The mature agent system is not the one that acts most often, but the one that preserves evidence, authority, uncertainty, and recovery.
#ai
V0M3 removed the permanently animated assistant, made the document and review state visually primary, and let generation appear only where an author deliberately invited it.
Porting V0M3’s source and tool boundary to the newly released MCP validated its capability adapters while exposing the identity, schema, permission, content-trust, and recovery work a protocol could not own.
On MCP’s launch, V0M3 treated interoperable discovery and invocation as an adapter opportunity while keeping identity, authorization, content trust, review, and effects under product policy.
V0M3 exported ordinary Markdown and durable assets first, with optional provenance beside them, so completed writing did not depend on a provider, model session, or proprietary editor.
V0M3 treated rejection as an ordinary expression of authorship, even when a generated passage was accurate, polished, and objectively useful by common review criteria.
V0M3 separated evidence assessment from prose mutation so a reviewer could identify unsupported, contradicted, or overbroad claims without silently becoming their author.
V0M3 made accepted model changes ordinary document revisions, preserving exact operations and lineage so removal, restoration, and explanation remained possible later.
V0M3 used a stable proposal contract without pretending that model providers shared context rules, tool behavior, privacy terms, streaming semantics, or recovery guarantees.
V0M3 treated tool names and descriptions as metadata while server policy determined the capability, resource scope, arguments, and effect actually allowed.
V0M3 revealed complete proposal operations only after their targets, evidence, schema, and rendering policy validated.
V0M3 became dependable when generated text entered a typed workflow with evidence, revision history, validation, and human ownership.
V0M3 replaced generated blobs with typed insert, replace, delete, move, and comment operations whose claims and targets could be judged independently.
V0M3 bound every assistance request to exact document state so a proposal could not land on text that changed while generation was running.
V0M3 kept the canonical document under explicit human control by making model output a typed proposal that could be reviewed, changed, rejected, or accepted.
V0M3 began when a useful generated paragraph lost its sources, task, prompt, and review state the moment I pasted it into a document.
K81R stored immutable question, evidence, model, prompt, claims, and verification revisions so refreshed knowledge produced a new answer instead of rewriting history.
K81R split end-to-end answer quality into corpus, retrieval, evidence, generation, and verification stages so improvements could not conceal new failure modes.
K81R kept authorization, retrieval scope, tool capability, citation validation, and output rendering outside model instructions.
K81R stopped revealing prose character by character and buffered generated output into validated claim units whose evidence arrived with them.
K81R applied access scope to lexical and semantic candidates before ranking so restricted material never entered snippets, facets, caches, or model context.
K81R turned insufficient evidence into an investigative interface showing available sources, missing relationships, contradictions, and narrower questions.
K81R stopped rewarding answers merely for linking sources and began checking whether each cited passage actually entitled the adjacent claim.
K81R made selected passages a visible, editable, versioned artifact rather than an invisible context window assembled just before generation.
K81R's polished mistake turned retrieval, evidence lineage, answerability, and refusal into core product behavior.
K81R treated direct search and evidence inspection as complete outcomes instead of hiding every useful result behind generated prose.
The November 2022 research preview made fluent but incorrect output an interface and evaluation constraint, not a footnote to model quality.