# ADR 0030 — RAVNOVES00 reference-source maturation before transfer

Date: 2026-07-27

Status: accepted

## Context

Mission Core has completed A0–A8 on the immutable RAVNOVES00 source and has a
source-catalog audit for the later E36 transfer gate. The immediate product
objective is not open-world autonomy or cross-route generalization. It is a
bounded near-industrial transport product operating at low speed in a known
location on known routes, with simple tasks such as movement from point A to
point B.

Attempting second-source transfer now would split effort before the
source-scoped perception architecture, evidence workflow and operator product
have reached a useful quality level. A new route and another camera view can be
recorded later, after the current reference path is stable enough to make that
transfer test meaningful.

## Decision

1. RAVNOVES00 is the only active physical reference source for the current
   product-maturation cycle.
2. E36 and TEST007 are deferred. The existing E36 catalog audit remains valid
   historical planning evidence, but transfer is no longer the current
   critical-path gate.
3. No new physical route or camera capture is required until RAVNOVES00 reaches
   the source-scoped product acceptance defined below.
4. All changes remain append-only immutable LAB results over the same
   RAVNOVES00 source. Accepted generations are never rewritten.
5. The next gate must define and freeze a reviewed development/validation split
   inside RAVNOVES00 before any new quality tuning.
6. The target is at least `90%` on each task-relevant validation dimension:
   object-versus-background decision correctness, camera↔geometry association
   correctness and current-versus-stale state correctness.
7. The `90%` target is source-scoped. It is not a claim of universal detector
   accuracy, cross-site generalization, safety certification or open-world
   autonomy.
8. Average quality cannot waive hard invariants: source accounting remains
   complete, missing evidence remains unknown, false free-space claims remain
   forbidden and degraded input must fail visibly.
9. The same reference cycle may extend reusable product controls, evidence
   viewers, LAB templates and comparison workflows under the existing Design
   Guideline and component architecture canons.
10. A second capture becomes the transfer gate only after the source-scoped
    release candidate is accepted and the operator explicitly authorizes new
    collection.

## Product envelope

The current product assumption is intentionally bounded:

- known or pre-mapped industrial location;
- declared routes and simple logistics missions;
- low-speed motion and conservative stopping on uncertainty;
- no adversarial environment, tactical autonomy or strategic replanning;
- no implication that a `90%` perception score alone authorizes actuation.

This narrower envelope reduces unnecessary complexity, but it does not remove
the requirement for explicit unknown/stale states, watchdogs, command expiry or
independent physical safety controls.

## Consequences

- The active critical path moves from E36 transfer to RAVNOVES00
  reference-source quality and product closure.
- The immediate next experiment is a source-scoped acceptance-contract LAB,
  not another data collection or another transfer audit.
- Future quality iterations may change algorithms, models and thresholds only
  through new immutable profiles and results evaluated against the frozen
  validation split.
- Product UI work is justified when it exposes a reusable task, method,
  evidence, result or comparison contract. Run-specific decorative or
  temporary interfaces remain forbidden.
- Generalization remains explicitly unproved until a later independently
  recorded source is admitted and replayed.