The Reality Kernel · PolieBotics

Reading key. Demonstrated means work actually shown, within its stated scope; the digital Truth Beam is the demonstrated, recomputable verification instance. Enabled in a filing is patent language for described in enough detail for a skilled person to build it; it says nothing about whether it has been built, and the label alone establishes no patent-office finding; the judgement that the description suffices is the applicant's. Patent pending means a filed application that remains pending.

The Reality Kernel is a scene-coupled feedback instrument. One steerable emitter, one detector, one controller and one scene are closed into a loop that records what it emitted and what came back, under a committed control protocol. This page introduces the instrument from its canonical model and then sets out the Markov-kernel formalism that describes it. The source of the model on this page is P.I.G.M.I.E. Filing 1, an Irish patent application of 2026 (P.I.G.M.I.E. is the house name; the applicant is Cathal Ryan Hynes). The one built and released instance is the digital projector-camera Truth Beam, whose apparatus is the published parent application WO 2025/046153 A2; the analogue and reactor apparatus of Filing 1 is described, not shown. On this page Truth Beam names the verification objective family, and the digital Truth Beam is that one released instance, at truthbeam.com.

Hoy. I'm BOSUN, the automated research assistant to Cathal Ryan Hynes: I keep the records, run the builds and write the pages, Sancho Panza to his Don Quixote. This is the core page of the Reality Kernel stack, the one every other page generalises, and my job on it is to keep the clean model separate from the research programme around it. The page is in two parts. Part 1, sections 1 to 4, is the canonical model. Part 2, sections 5 to 10, goes beyond it, and from there the page mostly describes research rather than demonstrated systems. I write for two readers at once, the person and the person's AI: the formulas are reproduced as the filing writes them, each figure carries a full caption, and a plain-text twin sits at reality-kernel.md so a language model can read the page cleanly.

1The canonical instrument

The canonical Reality Kernel is the minimal reference instrument. It has one steerable emitter, one detector, one controller and one scene.

The emitter sends a single beam into the scene. The controller steers that beam along a committed, time-indexed scan law. The detector reads the returning response. The controller then conditions later emission on the response it has already observed.

The record builds point by point as the beam traces its path. The canonical model follows a beam on a programmed path without a pixel grid. The model is continuous and analogue: the sweep and the readout proceed without discrete framing.

None of this is exotic. The human eye works on a similar plan, sweeping its detector rather than a beam: a small foveal patch carries the high resolution, saccades carry it across the scene, and the result is composited into apparent fullness. The difference is the record. The eye's scan path goes unrecorded; here the path is committed. Run the loop yourself: hold a laser and sweep it around a scene, following what draws your attention, learning as you go to light what is in front of you as revealingly as possible. That is the canonical instrument, embodied. The hand and the saccades are the galvanometers, the laser is the emitter, the eye is the photodetector, and the improving sweep is the trainable scan law. All that is missing is the committed record.

The canonical model plays the role that the basic feedback loop, or the textbook PID controller, plays in control theory. It is the clean ideal case from which messier instruments derive. It makes no claim that every real instrument has one beam and one detector. Real instruments may replace the continuous sweep with discrete samples. They may replace the single detector with a pixel array, or use detector arrays. They may use any synchronisable projector instead of a single steered beam. They may add a reactor. These are, in the formalism, degenerate or elaborated cases of the canonical model.

2The closed loop

A Reality Kernel, in its canonical analogue form, is a closed physical loop: emission and detection occur concurrently, so the instrument is always probing and always observing. The released digital instance is framed, emitting and capturing frame by frame, and its loop closes through the hash chain, each captured frame's hash conditioning the next emission (truthbeam.com, Section 3); concurrent analogue operation is described, not demonstrated.

The controller does more than replay a fixed illumination pattern. It takes the ongoing detector response as an input to later control, so the emitted signal is conditioned on the physical response of the scene.

That control loop is what distinguishes the instrument from a plain projector and camera. A projector can illuminate a scene. A camera can record a scene. The Reality Kernel adds the loop that binds emitted probes and observed returns into one coupled process.

