Sensory 2.0.0 — DPSH runtime

Released August 30, 2026 · Windows x64

Web artifact: Sensory-2.0.0-windows-x64.zip
SHA-256: 4eae97205789bbd8735adadaba19dae3188046e388207064482e583adba1186f

Sensory 2.0.0 introduces Cognia runtime 2: an event-driven DPSH backend for reproducible, falsification-oriented experiments with autonomous state neurons, nanosecond model time, explicit delays, stochastic firing and complete replay. The synchronous CUDA engine remains available as runtime 1.

Highlights

  • deterministic event queue and sparse delayed synaptic graph;
  • autonomous neurons with closed-form evolution and Ogata thinning;
  • analytical oscillators, phase modulation and phase-aware plasticity;
  • full .dpsh snapshots, named RNG streams, frozen replay and branching;
  • recorders, causal traces, trajectory analysis, ablations and matched controls;
  • STDP, learning gates, emergent workspace and distributed prediction errors;
  • declarative experiments and sweeps with versioned run registries;
  • multimodal shared-manifold measurements;
  • production DpshRuntime composition and Cognia-to-DPSH lowering;
  • realtime embodiment with population-coded actuators and a generic transport boundary;
  • CPU/CUDA bit-exact parity gate.

Validation

The CUDA validation gate passes 38 checks with no failures or blocked checks. The parity matrix covers 20 models × 5 seeds (100 runs), 3240 spikes and 1229 windows. Equality also survives snapshot/restore and active plasticity.

The rover embodiment regression additionally verifies:

  • I-9: attaching a motor sink leaves the event log bit-identical;
  • R0: live execution and offline trace replay have the same log hash;
  • moving one injection by 1 ns breaks that equality;
  • paced execution does not change model physics and reports overruns.

Runtime selection

cognia "2.0"
runtime "dpsh"

Runtime capabilities are validated during semantic analysis. A model requiring DPSH-only behavior is rejected by a backend that cannot execute it.

Compatibility

  • Runtime 1 remains supported and retains its synchronous execution model.
  • Runtime 1 .bin v4 and runtime 2 .dpsh v1 are separate formats.
  • Experiment registries reject cross-engine comparisons unless explicitly handled as a new series.
  • Models using competition now receive emergent inhibitory topology. Declare selector: argmax only for an intentional runtime-1 control condition.
  • Stochastic trajectories from older builds must not be compared as if they came from the same engine version.

Known limitations

  • GPU acceleration is scientifically equivalent but does not yet meet the original ≥10× whole-run target; queue and fan-out overhead remain optimization work under Z-9.
  • Realtime mode reports missed deadlines but cannot promise a deadline on arbitrary hardware.
  • The public package contains the desktop transport boundary; a physical robot still needs its device-specific relay and safety layer.
  • Sensory and DPSH remain experimental research software, not safety-critical control software and not a claim of machine consciousness.

Scale of the release

Compared with v1.0.0, this release contains 35 commits touching 181 files, with approximately 30,960 added and 459 removed lines. The thematic changelog is in changelog.md; implementation contracts live under Docs/runtime-v2.