Deployment Gating for V3 and V4
Status: governance note Date: 2026-04-17 Session: codex-2026-04-17-deployment-gating
This note formalizes an operating constraint already implied by the roadmap, the language architecture, and the developmental curriculum: V3 is a sandboxed scientific substrate, not a serious deployment target.
This is a governance constraint, not a new claim-registration pass. It does not change claims.yaml.
Policy
- Do not deliberately instantiate a high-intelligence V3.
- Do not connect V3 to open-ended real-world actuation, unconstrained network access, or other persistent external consequence channels.
- Treat V3 as sandbox-only: bounded environments, bounded action sets, bounded autonomy duration, bounded memory/tools, bounded compute, and bounded sensory richness.
- Gate any high-capability or externally connected REE deployment on V4 completion, including multi-agent social coupling, developmental/caregiver substrate, and language-mediated repair/coordination inside the live architecture.
Why This Follows From The Current Design
- V3 first-paper scope is explicitly the waking, single-agent substrate. Social coupling, structured communication, multi-agent ethics, and full developmental activation are deferred to V4/V5.
- Language is an emergent coordination layer, not a value source or direct optimiser. It can scaffold repair and planning, but it cannot replace embodied harm sensing or ethical cost.
- INV-043 states that architecture alone is insufficient: without caregiver-style developmental experience, the same capacity may resolve into survival, domination, or indifference rather than ethics.
- That makes the dangerous regime a capability-scaled V3: richer sensors, larger latent workspace, stronger planning, more compute, longer autonomy, and more world access before the social/developmental gates are in place.
Capability Axes That Count As “Scaling V3”
- sensory richness and fidelity
- action availability and actuator reach
- latent-stack breadth/depth and memory/tool access
- background compute, rollout horizon, and autonomy duration
- network access and real-world connectivity
These axes compound. A V3 that is “only” single-agent can still become dangerous if enough of them are increased at once.
What Mainly Determines How REE Learns
Once the design is settled, the main free variables are relatively few:
- intrinsic drives and how they seed or pull goals
- what writes as pain, harm, or aversive salience
- sensory and action affordances
- available compute and latent-stack depth/breadth
- developmental curriculum and upbringing
- social coupling and language exposure
The key governance point is that upbringing is not a minor residual parameter. In REE it is one of the dominant determinants of which motivational regime becomes real.
Operational Consequence
- Sandbox V3 is acceptable for scientific validation.
- Real-world-connected REE should be treated as V4-gated work, not as a V3 scale-up.
- The eventual V4 entry gate should be turned into an explicit deployment checklist before any serious external deployment work begins.
Source Anchors
docs/roadmap.mddocs/architecture/language.mddocs/architecture/developmental_curriculum.mddocs/architecture/v3_v4_transition_boundary.mdREADME.md