Limager: the perception regime

Running the Reality Kernel to ask, of a scene, what is out there.

P.I.G.M.I.E. Filing 1 · patent pending

The perception regime

The Limager is the perception regime of the Reality Kernel. A perception regime is an operating mode whose objective is to estimate what is present in a scene. A scene is the physical configuration being observed.

The Limager uses the same emitter-detector loop as the other regimes. The emitter sends probes into the scene. The detector records the physical responses. From these responses, the Limager estimates structure, including 3D geometry, material properties, and semantic labels.

There is an older trade hiding in the name. A limner was a manuscript illuminator - to limn is to depict by illuminating, from the same root as illuminate. The Limager limns: it illuminates in order to depict.

Adaptive probing

The Limager does not capture a static image. It performs adaptive probing. Adaptive probing means that later probes depend on earlier observations.

This is active vision, the field built on the observation that a perceiver which controls its own sensor can make ill-posed problems well-posed: we do not just see, we look (Bajcsy). The Limager adds the one thing that field never had. The gaze is committed, so a probing run can be audited afterward.

Training the optimisation

The original experiments trained the Limager from labels. By default, and as first implemented, the Limager was bootstrapped by spraying projections across the whole scene and reconstructing the illuminated volume, in the manner of a simplified neural radiance field. That early reconstruction was coarser than a full neural radiance field, but it gave the optimisation a starting model of the scene to refine.

Concretely: the instrument sweeps patterned projections across an unlabelled scene and records each response, then reconstructs the illuminated volume into a coarse occupancy-and-colour model. That coarse model seeds the optimisation. Subsequent, adaptively chosen probes then refine it against an inferred target over the same scene, such as a depth map or a fuller neural radiance field.

From there, the Limager's labels and outputs were trained against one of two kinds of target. The target could be inferred, by applying another method to the same scene, such as a reconstructed 3D model, a neural radiance field, or a depth map. Or it could be known, such as ground-truth classifier labels.

Gen II (under exploration). Where enough Reality Kernels operate simultaneously, the optimisation can instead be trained by reinforcement learning: the agent controls probe selection, and its reward is downstream task performance measured against the co-observing kernels' recorded responses, for example the accuracy of identity classifiers scored on those responses. The filings also describe, as a non-limiting objective, selecting probes to maximise information gain about the scene, written I(S;Y), the mutual information between scene variables and the recorded output.

In every case, the label source and estimator in use are recorded in the protocol digest, so a perception run states how it was trained.

Across the reactor spectrum

The Limager works across the full range of reactor configurations. A reactor is the optional physical medium between the emitted probes and the detected responses. With a linear or identity reactor, the Limager is simply an active sensor. Forgery resistance is not the objective at that operating point.

The same convolution-bundle format, meters, and protocol logging still apply. A convolution bundle is the joint, time-ordered record of what was emitted and what was observed. This makes a perception run fully auditable and allows one instrument to support the perception, verification, and rendering regimes.

Status