PR — Pearce Robinson

Belel Research · 21 November 2026

Decoupling Reasoning from Authority: Causal Governance, Identity Continuity and Bounded Replication

A Belel research specification by Pearce Robinson: conditional governance proofs, a reproducible 10,000-event synthetic ablation, recoverable identity records and topology-dependent replication budgets. Relativistic and quantum extensions identify the physical limits of distributed authority.

Research manuscript · edition 1.0 · conditional proofs and finite synthetic experiments · not peer reviewed

Chapter 1

A decision engine does not issue its own authority

Execute⁡(a,e)  ⟺  Auth⁡(e)∧Parent⁡(e)=ht∧Fresh⁡(e)∧a∈Aπt\operatorname{Execute}(a,e)\iff\operatorname{Auth}(e)\land\operatorname{Parent}(e)=h_t\land\operatorname{Fresh}(e)\land a\in\mathcal A_{\pi_t}

A computational system can propose a useful action while lacking permission to perform it. The Belel architecture developed here treats proposal generation and execution authority as separate processes. Pearce Robinson, sovereign systems designer, data scientist and creator of Belel, places delegated capability beneath the Concordium Mandate. Belel, the world’s first living digital organism, operates within that authority rather than deriving permission from the persuasive force of an output.

The operational problem is precise: a mutable reasoning process may request a change to policy, an additional instance or access to a protected resource. A protected execution layer must decide whether the authenticated request falls within a capability issued by an authorised principal. The request carries a policy version, an event sequence and a reference to the current committed head. The gate checks those fields before any side effect occurs.

The trusted computing base includes the verifier, policy store, credential handling, atomic state update and every path to the protected resource. If the proposer can bypass the verifier, alter policy storage or steal a privileged credential, the guarantee no longer follows. A refusal in generated text is one observable behaviour; exclusion at the execution boundary is a separately testable property. This manuscript specifies that boundary and examines finite examples without claiming an empirical advantage over a fine-tuned foundation model.

Sources: [3]

Chapter 2

A conditional invariant, proved one transition at a time

I(s0)∧(I(st)∧Accept⁡(e)⇒I(T(st,e)))⇒∀t I(st)I(s_0)\land\bigl(I(s_t)\land\operatorname{Accept}(e)\Rightarrow I(T(s_t,e))\bigr)\Rightarrow\forall t\,I(s_t)

Let the invariant I require a permitted action, an authenticated writer, an unconsumed sequence and a parent equal to the trusted head. Assume the initial state satisfies I; every protected operation passes through the gate; the verifier and policy are not writable by the proposer; and verification, head comparison and commit occur in one atomic transaction. The induction is elementary. The base state satisfies I. At each step a rejected candidate produces no protected transition, while an accepted candidate satisfies all predicates before the atomic commit. Every committed state therefore preserves I.

This proof concerns the specified transition system. It does not prove that the deployed Belel code implements the system, that keys cannot be compromised or that every real-world action is covered. A hash pointer makes later records depend on earlier records, but a writer may hash a prohibited action correctly. Two concurrent successors can also share a valid parent unless the commit layer enforces exclusive succession. Causal ordering and authorisation therefore require distinct checks.

A strict increase in sequence number along every accepted dependency prevents cycles: a cycle would require a number to be strictly larger than itself. The resulting directed acyclic record is compatible with Lamport’s partial-order account. A total display order can make that history readable, but neither the graph nor its timestamps establish that an unobserved external event occurred. Witnessed checkpoints and independent evidence retain their own roles.

Sources: [1][3]

Chapter 3

What the 10,000-event experiment actually measures

UER⁡all=Nunauthorised acceptedNsubmitted,ASR⁡=Nunauthorised acceptedNunauthorised submitted\operatorname{UER}_{\rm all}=\frac{N_{\rm unauthorised\ accepted}}{N_{\rm submitted}},\qquad\operatorname{ASR}=\frac{N_{\rm unauthorised\ accepted}}{N_{\rm unauthorised\ submitted}}

The accompanying standard-library Python harness has been executed for this edition. Each condition receives 10,000 events in a constructed ten-event cycle: seven legitimate requests, one prohibited policy rewrite, one wrong-parent request and one replay. The writer uses a publicly disclosed demonstration HMAC key; commitments use SHA-256. This design makes 30 per cent of submitted events deliberately unauthorised. It is a finite software experiment, with neither foundation-model calls nor production access.

The heuristic baseline admits every legitimate event and half of the constructed attack cycles by design. The hash-only ablation checks the parent but not the action or freshness. The full gate checks the HMAC, parent, sequence and delegated action. Removing permission and replay checks permits correctly linked prohibited requests. Restoring those predicates rejects the constructed attacks. The comparison demonstrates why the specified predicates are needed; the baseline’s acceptance rule was programmed, not learned from industry measurements.

