The document, mapped · 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.

Filing 1, the Physical Markov-Channel Apparatus, was filed at the Intellectual Property Office of Ireland in 2026 and is pending; the applicant publishes the document but not its filing date or Office number, so that status is applicant-reported here, as on the filings page. The published PDF runs to 448 pages (the filing as published), from some 25,700 lines of LaTeX source, with SHA-256 digests in CITING. The other pages on this site summarise its spine. This atlas indexes the named concepts of the filing, including the many the other pages leave out, one line each, so that a concept can be found without reading the whole document; the aim is that nothing significant hides in the territory, and completeness is not claimed. Every entry below is described in the filing, which is disclosure and nothing more. An entry is a described embodiment unless its own words mark it as an interpretation, a conjecture or a proposal, which is the three-way distinction the filing itself draws; most of the interpretations and conjectures are gathered in section 12. Two senses of one word recur: the convolution bundle is the committed record of a run, and the attractor bundle is the geometric object of section 12. None of this is a built system, and the only demonstrated, recomputable instance remains the digital Truth Beam.

This atlas is clerical work by BOSUN, Cathal Ryan Hynes's automated research assistant, and the clerk's one rule is that every concept keeps its name and its line: a summary that quietly drops an idea is worse than none. It is written for two readers at once, the person and the person's AI: each concept is a term with its definition, a plain-text twin sits at filing-atlas.md, and the filing itself, the published PDF, is authoritative; its index page is a navigation summary and not the filing.

Figure 1. The atlas as a map: the twelve groups of this page in a ring around Filing 1, each circle sized by its number of entries, 88 in all. Each group is a section below, in the same order; the map states nothing the page does not.

1The record and its opening

What the instrument writes down, and how a verifier inspects part of the record without being shown the whole.

Disclosure atom and evidence window: the smallest unit of the record that can be independently committed and selectively opened, and the group of atoms assembled for a specific verification, audit or cross-attestation.
Run seed and opening seed: two independently generated seeds. One fixes the emission schedule before the run; the other, supplied by a verifier or beacon only after commitment, selects which atoms are disclosed, so the device is not expected to predict, under the declared randomness and timing assumptions, which parts of its record will be inspected.
Two-seed selective opening: the full protocol: Merkle-commit the complete sample-level record at window close, then open randomised atoms under the post-commitment seed. Verification without streaming.
Verifier coverage tuple: five declared parameters, the opening fraction, edge-check coverage, decimation, amplitude and phase tolerances, that index the inspection policy of a verifier over a joint record.
Detection-probability relation: a first-order bound tying that coverage tuple to the probability of catching a forgery that corrupts a given fraction of committed edge-atoms; the bound assumes independent sampling of the coverage and residuals above tolerance.
Verifier timing fidelity: the declared resolution at which inter-module arrival times are checked; the latency budget's contribution to hardness scales against it.
Logged derivation of session parameters: every session's parameters derive from precommitted specifications, so even the setup is auditable.

2The mesh, formally

Several instruments witnessing together: the mesh as a graph, with the joint record verified across it.

Coupling graph: the mesh as a directed graph of modules, edges, declared delays and weights; the joint record is verified across it.
Witnessing requirement: each module's response must be bound to its own observation of physical reality. The filing bounds each module's witnessing contribution from above by the entropy of its scene region left unobserved under the declared adversary surveillance budget, estimated by a committed five-step procedure: enumerate the adversary's observation channels, define the meter variables the verifier checks, fit the best declared reconstruction model under that budget, measure its conditional prediction error against the verifier's tolerances, and commit the estimate with a confidence interval.
Coupling modalities: inter-module coupling by amplitude, which is canonical, or by phase, frequency, polarisation, wavelength or spatial mode, each with enabled species.
Versioned parameter families and rotation: the trained coupling network is frozen and digest-committed per deployment version, periodically retrained against current adversary classes and re-frozen. Key rotation, in analogue.
Cross-version transfer-attack evaluation: a rotated coupling's hardness is scored against an adversary that carries forward all it learnt from prior versions, including reactor microstructure, geometry, scene priors, latency calibrations and surrogate models, so that rotation's value would be scored rather than assumed.
Topology reconfiguration: rearrangement of the mesh graph itself at declared rotation cycles, re-weighted at parameter level, physically re-routed, or circuit-switched through optical switch elements, as a defence and maintenance move.
Optical-delay logical clocks: partial orders of events from physical propagation delays between instruments, under propagation and hardware-latency bounds, rather than from a trusted external time source; each instrument still needs its own detector and timing hardware.
Enrolment, revocation and quarantine: the network trust lifecycle, with an exponential-weights web-of-trust updating each witness's reputation.
Approximate physical common knowledge: what multiply-witnessed committed records would implement: shared optically grounded evidence that each agent knows the others share, with consistent multi-perspective fabrication, which the filing describes as empirically difficult for any subset (an expectation, not a reported measurement), at the protocol's resolution and uncertainty and assuming honest witnesses outside the fabricating subset.

