Cognia 7.4.0 — virtualized perception and public percept reporting

Applies to: dpsh, transformer, orchestration

Prepared November 4, 2026

Summary

Cognia 7.4.0 promotes the reusable engineering results of the TRV and VQCA research programs into an experimental public language and runtime surface. It does not introduce a conscious engine or a preselected virtualized_path. Virtualization must emerge from ordinary delayed edges, refractoriness, adaptation, recurrence and current network state.

Language and DPSH

  • Relational transduction fields expose explicit spatial neighbourhoods without reducing a camera frame to a tiny input grid.
  • Named routes, loops and causal lineage support replay, cut, delay and scramble interventions over a shared DPSH substrate.
  • Clearance modulation and route-availability hysteresis provide a bounded local control mechanism.
  • Signed prediction-error gates, captured eligibility and capacity-weighted credit support specialization without assigning every center the same loss.
  • Temporal banks expose route lag, learned precision and calibrated confidence.
  • Workspace and coalition observers read emergent activity without declaring a privileged consciousness node.

Bounded transformer report loop

  • report_contract limits inputs, vocabulary, token budget and confidence.
  • The transformer sees only committed allowlisted workspace evidence; raw scenes, route IDs, branch IDs and ground truth are excluded.
  • proposal_return sends an evidence-bound proposal back as a new DPSH modality. The recurrent network decides endorse, reject or stale.
  • User-defined concept vocabularies remain independent of the fixed protocol verdicts.
  • cognia_system_run verifies artifacts and hashes before creating engines, runs DPSH and transformer components through the canonical scheduler and emits runtime, causal trace, output and report artifacts.

Examples 44 and 45 are public end-to-end fixtures. Example 45 accepts a live camera stream, forms a distributed DPSH percept, obtains a bounded transformer identification and returns it to a separate DPSH endorser.

Resident CUDA visual streaming

The DPSH CUDA backend can keep neural state, graph data and the event queue in VRAM across successive camera frames. Timed sensory inputs are appended on the device and only compact spike/readout and causal observation data cross back to the host. The implementation bounds resident logs and checkpoints, preserves pending events during materialization and poisons a failed resident session so that execution cannot continue from divergent host and device states.

Compatibility and status

The virtualized-perception features remain experimental. A functional percept, context-sensitive report or recurrent endorsement is not proof of subjective experience or qualia. Applications must establish their own causal controls and must not use this release for safety-critical decisions.

Existing 7.3.1 models and artifacts remain readable unless they depend on an internal experimental snapshot format explicitly replaced by the new resident or scheduled-report version. The Windows package targets the documented Blackwell/CUDA 13.2 configuration.

Release acceptance

Publication requires a clean Windows Release build, the complete Cognia, DPSH, transformer and orchestration validation suites, CPU/CUDA parity for the promoted resident path, Examples 44 and 45, package-content and documentation gates, and a smoke test from the final packaged archive.