Digital Employee
Position, responsibility, workflow, runtime, governance, evaluation and managed digital work.
A Successful Replay Is Not a Standing Right to Execute
Executable memory can reduce the cost of repeated work, but historical success creates only a replay candidate. Every run must reprove compatibility, business authority, and external effect state.
A Correct Answer Does Not Mean the Work Is Complete
Continuous-execution evidence shows that semantic quality, critical actions, and timeliness can diverge. Completion for a digital employee should therefore be an evidence vector over obligations, time, effects, and acceptance—not a single success bit.
An Agent Is Not an Authority Boundary: What Cross-Substrate Failures Reveal
A new controlled study shows that identical planner-visible workspaces can require opposite safe publication decisions when actor-attempt authorization lives elsewhere. Its most important result is not that models reason perfectly once permissions are shown, but that execution-time authority checks can stop fixed unsafe intents without changing the model's plan.
Persisted Data Does Not Become Durable Memory
Controlled evidence shows that lifecycle routing can reduce temporary overwrite of permanent facts, but it does not decide who may promote an observation into durable state. Safe memory keeps observation, classification, write path, and read authority separate, with explicit promotion and revocation.
More Admission Checks Do Not Mean Safer Admission
A provider-valid API-key lane was rejected because the host lacked a CLI needed only by another lane. Correct admission is a proof bundle bound to the selected execution lane, its verifier, and evidence freshness; business authorization remains separate.
The State Rewound; the Effect Did Not
Aligned recovery can return model context and a controlled workspace to one checkpoint without undoing remote messages, records, or transactions. End-to-end recovery needs a separate external-effect ledger and resume gate.
Weekly 009 — Completion Must Be Typed: One Green Check Cannot Close the Whole System
Fourteen published Daily Research notes from September 7 through 13 repeatedly show that a local terminal state cannot be promoted into a global terminal state. This brief proposes Typed Closure: State, Authority, Effect, and Coverage need separate closure evidence, while Unknown must remain a valid durable state.
Complete Evidence Is Not Execution Authority
Complete evidence can make a planning problem solvable without granting permission. A governed digital employee must separate evidence, proposed intent, and current authority, then enforce an intent-bound check before any external effect.
A Second Model Is Not an Independent Control
A reviewer can be a separate model invocation and still share the actor's mistaken premise. Independent control requires separate scope facts, call-time effect authority, and an enforcement path that can deny the action.
A Plan Is Not a Commitment: What TROVE Changes About Agent Orchestration
TROVE treats a multi-step path as a provisional route, commits one top-level skill, then Retains, Inserts, Replaces, or prunes the remainder from boundary outcomes. The broader lesson is bounded: planning may look farther ahead than execution commits, while external Effect Authority still requires separate governance.
