Constitutional substrate

The constitutional substrate is the lowest layer of AIEP. It defines what a valid instruction is, how alternatives are generated, how they are selected, and how every decision in the system is made tamper-evidently traceable. Every higher layer in the AIEP stack depends on the constitutional substrate being present and intact.

The core patent is GB2519711.2, filed 20 November 2025.


The instruction object

An AIEP instruction is not a natural language string. It is a structured machine-readable object with the following required fields:

FieldPurpose
intentionThe canonical statement of what the instruction is trying to achieve
parametersScoped, typed parameters that bind the instruction to its context
contextThe evidence context in which the instruction is valid
safetyEnvelopeAn identifier referencing the safety constraints that bound execution
protocolVersionA reference to the public protocol registry version in force at time of publication
divergenceGraphThe embedded graph of alternative canonical interpretations

The object is hash-bound at publication. Any alteration to any field produces a different hash — making tampering detectable by any retrieval agent.


The DivergenceGraph

The DivergenceGraph is the constitutional substrate’s core innovation. It embeds alternative interpretations of an instruction directly within the instruction object — not as ambiguity, but as structured, auditable alternatives.

Structure:

ElementDescription
Primary canonical instructionThe current empirical consensus interpretation
Alternative instructions (2–8)Generated exclusively using the defined divergence operators
Parent-hash chainEach alternative is cryptographically chained to its parent
Weighted scoringRegistry-retrieved coefficients determine which alternative is selected under any given context

The DivergenceGraph does not allow arbitrary alternatives. Alternatives may only be generated using the defined operator set — this is what makes the graph bounded and reproducible.


Divergence operators

Nine operators are defined. Each transforms the primary instruction in a specific, bounded way:

OperatorTransformation
D1Synonym substitution — replace a term with a semantically equivalent term
D2Parameter relaxation — widen a constraint
D3Parameter tightening — narrow a constraint
D4Context promotion — elevate a contextual condition to a primary condition
D5Context demotion — reduce a primary condition to contextual
D6Evidence substitution — replace the supporting evidence reference
D7Scope expansion — extend the domain of applicability
D8Scope contraction — restrict the domain of applicability
D9Temporal shift — shift the validity window of the instruction

These nine operators, applied to any valid instruction, produce the complete bounded space of alternative interpretations. No alternative outside this space is valid under the constitutional substrate.


Deterministic selection

When an AI agent retrieves an instruction, it must select from among the primary and alternative interpretations. This selection is deterministic — given the same context, schema version, and registry coefficients, every agent produces the same selection.

The selection algorithm:

  1. Retrieve the instruction object from the Mirror endpoint
  2. Retrieve weighted scoring coefficients from the protocol registry for the current protocolVersion
  3. Evaluate each alternative against the agent’s context parameters
  4. Compute a deterministic score for each alternative under the retrieved coefficients
  5. Select the highest-scoring alternative — or the primary if no alternative exceeds it
  6. Record the selection as a signed artefact in the evidence ledger

This is not ranking by probability. It is selection by deterministic evaluation against a declared, retrievable scoring function.


Canonical primitives R1–R8

The GENOME kernel defines eight canonical primitives that form the frozen core of the constitutional substrate. Every component at every layer of AIEP is built from these primitives:

PrimitiveFunction
R1Canonical serialisation — deterministic encoding of any artefact to a stable byte representation
R2Hash binding — SHA-256 of the canonical form, producing the artefact’s unique identifier
R3Provenance declaration — issuer identity, timestamp, and schema version bound to every artefact
R4Schema pinning — artefacts reference the exact schema version they were validated against
R5Divergence operator application — bounded transformation of instructions using D1–D9
R6Registry retrieval — deterministic fetch of scoring coefficients and schema definitions
R7Ledger append — tamper-evident appending of artefacts to the evidence ledger
R8Fail-closed gating — suppression of execution when any invariant is not satisfied

These primitives are frozen. A new version of the GENOME kernel does not alter R1–R8 — it only adds above them. This is what makes artefacts produced under different kernel versions interoperable.


GENOME_LOCKFILE.json

Every GENOME-compliant deployment commits to the kernel version at build time via GENOME_LOCKFILE.json. This file declares:

  • The kernel version in force (kernelVersion)
  • The canonical primitive set bound to that version (primitiveSet)
  • The schema versions the deployment accepts (schemaVersions)
  • A hash of the lockfile itself, signed by the deployer

A deployment that alters its lockfile after build produces a different hash — and any artefact it subsequently produces is traceable to the post-alteration state. This makes configuration drift detectable.


Why this matters

Most AI systems have no substrate. Instructions are text. Alternatives are probabilistic. Selection is opaque. The constitutional substrate replaces this with:

  • Instructions that are structured objects, not strings
  • Alternatives that are bounded and auditable, not sampled
  • Selection that is deterministic and retrievable, not probabilistic
  • Every decision traceable to the kernel version and registry state in force at the time

An AI system reasoning against the constitutional substrate is not a better language model. It is a different kind of system — one whose decisions can be investigated after the fact.


Architecture · GENOME SDK · Protocol · Patents — GB2519711.2