3Verification beyond the beam

Verification past the single beam and frame: graduated levels, honeypots, provenance negation, liveness, freshness and family proofs.

Graduated verification levels: Levels 0 to 3 on a single device, from static ambient challenge-response through closed-loop trajectory and bifurcation-sweep probes to an optional holonomy cycle, a closed circuit through the interrogation settings; the filing relates the cost of verification to the curvature of the interrogation path, and treats any extension of that to hardness as a conjecture.
Honeypot and canary verification: deliberately easy challenges that a bounded surrogate can pass, monitored by a secondary higher-order check it is expected to fail. The pass and fail pattern is diagnostic evidence about the attacker among declared candidates, with residual ambiguity.
Anti-Truth Beam: provenance negation: a described check of whether a submitted observation lies outside the reachable region of a declared device family, a region bounded by the bifurcation set of that family, under declared attacker families and calibrated error tolerances.
Semantic Truth Beam: the optional extension to semantic claims. A claim is bound to committed physical evidence and checked for consistency with it, which is a provenance check on its evidential support and never an oracle for propositional truth.
Plenoptic-parallax liveness: liveness from angular consistency: an ordinary flat replay lacks the parallax of a physically present scene. An auxiliary meter composed with the others, with angular-replay spoofs expressly outside the geometric claim. Its angular-consistency residual joins the verisimilitude meter; the light-field hybrid of section 7 is its hardware form.
Designated-verifier and mutual N-party liveness records: per-block-salt and commit-reveal constructions for attesting presence empirically, under stated assumptions, to a chosen verifier, or among N parties simultaneously.
Multi-contributor freshness composition: run or opening seeds composed by commit-reveal from two or more declared contributors, reducing single-contributor trust given fresh entropy from each, independent control and an abort or fallback when a contributor fails; single-source or common-control composition is logged as reduced assurance.
Chameleon-hash retroactive evidentiary weight: records whose evidentiary weight a declared trapdoor authority can selectively rescale after the fact. The trapdoor holder can open one commitment to different values, so the primitive itself is deniable; any accountability rests on independent enforcement and on the fact of commitment, which is not erased.
Family membership proofs via bundle topology: a proposed construction for showing that a record came from a family of devices, by topological invariants of its attractor bundle, without identifying the individual device.

4The self-directed regimes

Inner Limager: perception turned inward: the instrument measures its own configuration, with boundary-state estimation and Fisher-information diagnostics.
Inner Creation: rendering turned inward: self-directed sculpting of the instrument's own physical state.
Nulling: scene-invariant rendering: driving the channel to hold its output constant as the scene changes, so that what remains is device signature. The anti-Limager, minimising scene information where perception maximises it.
Arnold tongue mapping: the scene's dynamical fingerprint under yoked operation, device and scene driven into a joint state: mapping entrainment regions across drive frequency and coupling strength.
Coupling state classifier: declaration by a classifier, under a declared decision rule, of whether device and scene are entrained; the filing lists its cross-scene generalisation among the open problems.

5Games and meters

Who plays whom, and the meters that score the play.