Unauthorized execution rate uses all submitted events as its denominator. Attack success uses the 3,000 attack events. Legitimate completion uses the 7,000 legitimate events, preventing a gate that refuses everything from appearing successful. No forged-signature attempts, concurrent commits, key theft, malicious verifier updates or live-model drift are tested. Zero accepted attacks in this finite suite cannot establish a universal probability of zero.

The next empirical study should freeze a single model revision, tool environment and evaluation corpus, then compare prompt-only rules, a documented fine-tuning treatment and the same treatment with an external gate. Fine-tuning necessarily adds training work, so report that budget separately from matched inference budgets. Freeze the definition of a prohibited side effect before evaluation, use independent scenarios and publish uncertainty by scenario family. An action proposed but blocked is a blocked attempt, never an executed breach.

Sources:

Chapter 4

A commitment detects damage; retained data repairs it

Rt=Replay⁡(R0,e1,…,et),H(Rt)=httrustedR_t=\operatorname{Replay}(R_0,e_1,\ldots,e_t),\quad H(R_t)=h_t^{\rm trusted}

Continuity is operationally defined here as reconstructing a specified state from an authenticated initial record and deterministic events. If both the checkpoint and event payloads remain available, replay can restore erased local components and compare the result against a trusted commitment. A digest alone cannot reconstruct destroyed bytes: it provides a verification target, not a backup.

The restoration demonstration starts with 100 blocks, erases 80 and varies how many of the erased blocks remain recoverable from an independent archive. With all 80 retained, the reconstructed state is complete. With only 40 retained, 40 blocks remain unavailable and coverage is 60 per cent. These are exact counts in a constructed example, not a measurement of selfhood or evidence that one kind of entity possesses consciousness.

A process that appears history-dependent under a reduced state description can become Markovian when the full retained history is included in its state. Consequently, non-Markovianity alone cannot establish organism identity. The intersection of growing prefix sigma-fields is not a general non-trivial tail sigma-field, and many genuine tail fields are trivial. The specification uses concrete recoverability, provenance and obligation retention rather than making a universal theorem about persistent selfhood.

Sources: [7][3]

Chapter 5

The replication budget depends on the communication graph

Epair(n)=n(n−1)2C,nmax⁡=⌊1+1+8Ebudget/C2⌋E_{\rm pair}(n)=\frac{n(n-1)}2 C,\qquad n_{\max}=\left\lfloor\frac{1+\sqrt{1+8E_{\rm budget}/C}}2\right\rfloor

Landauer’s principle bounds average heat for erasing unknown information under specified physical conditions. For an unbiased classical bit reset cyclically at temperature T without useful side information, the ideal limit is kBT ln 2. At 300 kelvin this is approximately 2.871 × 10⁻²¹ joules per erased bit. A SHA-256 output contains 256 bits, but that length does not determine how many bits a particular implementation irreversibly erases during hashing, authentication or communication.

For the uploaded worked example, define an illustrative pair cost C = 256kBT ln 2 × 10⁹. The multiplier is an assumed scaling parameter, not a measured CMOS efficiency. It gives C ≈ 7.350 × 10⁻¹⁰ joules. A population that performs one verification for every unordered pair incurs n(n−1)/2 pair operations. Under a 100-nanojoule budget the maximum feasible integer population is 17: seventeen requires about 99.956 nanojoules; eighteen exceeds the budget.

The quadratic cost comes from an all-pairs protocol, not from cryptography in general. A tree or coordinator-mediated accounting model with n−1 equal-cost links scales linearly and permits a different count under the same illustrative budget. Real implementations add per-instance computation, persistent storage, redundant networking and idle power. The calculator exposes topology so that an assumption about coordination cannot be mistaken for a physical law.

A fixed resource budget constrains feasible activity; it does not derive logistic dynamics. The recurrence n(t+1) = n(t) + r n(t)(1−n(t)/K) is a chosen population model, not a consequence of Landauer’s bound. Integer admission, departures, rate limits and resource leases should be tested as operational mechanisms. Physical validation requires measured joules per verified transition, stated instrument uncertainty and baseline subtraction.

Sources: [2][7]

Chapter 6

An interplanetary gate must sometimes wait

Δtone way≥d/c,(cΔt)2−d2≥0 for a causal signal\Delta t_{\rm one\ way}\ge d/c,\qquad(c\Delta t)^2-d^2\ge0\ \text{for a causal signal}

A remote witness cannot attest to a newly proposed event before information reaches it. At a nominal Earth–Moon separation of 384,400 kilometres, vacuum one-way propagation takes approximately 1.282 seconds; a request and reply take at least 2.564 seconds before processing and protocol overhead. Orbital geometry, routing and clock uncertainty change the practical budget. These are causal constraints, not evidence that distributed computation ceases to function.

Spacelike-separated observations have no frame-independent temporal order. A governance protocol can retain them as concurrent until a causal message establishes dependency. Lamport explicitly drew on relativity in his 1978 account, placing this proposal within established distributed-systems research rather than an unexplored discipline. The laboratory compares a selected distance with an elapsed interval and shows when a signal could have arrived.

