Connectome Entity research · active research pauseFrozen empirical evidence · computational model
DeltaX EvaluateCONNECTOME ENTITY · RESEARCHSee evidence

Research proof 01 · Phase IV-C held-out benchmark

The governor should not need to own the intelligence.

The Connectome Entity uses an idealized whole-CNS connectome-derived rate-network substrate to generate candidate behavior. DeltaX receives those candidates externally and evaluates only what the substrate actually proposed.

See the architecture ↓Open public research ↗

This is a computational research model, not a living brain, a biological recreation, or evidence of high-level cognition.

Where DeltaX sits

Generation and governance stay separate.

Environment → sensory input → connectome-derived recurrent network → candidate behavior → DeltaX governor → selected or refused result → separately simulated consequence.

Environment and sensory inputConnectome-derived candidate behaviorDeltaX candidate gateSelected or refused resultSimulated consequence

The substrate supplies the available behaviors. DeltaX is constrained to that finite field. This separation is relevant to agent designs that change models or planners while keeping the decision boundary explicit; this one experiment does not validate robotics, other AI agents, or every substrate.

Frozen evidence · Phase IV-C

What the held-out study actually recorded

9,000closed-loop episodes across six controllers
100held-out seeds in 15 environments
199,500 / 199,500candidate-constrained selections matched an active proposal
0fallback triggers and selection errors in this benchmark
11 / 15environments solved with 100% seed parity in compatible environments
1,200failures in four incompatible environments attributed to substrate candidate absence

The 9,000 episodes include 4,500 DeltaX-controller episodes and 4,500 baseline episodes. Candidate matching and zero selection errors describe this benchmark, not a universal error rate or a production API SLA. The public package lists 19 primary evidence artifacts with SHA-256 digests.

The revealing failure

When right steering was required, the candidate was absent.

Four environments required rightward steering. In this rate-network instance, the substrate did not generate the needed turn_right candidate. DeltaX did not fabricate it or silently substitute another action. All 1,200 DeltaX failures in those environments were classified as SUBSTRATE_CANDIDATE_ABSENCE. The system did not solve those environments.

This identifies a capability boundary in the substrate while preserving the governor’s candidate boundary.

Research status and limits

Active research pause

The pause freezes the published evidence boundaries. Phase IV-D plasticity work is development-only; held-out plasticity confirmation has not been run. The public repository verifies the governance integration contract and committed traces but does not reproduce the private DeltaX runtime.

  • Idealized computational connectome-derived rate model; no biological equivalence.
  • No claim of a living brain or high-level cognition.
  • No claim that all 15 environments were solved.
  • Selection does not itself authorize execution.

Inspect the source