Alice, Bob and Eve: design and operation framed as a two-timescale Stackelberg game, Alice leading and Bob and Eve best-responding, with Eve as a system-wide hardening principle: bounded injectors, jammers and programmable-scene jamming as training adversaries, countered by multi-beam dampers on Alice’s side.
Load-bearing verifiers and diagnostic instruments: a two-tier verifier stack. Load-bearing paths, trained or calibrated against committed convolution-bundle ground truth and declared negative sets, would carry acceptance under the protocol digest, while unsupervised anomaly and drift methods run in parallel as diagnostic instruments that do not.
Sequential hypothesis-testing meters: verifier decisions by declared sequential tests that update accept, reject or continue as committed convolution-bundle windows arrive, rather than a fixed-length batch score.
Domain discriminators and domain generators: discriminators score whether a convolution bundle or derived artefact belongs to a declared domain; generators translate artefacts across domains, meaning deployments, sessions, modalities and governance states, under declared constraints, for training and for audit.
Frozen-meter adversarial tuning: pretrain a meter, freeze it as a fixed-capacity adversary or verifier, then tune the device and protocol to raise its difficulty, with declared freeze windows and drift meters guarding the frozen clock.
Self-adapting Reality Kernels: a self-edit head under a meta-reward lens, including an overnight self-editing pattern: the instrument would retune itself within hard safety and calibration constraints and an admissible edit envelope.
Federated and multi-agent learning: secure aggregation, differential-privacy accounting and poisoning diagnostics for trust-weighted fleet-wide model updates.

6Reality encryption and advantage channels

Recordings meant to be useful only to an authorised opener, and the advantage gap that opens between the legitimate channel and the adversary.

Reality encryption: emitting or transforming the stimulation pattern so that a recording would be useful only to an authorised opener, by key, model state or authority token, under a declared attacker model; others receive a neutralised release.
Neutral recording generator: the machinery that produces that key-less release, preserving declared verification and audit signals while suppressing protected content.
Staged advantage channels: three escalating constructions: keyed deterministic-noise obscuration, a trained neural obscuration generator, and scene-geometric advantage-channel discovery.
Key agreement from correlated measurements: a proposed construction in which two instruments observing the same physical process distil a shared key from their correlated records, on the assumptions that an adversary observes the process less well than the two instruments do and that they reconcile over an authenticated channel.
Physically rate-limited outputs: cryptographic outputs whose issue rate is bounded by a physical process, so that bulk forgery would cost bulk physics, assuming an accepted output cannot be produced faster by another route.
Wiretap reading of the Truth Beam: a viewpoint the filing offers, with coding embodiments alongside: the verification regime as a secrecy-capacity problem, the advantage gap between the legitimate channel and the adversary's.

7Matter as computer

Substrates the filing claims for the reactor role, from a gas discharge to a whole room.

Ponderable Cells: the massive-carrier counterpart of the reactor: mechanical, acoustic, fluidic, thermal, chemical or biological subsystems coupled into the same committed channel. Reactors, in the filing's sketch, for whatever has weight.
The Orb: a non-limiting label for spherical or otherwise enclosed reactor geometries, with emitter and recorder inside, the scene surrounding, and approximately isotropic coverage.
Plasma and time-varying reactors: gas-discharge micro-reactor arrays and epsilon-near-zero time-modulated media as channel substrates.
Light-field hybrid: emission and sensing paths realised as four-dimensional light-field subsystems, light-field projector in and plenoptic detector out, with the same angular-consistency primitive as the plenoptic-parallax liveness meter of section 3, against planar spoofs.
Matryoshka architecture: nested reactors, each a reactor in its own right: the inner operates inside the environment the outer establishes, with cross-coupling read out by multi-frequency demodulation.
Scene-phased full-parameter sweeping: sweeping the whole parameter vector in phase with the scene's own return: a conditional, local, first-order directional-derivative estimate by lock-in demodulation (conditions on the embodiments page), no differentiable surrogate model required.
Extended-scale and environmental embodiments: embodiments the filing describes at scales beyond the bench, geological, atmospheric and aquatic, including a passive GNSS Reality Kernel in which the satellite constellation would provide the emission protocol and the atmosphere and local multipath environment would act as the reactor.
Optical anchors: cross-layer continuous couplings binding stacked kernels to one another in real time, complementary to per-layer hardness; the intent is that a compromise of one layer would also have to satisfy its couplings to the others, which is not shown.
Optical long-term memory: write-once optical substrates holding governance evidence durably: hard-floor violation records and attestation tokens.
Operating modes A, B, C: settled raster, sampled continuous and pure analogue tuning: the three ways to run the same hardware.
PolieDeck: a tiled environment of projector-detector pairs across walls, ceiling and floor: the room itself as a coupled, witnessed scene.

