205. What a runtime can do is a set of separately evidenced facts, and detection never authorises
Context
The 10-16 browser design and the 10-17 local-inference design both met a single optimistic 'supported' flag: a browser able to run a task is not allowed to; a model name does not prove a context window.
Decision
A capability of a runtime is declared, compiled, detected, authorised, verified and available as separate facts, each with provenance and time; routing reads the facts it needs and states which; inference never replaces authorisation.
Alternatives rejected
- One boolean per capability
- Capabilities guessed from a runtime or model name
Consequences
Generalises ADR 0174 (on an unlanded branch)'s evidence levels (in flight) and ADR 0170 (on an unlanded branch)'s untrusted browser; I3177 and I3199 read it.
Sources
Promoted on 2026-10-18 from the owner's ChatGPT project (workspace chatgpt-majordomus, issue I3100). The owner has not accepted it; it is proposed for the owner's decision.
chatgpt-majordomus:c275#4