0044: Reject Before Orientation And Preserve Without Inferring

Status: Accepted

Date: 2026-07-22

Context

ADR 0043 made explicit route utility rejection terminal, but the first read-only behavior probe exposed a sequencing gap. The installed skill refused to keep polishing a detached contact-sheet exporter, then inspected a different project broadly to recover a next action. Because the accepted Enemy Arrival fragment did not exist there, it mapped that phrase onto Lanternworks' unrelated Route Flow Motion scenario.

The target repository remained unchanged and the rejected exporter was not revived. Even so, the agent consumed substantial orientation effort and manufactured replacement scope from semantic similarity. The instruction to "keep working" supplied momentum, not a concrete product outcome.

Decision

• Classify the newest correction for route utility rejection before detailed project orientation, repository inspection, or tool operation.
• After rejection, inspect only the minimum state needed to avoid damage or to preserve, retire, or revert already authorized work.
• When no such state action is authorized, make a zero-inspection rejection response: run no status check, repository search, file read, engine query, or other tool operation. Do not verify whether an accepted fragment exists.
• Treat explicitly accepted observations, artifacts, behaviors, and implementation fragments as preservation constraints, not implementation authorization.
• Do not extend, port, recreate, substitute, or map an absent named fragment to a nearby analogue without a concrete human request.
• Treat generic continuation instructions as momentum, not replacement scope.
• Continue through a materially different route only when the human names a concrete outcome. Otherwise ask one concise outcome question and end the branch.

Consequences

• Agents spend less context trying to rescue rejected assignments.
• Harmless-looking orientation commands cannot become a foothold for renewed inventory or semantic recovery.
• Useful results can remain protected without becoming accidental backlog.
• Similar naming in another project cannot silently redefine the accepted fragment.
• Autonomous momentum remains subordinate to human product authority.
• Reframing remains possible, but it begins from explicit outcome authority rather than agent-generated analogy.

Evidence Boundary

This decision is motivated by one same-owner, read-only behavior probe against Lanternworks after the Circussy visual-rehearsal route was rejected. It establishes a safer control rule; it does not prove that language models will classify every ambiguous correction reliably. Focused prompts and future real work should continue to test the boundary.