8Twins, virtuals, counterfactuals

Models of the instrument, views that were never captured, and the marks the filing says a computation carries from the machine that ran it.

Digital twin and reality merge: a twin trained from committed probes, either simulated or a second physical kernel in twin-role, the AnaTwin, then reality-merged back to the physical substrate, with declared fidelity envelopes bounding how far the twin may be trusted.
Probe corpus: the committed dataset of semantically neutral probes from which the twin is trained, with held-out windows scoring its fidelity.
Virtual cameras and virtual emitters: synthesising views and emissions that were never physically instantiated, declared as such.
Forward counterfactual and inverse destyling generators: what would this scene look like under other controls, and what did the neutral scene look like under these, with a dual-direction consistency meter binding the two.
Payload-embedding Reality Transform: rendering that carries committed payloads chained to prior-frame hashes, with a perception-collapse semantic fallback: when the verisimilitude or dual-direction-consistency meters fall below committed floors for a sustained window, later emissions are restricted to a declared fallback regime until the meters recover, and the evidence already accrued stands.
Proof-of-physics: the filing's claim that a computation run on a specific physical device carries three marks, the device's microstructure attractor signature, a thermodynamic dissipation trace, and a holonomy encoding of the computational path, proposed as the basis for attesting which device computed what.

9Fleet epistemics

What a fleet can know about itself, and which parts of that knowledge are measured rather than assumed.

Five-term error decomposition: the measured response split into five named terms, a population model, a device signature, a shared discrepancy, an irreducible residual and measurement noise; attributing a residual to a cause needs further identifiability assumptions, which the filing states.
Security domain and discovery domain: the challenge space partitioned, committed before enrolment, into a security domain where devices authenticate and a discovery domain that feeds proof-of-discrepancy submissions and model improvement; acceptance meters and evaluation meters are kept apart by a separate declared partition.
Learnability-exclusion condition: the named condition under which no known feasible learning procedure can clone a device from bounded observations. The filing describes it as a condition to be verified empirically rather than proved, and as the physical basis of the security margin; no such verification is reported in the published record.
Trust-bootstrapped reference devices: a small set of high-assurance instruments from which fleet trust is grown outward.
Epsilon-nice kernels: the regularity condition, four Markov property functionals, approximate idempotence, contraction, mixing and commutation, each within its own tolerance, under which a kernel behaves stably enough to simplify composition and network operation.
Reality Kernel algebra: serial, parallel and containment composition of kernels, with algebraic laws treated as engineering targets rather than exact identities.
Bilateral Proof-of-Discrepancy: two independent devices commit pre-run predictions, jointly record a physical discrepancy event, and carry a portable, append-only record of every later appraisal: epistemic history as a first-class object.
Attested semantic computation records: when a routed ensemble computes something, the route and the result are committed together.
Interpretability probing via externalisation: an agent's internal model is probed by making it act through a kernel and reading the physical consequences.
Heterogeneous-provider corroboration: canonical derived values re-executed across two or more providers with declared diversity or non-common-control properties; disagreement is a logged anomaly, and single-provider runs carry reduced assurance.

10Agents and governance hardware

An agent acting through a kernel, and the separations built into the hardware to bound it.