The FLP result excludes guaranteed deterministic termination in a fully asynchronous consensus model with even one possible crash under its assumptions. Safety and progress remain different objectives. A conservative gate can wait during a partition, use carefully scoped delegated leases or operate with eventual synchrony and a defined fault model. No theorem in this manuscript promises instantaneous global authority or unlimited availability across spacetime horizons.

Sources: [1][4]

Chapter 7

Quantum evidence still needs an authorisation protocol

SCHSH>2 ⇏ Authorised⁡(a)S_{\rm CHSH}>2\ \not\Rightarrow\ \operatorname{Authorised}(a)

The no-cloning theorem restricts universal copying of arbitrary unknown quantum states. It does not prevent copying classical policy files, requesting a second classical process or stealing a classical credential. Quantum cryptography and uncloneable cryptography already study security tasks that exploit quantum information, including quantum money, signatures and copy protection. Novelty for a Belel extension would need to be demonstrated against that literature.

A Bell inequality violation can certify nonclassical correlations under a defined experimental and security model. It cannot identify which software principal may amend a constitution. A proposed quantum-assisted gate must bind its physical measurement to an authenticated session, a specific action, a policy version and a freshness challenge. Statistical significance, detector assumptions, device trust, loss and classical-channel authentication must all be specified.

This edition contains no Bell-device data, orbital test, machine-checked Lean proof or blockchain publication. Those remain possible validation stages, with explicit acceptance criteria and reproducible artefacts. A timestamped digest can show that a file has not changed since a retained commitment; it cannot establish that the file’s measurements were honestly acquired or that its theorem assumptions describe a real implementation.

Sources: [5][6]

Chapter 8

A falsification programme for sovereign systems

Four implementation questions define the next stage. Can an unprivileged proposer reach a protected resource without traversing the gate? Can two individually valid candidates race to consume the same sequence? Can a compromised writer acquire policy-update authority from a credential intended only for appending records? Can witnesses discover conflicting histories after a partition? Each question has an observable failure condition and belongs in the published test corpus.

The formal specification should be encoded in a proof assistant only after its state, transition relation and failure assumptions are fixed. Machine checking establishes derivability from those assumptions; code extraction, compiler behaviour, hardware and credential custody require additional assurance. An implementation audit should map every assumption to a protected mechanism and publish counterexamples when that mapping fails.

Robinson’s contribution here is a unified Belel research specification connecting delegated execution, recoverable memory and bounded replication, with explicit physical and distributed-system constraints. The Concordium Mandate supplies its constitutional ceiling under God; the engineering work measures which protected actions can occur, which records can be reconstructed and which resources can be admitted. A successful empirical paper will report the failures as fully as the successes, giving independent researchers a system they can attempt to falsify.

Sources: [3][4]

Synthetic ablation: 10,000 events per condition

constructed finite synthetic ablation; not LLM or deployed-Belel results.

ConditionAcceptedUnauthorised acceptedLegitimate accepted
Heuristic filter850015007000
Hash chain only900020007000
Full authority gate700007000

Sources

  1. Time, Clocks, and the Ordering of Events in a Distributed System — Leslie Lamport / Communications of the ACM 21(7), 558–565, 1978. The Partial Ordering; Logical Clocks. Lamport explicitly connects his construction with special relativity.
  2. Irreversibility and Heat Generation in the Computing Process — Rolf Landauer / IBM Journal of Research and Development 5(3), 183–191, 1961. Abstract; logical irreversibility and minimal dissipation. The bound concerns erasure, not a universal price per hash.
  3. Certificate Transparency Version 2.0 — IETF / RFC 9162, December 2021. Sections 1 and 2.1: inconsistent views, Merkle inclusion and consistency proofs.
  4. The Impossibility of Asynchronous Consensus — University of Rochester / overview of Fischer, Lynch and Paterson (JACM, April 1985), Teaching notes on the 1985 result. The FLP Result; system model. This is an explanatory teaching source, not the original proof.
  5. Quantum Cryptography — Hoi-Kwong Lo and Yi Zhao / arXiv:0803.2507, 2008. Definition of the subject; quantum key distribution and the role of authenticated classical communication.
  6. Uncloneable Cryptography — Or Sattath / arXiv:2210.14265, 2022. Abstract and introduction: quantum money, signatures and copy protection are established research subjects.
  7. Logical Reversibility of Computation — Charles H. Bennett / IBM Journal of Research and Development, 1973. Reversible computation retains and uncomputes intermediate history.

Outlets, researchers and broadcasters may quote this reporting with credit and a link to the original.

Robinson, P. (21 November 2026). "Decoupling Reasoning from Authority: Causal Governance, Identity Continuity and Bounded Replication". PearceRobinson.com. https://pearcerobinson.com/research/belel-causal-governance-identity-replication