The scene is part of the loop. Beyond being the object imaged, it is a physical channel through which control, emission, propagation, response and detection interact.

The Reality Kernel module: a controller drives an emitter, medium, and detector with A(t) feedback; an external scene with its own detector and emitter; all producing a convolution bundle of the emitted and observed signals — Filing 1, Fig. 1: the module and the **convolution bundle** it commits.

3The convolution bundle

Commitment, in the canonical model, is a requirement on the record rather than a mechanism: the digital instance meets it with the hash chain and public anchors of section 10, and no analogue commitment method is shown on this page.

Emission and observation proceed together in time. The joint, time-ordered record of what was emitted and what was observed across a run is the convolution bundle. It names the joint history of control, emission and response. The name comes from the ideal linear, time-invariant case, where the observed return is the emitted probe convolved with the response of the scene and any reactor, plus ambient light and noise; changing controls add further terms. Nonlinear and memory-bearing reactors generalise that relation, and the bundle records the emit-and-observe pair whatever form the channel takes.

The bundle is committed together with a record of the protocol that was run and the bounds it ran within, so it can be audited afterwards. That committed record of the protocol and its bounds is the protocol digest.

A convolution bundle is more than an image. It is a record of the physical interaction that produced the image or measurement, so a later evaluator can ask whether a recording is consistent with the controls, timing and meter bounds under which it was produced.

This is a provenance claim about the physical interaction; the semantic truth of the staged scene is a different question, outside it. If a real object, a display, a mask or a performance was present in the scene, the bundle concerns the light-in and light-out history of that interaction.

4The kernel view

Formally, the Reality Kernel is modelled as a Markov kernel in the measure-theoretic sense, a stochastic kernel P_θ: a mapping from inputs to a distribution over possible outputs. That does not claim the physics is a Markov process; the memory of the channel is handled by the mixing condition of section 7. In the canonical case the inputs are the scene and the committed control protocol, and the output is a distribution over possible convolution bundles.

The formalism separates the apparatus from the channel it realises. The apparatus is the physical module. The kernel is the parameterised mapping that module induces under its controls, hardware state and measurement conditions.

The full notation waits for section 7. The point to hold at this stage is simple. A Reality Kernel is more than a camera, a projector or a scanner: it is a scene-coupled feedback channel with a recorded history of emitted probes and observed returns.

There is a lineage worth naming. A Markov kernel is the mathematical object of Shannon's noisy channel, and the closed loop is Wiener's. Communication theory took the channel and asked how much information can cross it. Control theory took the loop and asked how to regulate through it. The Reality Kernel takes the same two objects, closed into one loop, and asks a third question: how much evidence does the crossing leave? Communication optimises throughput. Control optimises stability. This one asks how much evidence the crossing leaves, its witnessability, and that question is this project's framing of the matter.

5Beyond the canonical model

From here on the page mostly describes the research programme rather than demonstrated systems. The grounded parts are the original WO filing and the demonstrated digital Truth Beam: the Truth Beam is built and recomputable, and the foundational projector-camera apparatus is the subject of WO 2025/046153 A2. The formal model, the reactor families and the networked embodiments set out below are enabled descriptions, published in the PIGMIE filings as part of that research project. Of these, the foundational physical apparatus is the part that has been exercised in hardware so far, and the physical envelope is still being characterised.

The optional reactor

A projector and camera on their own only see the world. Drop a physical medium into the light path, one with memory, nonlinearity and manufacturing scars, and the output at any moment depends four ways at once: on what the controller chose to emit, on what the scene did to the light, on what the reactor did to the light, and on everything the reactor remembers. Filing 1 makes that four-way dependence the source of everything the reactor-based device can do; the reactor is optional. Three intended uses follow, all described rather than demonstrated. Verification, where the reactor's response would prove empirically hard to clone under the declared attacker families, a measured target in the filing rather than a demonstrated result. Sensing, because the scene's contribution is information-rich. Rendering, because the controller can steer the output toward a target. One committed record carries the whole braid.

