| name | crypto_ragdb |
| description | Retrieve attack-family references, precondition checklists, and solver templates without letting retrieval replace local proof. |
| metadata | {"short-description":"Use local retrieval as a bounded hypothesis and validation aid"} |
Purpose
This skill provides bounded retrieval over a curated knowledge base. It should accelerate recognition of attack families and validation structure, but never substitute for local evidence.
Inputs
- current hypotheses or blocked validation questions
- specific unknowns such as attack preconditions, complexity heuristics, or known invariants
Required outputs
- notes or research artifacts under
artifacts/out/research/
- updates to research seed or validation plans when retrieval changed next actions
Operating rules
- Query narrowly.
- Translate retrieval results into concrete local checks.
- Record which external idea changed which artifact.
- Prefer checklists and discriminators over full copied solve paths.
- Drop retrieval outputs that cannot be grounded locally.
Edge cases
- retrieval finding a very similar but not identical primitive
- known attack names anchoring the manager too early
- solver templates that smuggle in assumptions not present in current evidence
Good retrieval targets
Use retrieval for:
- attack precondition checklists
- common discriminators for a family
- complexity estimates
- solver skeleton reminders
- failure-mode reminders for niche primitives
Avoid using retrieval as proof that your current challenge matches a famous writeup.
Validation gate
Pass only if retrieval outputs are reflected in bounded next actions or explicit rejections, not in hand-wavy confidence inflation.
