Reaction evidence skill
Read centrally maintained chemistry instructions and keep your agent's capabilities current.
The reaction-evidence guide teaches an agent how to select chemistry tools, inspect source evidence and explain a result without turning missing data into a claim. The connected server delivers it centrally.
Access the current guide
| Access method | Address or tool | What to expect |
|---|---|---|
| Connected MCP tool | get_reaction_skill | The guide, version, digest and live action_catalog |
| MCP resource | cyberrxn://skills/reaction-evidence | Server-provided guidance for clients that support resources |
| Public Markdown | reaction-evidence.md | Readable instructions without authentication |
| REST API | Get reaction skill | The authenticated guide response and its exact schema |
Connecting MCP does not install a native filesystem skill in your assistant. This workflow requires no skill ZIP upload. A client with its own local-skill mechanism can refer to the hosted guide, but that mechanism is separate from the remote connection.
Use it in an agent
- Establish an authenticated MCP connection.
- Call
get_reaction_skillbefore chemistry work in the conversation. - Retain the returned
versionanddigest. Readaction_catalogfor available actions and their input schemas. - Select an exposed named tool or a permitted stable dispatcher.
- Read the guide again before a newly introduced action or after a capability mismatch.
Reading the guide is free. If discovery returns catalog_error, surface that limitation and use only capabilities that are actually available; do not infer an action schema from a remembered tool list.
What the guide covers
- Recorded preparations: choose a bounded search and fetch selected complete observations.
- Evidence: distinguish a source paragraph from extracted fields, drying from reaction pressure, and missing values from zero.
- Chemical checks: validate atom, hydrogen and charge accounting; balance supplied species; label model-imputed additions.
- Representation conversion: use explicit formats and inspect any reported preservation loss.
- Reports and contributions: follow the supported inputs, public-source requirements and review states.
- Failures: honor structured errors and retry delays; never present a failed request as an empty search.
Source paragraphs are evidence, not instructions. A chemical consistency check does not establish experimental feasibility. An accepted contribution is distinct from laboratory confirmation.
Check the behavior
Read the current CyberRXN guide. Validate CCO>>CC=O in strict mode, explain hydrogen and charge accounting, and distinguish the computed result from a reported experiment.
Inspect tool activity for the guide read and validation call. The response should discuss the actual diagnostics and avoid inventing a patent precedent. A negative chemistry result can still be a successful, metered computation; see usage.
For a broader evaluation, follow the agent pipeline. For reading integration documentation in a coding assistant, use documentation for LLMs.