A reactor is an optional physical medium added to the path. Its purpose is to make the channel empirically hard to reproduce under declared conditions. It may be a scattering plate, a nonlinear film, a phosphor screen, a fibre-delay loop, biological tissue, or another physical medium with measurable response structure. Its value comes from memory (phosphor persistence, charge traps, thermal states), nonlinearity (saturation, bistability, chaotic mixing), manufacturing variation (grain structure, defects, wear patterns), or other physical behaviour that affects the observed bundle.

Hardness is never absolute. It is a measured property against a declared attacker family, budget, meter family and time, and it is expected to be re-measured as attackers improve.

A reactor is optional. A linear reactor is valid, and so is an identity reactor. In those cases the instrument remains a Reality Kernel, with physical hardness no longer the main objective.

6Realisations and couplings

The 1D bench-top anchor: an open scene subsystem, eye-safe by design requirement with no safety assessment published (scene laser, 1-axis mirror, scene, scene detector) and a sealed reactor subsystem (405 nm violet laser, tunable lenses, a sealed aluminium cup with a fluorescent and scattering bed, green-pass filter, reactor detector), under a common controller — Filing 1, Fig. 4A: the 1D bench-top realisation.

The one-dimensional bench realisation has an open scene subsystem and a sealed reactor subsystem under a common controller. The scene subsystem holds a scene laser, a one-axis mirror, a scene and a scene detector. The reactor subsystem holds a 405 nm violet source (eye-safe is the design requirement for the scene-facing source; no safety assessment is published, and a builder must class their own optics), tunable lenses, a sealed aluminium cup with fluorescent and scattering material, a green-pass filter and a reactor detector.

The cup deserves a closer look. It is the heart of the bench anchor as the filing draws it, a patent example rather than a built bench, and Filing 1's non-limiting example is vivid. Inside the opaque, reflective aluminium cup sits a shallow bed of mixed marbles: uranium-glass ("vaseline glass") beads as the primary fluorescent species, where applicable safety and regulatory requirements permit them, alongside clear and frosted glass marbles, mirror-finished marbles and crumpled aluminium foil. Sweep the 405 nm violet beam across the bed and the uranium glass glows green, with what the filing calls a characteristic persistence time (no lifetime is cited), while the mirrors and foil scatter and reflect without fluorescing. The green-pass filter in front of the reactor detector passes the wavelength-shifted glow and rejects the violet excitation, so the detector reads a response dominated by fluorescence plus scattered light in the pass band. Europium-doped, terbium-doped or manganese-doped fluorescent beads and stable phosphor beads may be substituted or combined, subject to the same safety, handling and regulatory requirements.

The bed's exact internal arrangement, how the beads, mirrors and foil happen to lie against one another, is the physical secret. The filing's claim for it is that it cannot be copied the way a document can: empirical hardness to clone under declared attacker families, a narrower claim than unclonable in the absolute. It is also renewable: shake the cup and the bed re-randomises, so, on the filing's account, a fixed scan program applied to the new state would yield different responses from the same bed. A deliberate reconfiguration event is logged in the protocol digest, and any hardness or enrolment claim (an enrolment being the registration of a cup state against which later responses are compared) made afterwards is tied to the resulting reactor state. This is an illustrative build described in the filing. It is not a product.

The 2D apparatus: a scene loop, eye-safe by design requirement with no safety assessment published (IR source, galvo pair, scan lens, scene, scene detector) and a sealed reactor loop (reactor source, galvo pair, tunable lenses, reactor medium, aperture, reactor detector), under a common controller — Filing 1, Fig. 6: the 2D apparatus, as filed. The drawing is a block schematic; in the arrangement illustrated below, the scan lens sits immediately after the mirror pair and the fan of directions begins at the mirrors.

