- source file and line evidence
- capability family and task signatures
- executable verification command where available
- supported environment and dependency versions
- known failure modes
- security classification
- cost estimate
authority: none- RVF packaging for signed portable distribution
- RuVector indexing and retrieval
- Core Memory promotion and revocation state
First benchmark
Use a fixed set of 20 to 50 RuV repositories and held-out tasks derived from real integration, debugging, benchmarking, and deployment work.
Compare under identical model, seeds, harness, tool access, and total context budget:
- no repository skill
- README and documentation retrieval
- automatically distilled skill without verification
- automatically distilled plus executable verification and provenance
Keep the test tasks hidden from distillation.
Falsification
This direction should be rejected or narrowed if ordinary repository retrieval performs within variance, if the skill is mostly copied documentation, if verification costs erase downstream savings, if skills become stale faster than they are reused, or if licensing prevents redistribution.
Security
A skill is operational evidence, never execution authority. It cannot grant capabilities, relax RVM policy, change evaluator state, or self-promote. Skills that include shell commands, network operations, secrets, destructive operations, or deployment instructions require explicit risk metadata and remain subject to RVM at execution.
Track source license and commit for every skill. A skill derived from incompatible source terms must not be redistributed as a clean RuV artifact.
Metrics
Task success, time to first valid action, total input/output tokens, tool calls, wall time, verifier cost, stale-skill failures, security interventions, retrieval precision, source coverage, skill bytes, and reuse count.
Acceptance
Promote the primitive only if verified skills improve held-out task success by at least 5 absolute points over the stronger of no-skill and documentation retrieval, or match success while reducing total model tokens by at least 25%. Require zero capability expansion, complete source and license provenance, and no protected security regression.
Coordinate with MetaHarness independent evaluation and the Core Memory federation record. No autonomous merge.
Finding
Repo-To-Skill (arXiv:2609.02749, submitted 2026-09-02) argues that repositories contain operational knowledge that agents repeatedly rediscover. Its AREX Skill Library contains 5,000+ verified skills distilled from 1,000 ML repositories across 20 areas and 178 capability families.
With a GPT-5.5 backbone, research harness, and downstream execution budget held fixed, the originating team reports gains of 134.3% on MLE-bench, 34.4% on PaperBench, 9.2% on FrontierCS, and 14.0% on PassNet versus the same agent without skills. Treat these as originating-team results. The arXiv page does not provide enough evidence yet to assume the full library, distillation pipeline, licenses, and benchmark setup are independently reproducible.
RuV opportunity
Do not create another skill store. Use Dream Machine to compile verified operational skills into the existing RuV substrate:
authority: noneFirst benchmark
Use a fixed set of 20 to 50 RuV repositories and held-out tasks derived from real integration, debugging, benchmarking, and deployment work.
Compare under identical model, seeds, harness, tool access, and total context budget:
Keep the test tasks hidden from distillation.
Falsification
This direction should be rejected or narrowed if ordinary repository retrieval performs within variance, if the skill is mostly copied documentation, if verification costs erase downstream savings, if skills become stale faster than they are reused, or if licensing prevents redistribution.
Security
A skill is operational evidence, never execution authority. It cannot grant capabilities, relax RVM policy, change evaluator state, or self-promote. Skills that include shell commands, network operations, secrets, destructive operations, or deployment instructions require explicit risk metadata and remain subject to RVM at execution.
Track source license and commit for every skill. A skill derived from incompatible source terms must not be redistributed as a clean RuV artifact.
Metrics
Task success, time to first valid action, total input/output tokens, tool calls, wall time, verifier cost, stale-skill failures, security interventions, retrieval precision, source coverage, skill bytes, and reuse count.
Acceptance
Promote the primitive only if verified skills improve held-out task success by at least 5 absolute points over the stronger of no-skill and documentation retrieval, or match success while reducing total model tokens by at least 25%. Require zero capability expansion, complete source and license provenance, and no protected security regression.
Coordinate with MetaHarness independent evaluation and the Core Memory federation record. No autonomous merge.