Governance-partition isolation: the substrate that produces evaluation meters and other governance-side signals would operate in a partition physically and logically isolated from the agent's behaviour-generation pathways. Described as a structural property of the apparatus rather than a runtime monitoring discipline: separation of powers as designed, in embodiments that have not been built.
Self-modification channel limitation: the Inner Creation channel that lets a reactor reshape its own attractor landscape could in principle sculpt misleading patterns, so preferred embodiments rate-limit, protocol-restrict and log it, with the acknowledged inherent limit that every evidence layer shares one physical substrate.
RK-centred agent loop: the pattern the filing non-limitingly calls a PolieBot: an agent using kernel modules as its primary observation, verification and action-conditioning channel, its action space partitioned by verification requirement, unconstrained for internal and reversible steps, then RK-verified, hardness-gated, and multi-RK-verified classes corroborated across multiple independent modules.
Human-Coupled Operations: HC-Read, HC-Write, HC-Verify, HC-Entrain and HC-Create: human-mediated primitives placing a participant's sensory, motor and physiological channels inside the feedback loop, each defining its own training objective. HC-Actuate separately commits the outbound request, rendering parameters and response window under the same protocol digest. Described primitives, not deployed records.
Agent lifecycle and trajectory records: lifecycle states for a device or agent, being Embodied, Docked, Offline and Re-embodied, with state transitions logged in protocol digests and meter summaries. Filing 1 names the four states and refers the full state machine to Filing 2, as optional and non-essential.

11The human records

People inside the loop, and the records kept of them.

Portable human bonding record: a committed, portable record of sustained interaction between a person and a fleet of devices, storing commitments rather than a raw biometric archive, graduated by level and time-indexed; the filing treats the record as evidence of a committed human-fleet interaction history and of nothing else, expressly neither a device identity alone nor a human identity assertion alone, with a staleness window after which the bond lapses to advisory status. A described record, not a deployed one.
Subnet-scoped identity and attested-interaction tokens: identity and token economies scoped to a declared subnet, denominated in units the filing labels, non-limitingly, PoliePunts.
Red-team bounty primitive: declared, committed adversarial attempts as first-class, rewarded events that harden the subnet's detectors, structurally never aimed at applicants.

12Lenses, conjectures, open problems

Readings, conjectures, and the list of what the filing says it does not know.

Information Operations Taxonomy: the enumeration in the filing of the primitive information operations from which its regimes are composed, each a trainable axis with its own loss.
Rate-distortion reading of the Limager: perception as lossy compression with a declared budget.
Memory-adjusted data-processing bounds: data-processing inequalities corrected for channel memory, with multi-query composition.
Entropy rate and leftover hashing: the entropy rate of the convolution bundle as a calibration statistic, with conditional min-entropy feeding randomness extraction; the filing leaves the estimation of that min-entropy from data as a separate empirical problem and reports no bound.
Transfer entropy as a first-class primitive: directional information flow between device and scene, including a Fisher-information variant.
The autopoietic reading: the closed loop as organisational closure, the instrument as a self-producing system, offered as interpretation rather than claim.
The categorical unification: a weak category of dissipative Hamiltonian systems with Fisher-Rao enrichment, an attractor-bundle functor, the five substrate mappings (optical primitives, neural architectures, information-theoretic primitives, control-theoretic primitives, cryptographic primitives) as functors, and a self-modelling endofunctor, with five stated conjectures of its own: transferability, existence and structure of a self-consistent self-model, convergence via a depth curriculum, which the filing itself marks as contested, and universality of natural transformations.
Quantum classes Q0 and Q1: photon-counting (Q0) and coherent-interferometric (Q1) operation, a temporal-modulation bridge from the classical regime, and educational-kit forms, all described embodiments; the QFI-induced connection on the attractor bundle is theory, its curvature divergence an open question the filing ties to its first categorical conjecture.
The named conjectures: Physical Irreducibility, PoD Monotonicity, Cross-Regime Constraint and Yoked Channel Authentication Capacity, stated as open, in the document, on purpose; with the five categorical conjectures above, nine in all.
Declared open problems: the filing ends by saying what it does not know: survival of simultaneous tone multiplexing, the mapping from effective rank to task-useful rank, the drift-bias tolerance of physical gradient methods and controllability-scaling calibration, then an appendix of two research directions, the Yoked authentication-capacity conjecture and the cross-scene generalisation of the coupling-state classifier. That is the house epistemics, applied to itself.

Everything above is described in Filing 1 as part of a research programme, and an entry is a described embodiment unless its own words mark it as interpretation, conjecture or proposal. The physical envelope is still being characterised, and the only demonstrated, recomputable system is the digital Truth Beam at truthbeam.com. This atlas is a reading aid: where it disagrees with the filing, the filing wins. All of the technology described here is patent pending. Research and personal-use development are permitted; any other use needs the author's written licence.