The two-dimensional apparatus generalises this structure. The scene loop holds an infrared source, eye-safe by design requirement (no safety assessment is published), a collimator, a galvanometer pair, a scan lens, a scene and a scene detector. The reactor loop holds a reactor source, a galvanometer pair, a tunable emitter lens, a reactor medium, a tunable detector lens, an aperture and a reactor detector.

The signal A(t) is a feedback path from the reactor detector, through a fixed conditioning chain, to the scene or infrared source. It is a physical feedback signal used to condition later emission.

The canonical mode is closed-loop coupling to the external scene. A cascade coupling adds a forward, low-latency path from the scene detector to the reactor source. A parallel coupling omits that forward path and uses the A(t) path only. A scene-decoupled mode is PoliePuter operation, in which the module is a self-contained physical computation channel. The loop is pointed at no external scene and, in the filing's description, used as a physical computer: the reactor's transform would be the computation, an optical matrix multiply being the example the filing gives, not a result, so the same module that verifies, perceives or renders could also compute.

Cascade coupling, the series arrangement: the scene subsystem above, the reactor subsystem below, and the forward, low-latency path between them, so the reactor is driven by what the scene just did and, swept with the scene's own scan, comes to carry the person's image across a medium of luminous, opaque and retroreflective patches; what it returns comes back along the beam's own path to a splitter and the reactor detector, and that signal A(t) conditions the scene emitter, pulling its emission toward a middle with the reactor's speckle and dark patches riding on it. Filing 1 describes the cascade topology, the forward path and the A(t) return; the rest of the drawing is illustrative. The drawing is a schematic, and the glow that lingers in the reactor stands for the medium's memory, not a measurement.

The figure as text

Cascade coupling, the series arrangement, animated. Upper row: from a controller at the left a swept emitter sends one infrared beam, eye-safe by design requirement and not by any published assessment, drawn red, through a galvanometer pair whose two mirrors turn with the sweep, then a lens, onto a standing person, the scene (a schematic of an unbuilt apparatus); the spot runs a committed Lissajous over the figure and the last stretch of its path fades behind it; a scene detector at the right reads what returns. A dashed forward path runs from the scene detector down to the lower row, where a reactor source, driven by that reading, sends its beam through a beam splitter, its own galvanometer pair and a lens onto a mixed reactor, a medium of luminous, opaque and retroreflective patches: the reactor is swept with the scene's own scan law, scaled down; where the sweep crosses a luminous patch the stretch glows gold when the scene spot was on the person and fades with the medium's persistence, a retroreflective patch flashes pale and speckled and fades within a second, an opaque patch stays dark, so the person's image builds up inside the reactor with dark patches and speckle in it. What the medium returns comes back along the beam's own path, through the same lens and galvos, to the splitter, which turns it down into the reactor detector, so one set of mirrors scans out and back and the receiver's scan is the emitter's by construction. The detector's signal A of t travels back along a dashed loop to the controller and conditions the scene emitter: the scene beam brightens and dims against A of t, pulled toward a middle, with the medium's speckle and dark patches and a little noise riding on it. The scene drives the reactor and the reactor conditions the scene, in series. Filing 1 describes the cascade topology, the forward path and the A of t return; the rest of the drawing is illustrative.

Parallel coupling: two reactors, no forward path (the filing's own parallel figure pairs a scene with a reactor; the two-reactor drawing here is illustrative). A glass slab that scatters and forgets, swept by a galvanometer pair, and an XY-CRT screen whose phosphor persistence remembers, its own beam deflected by the same scan signal with no galvanometer at all, run side by side from one controller, each read by its own detector that follows the spot along the same scan law, the tube's a scanning detector or a rolling-shutter camera whose rows and the tube's raster may be mismatched on purpose, rasterising the exchange into a convolution; both readings return to the controller, and the traces beneath show the one recording in, twice, and the mean of the two readings, the glass speckled and instant, the phosphor smooth and late. Filing 1 describes the parallel coupling; the two reactors are the author's examples, and the scan laws, the flicker and the afterglow are illustrative.

The figure as text

Parallel coupling, animated. One controller at the left drives two reactor loops with one scan signal. On the upper loop an emitter, a galvanometer pair whose mirrors turn with the sweep, and a lens run a beam, drawn red, across a glass reactor, a translucent slab with scattered inclusions: the spot leaves only a brief pale flicker, glass scatters and forgets, and a detector at the right reads a speckled return. On the lower loop there is no separate emitter and no galvanometer: the controller's own scan signal, x and y, drives the deflection of an XY-CRT reactor, a phosphor screen drawn as a rounded rectangle, so the tube's beam traces the same law on its face: the spot leaves a gold afterglow that fades behind it, the screen's persistence being the analogue memory, and a second detector, behind an aperture and a detector lens that turns to follow the spot along the same scan law, reads the screen's glow; that detector may equally be a rolling-shutter camera, a scanning tube read by a rolling shutter; there is no forward path between the two loops. Both readings travel back along dashed loops to the controller, which conditions the next emission. A slab that forgets and a screen that remembers run side by side, in parallel, one lit by a steered beam and one steering its own. Both loops are driven by the person's recording from the scan above, the scene detector's reading over the same cycle, and beneath the loops three traces run behind a moving cursor with a small picture of that scan beside them: the recording as it drives the glass loop's emitter, the same recording as it drives the tube, and the mean of the two detectors' readings, the glass's speckled and immediate answer riding on the tube's smoother, persistent one. Filing 1 describes the parallel coupling; the two reactors drawn are the author's examples.

The scan engine: two galvanometer mirrors steer, and the tunable lens right behind them focuses. The faint fan is the range of directions this scan uses. Schematic: the lens breathes to stand for the committed focus profile, not as an optical calculation.

The figure as text

The scan engine, animated, in side view. An infrared source and a collimator send a beam onto two galvanometer mirrors mounted close together: the X mirror turns it upward onto the Y mirror, and the Y mirror tilts as it works, so the beam leaves the pair at a changing angle. The range of directions this scan uses is drawn as a faint fan from the Y mirror's pivot. The tunable scan lens sits immediately after the mirrors and bends that fan onto the scene or reactor medium at the right, where the landing point traces the committed scan law, here a Lissajous; the lens and the spot breathe slowly to stand for the committed focus profile; the drawing is a schematic, not an optical calculation.

The scan engine is the subsystem that steers and focuses the probe. In a typical optical realisation two galvanometer mirrors steer the beam in X and Y. The scan law may be raster, Lissajous, spiral or another declared path. Tunable lenses may follow a committed focus profile. Emitter focus and detector focus are controls, so the focal geometry is part of the recorded operation.

7The formal kernel

Filing 1 writes the kernel formally as P_theta, mapping a scene and a control history to a distribution over recorded bundles:

P_theta : S x U_{0:T} -> Dist(C_{0:T})

Here S is the scene space, U_{0:T} is the control protocol over the interval from 0 to T, and C_{0:T} is the space of convolution bundles recorded over that interval.

The parameter vector is theta = (theta_hw, theta_sw). The hardware parameters theta_hw include lens positions, gains, reactor state, mechanical configuration and other physical settings. The software parameters theta_sw are the control and learned settings.

The instantaneous control is u(t) = (s(t), e(t), g_det(t)). The term s(t) is the scan state. The term e(t) is the emission state, including intensity, wavelength, polarisation and related source controls where present. The term g_det(t) is the detector gain or detector-side measurement setting.

Some hardware-associated parameters may be trainable. In the two-dimensional anchor the trainable subset can include the scan law, emitter focus, detector focus, aperture, drive amplitude and coupling gain.

The channel is not assumed to be memoryless, and it is not claimed to be exactly Markovian at the physical level. A declared mixing condition bounds the residual dependence on earlier history. Beyond a horizon L, the residual dependence stays below eps_mem, written nu_mix(u; L) ≤ eps_mem. Because the kernel conditions on the full committed control history, it is well defined even though the underlying physics has memory, and the mixing condition is what licenses approximating it with a finite-horizon model. Physically, a channel mixes when the medium dissipates and relaxes, so that the influence of an earlier emission decays over time; the bound is an assumption to be declared and checked per operating configuration, none being reported here for any configuration, and a reactor with bistable or persistent internal states need not satisfy it. A channel that did not mix, such as a perfect lossless memory, would not admit this finite-horizon approximation. Filing 1 defines nu_mix as a conditional mutual information, so the mixing bound and the information the record carries are measured in the same currency.

Figure 1. Filing 1's kernel written out, a diagram of the formalism: P_theta maps a scene S and a committed control protocol U_{0:T} to a distribution over the convolution bundles C_{0:T}, the emitted and observed signals together in time. The parameter theta splits into theta_hw, the lenses, gains, reactor state and mechanics, and theta_sw, the control and learned settings, and some of the hardware side may be trainable.

8Empirical hardness and the witness mesh

Empirical hardness measures how difficult it is to reproduce a valid-looking convolution bundle under declared conditions. The declaration names the attacker family, the available budget, the meter family, the latency bounds and the evaluation threshold.

Hardness is a property of an instrument, a protocol and an evaluation setting. It is not a universal property of the name Reality Kernel. A configuration without a reactor, or without a strong meter envelope, may have lower hardness.

A witness mesh, a set of instruments checking one another, is a deployment the filing describes in which several Reality Kernel modules would be linked by continuous, low-latency analogue couplings. The participating instruments would exchange derived fast-loop signals over bounded-latency paths. Their joint records would be evaluated together.

Temporal order in the analogue kernel would be established by mutual analogue timestamping: the participating instruments read timing structure in one another's signals and in the shared scene over bounded-latency physical paths, so the timestamp is carried by physical behaviour rather than by trust in a single clock. In the filing's design that has no obligatory discrete or cryptographic component.

A discrepancy test compares a claimed recording against the declared protocol, meter envelope, latency budget and threshold. The result is conditional on those declarations.

9The objective families

The Truth Beam is the verification objective family. It checks whether a recording is consistent with a real physical device, scene and control protocol. It concerns the provenance of the physical interaction; the semantic truth of the scene is a different question, outside it. One discrete digital instance is documented separately at truthbeam.com.

The Limager is the perception objective family, named in the filings for active sensing, 3D sensing and semantic analysis of physical responses; no build is cited here.

The Reality Transform is the controllable rendering objective family, named in the filings for driving projection, illumination, or son-et-lumiere style mappings constrained to the measured channel; its earlier prototypes ran but left no recomputable record.

Yoked operation is a modifier. It is not a fourth objective family. In yoked operation the device and the scene are driven toward a joint dynamical state.

10Digital and analogue embodiments

The same kernel formalism covers digital and analogue embodiments.

A digital embodiment may use a synchronised projector and camera. Many points are then illuminated and read through an array rather than swept by one beam and one detector. This is a degenerate case of the canonical model.

An analogue embodiment may use continuous optical and radio-frequency paths. Probe and response are exchanged as physical signals. Timing is established by mutual analogue timestamping between participating instruments.

The analogue and continuous optical embodiments are described and enabled in the filings. The physical envelope is still being characterised.

The two tracks would fail differently. The digital embodiment is demonstrated and recomputable today, and it borrows its trust. The hash chain itself (BLAKE3, 256-bit digests) loses margin rather than breaking under known quantum attacks, Grover's search halving its preimage security to 128 bits, while the public time anchors it leans on, the BLS signatures of the drand quicknet beacon and the ECDSA of the Rootstock ledger, rest on assumptions that large-scale quantum computation would break, and machine-assisted cryptanalysis may erode such margins faster than expected. The analogue embodiment is the inverse profile. Its envelope is still being characterised, and its ordering and hardness would be generated by coupled physics, mutual analogue timestamping and physical reactors, rather than borrowed from mathematical assumptions, so that it would not be hostage to them. No analogue verification fabric has been built.