P15 — AIEP — Deterministic AI Agent Evidence Query Optimisation (ts)

Applicant: Neil Grassby
Inventor: Neil Grassby
Classification: withheld — internal
Priority / Alignment: Architecturally aligned with GB2519711.2

Abstract

A deterministic query optimisation system for artificial-intelligence agents querying externally published artefacts, wherein query canonicalisation, ordering, and resolution are governed exclusively by versioned registry rules. Indeterminate optimisation results in execution denial. The invention prevents heuristic or probabilistic optimisation from introducing non-reproducible query ordering into governed evidence retrieval pipelines.

Field of the Invention

[0001] The present invention relates to deterministic execution control in artificial-intelligence agent systems.

[0002] More particularly, the invention relates to a deterministic query optimisation mechanism for AI agents retrieving externally published evidence artefacts within an Architected Instruction & Evidence Protocol (AIEP) system as defined in United Kingdom patent application GB2519711.2.

Background

[0003] Artificial-intelligence agents performing evidence-based determination must retrieve artefacts from external sources in a reproducible and deterministic manner.

[0004] Conventional query optimisation systems employ heuristic ranking, learned retrieval models, or probabilistic scoring to order candidate artefact retrievals.

[0005] Such approaches introduce non-determinism into evidence retrieval: identical queries against identical artefact sets may produce different retrieval orderings depending on runtime state, model version, or environmental factors.

[0006] Non-deterministic retrieval ordering compromises the reproducibility requirements of AIEP-governed evidence substrates and may produce divergent determination outcomes across distributed nodes.

[0007] There exists a need for a query optimisation mechanism that enforces deterministic, registry-governed ordering of evidence queries such that identical queries against identical artefact sets produce identical retrieval sequences across all nodes and execution cycles.

Summary of the Invention

[0008] The invention provides a deterministic query optimisation system for AI agent evidence retrieval within an AIEP substrate.

[0009] The method comprises:

(a) receiving an evidence query from an AI agent;

(b) canonicalising the query in accordance with schema-defined canonical form;

(c) computing a QueryHash over the canonical query representation;

(d) retrieving applicable ordering rules from a versioned QueryOptimisationRegistry;

(e) applying registry-defined ordering rules deterministically to produce a canonical query execution sequence;

(f) executing the query sequence in canonical order;

(g) generating a QueryExecutionRecord comprising QueryHash, applicable registry version, and execution sequence; and

(h) suppressing query execution in a fail-closed manner when any optimisation step produces an indeterminate state.

[0010] The technical effect is reproducible, fail-closed evidence query behaviour ensuring identical retrieval sequences across distributed AIEP nodes operating over identical artefact sets and registry versions.

Brief Description of the Drawing

FIG. 1 — Deterministic Processing Pipeline

   ┌───────────────┐   InputType   ┌─────────────────────┐
   │  Input Object │──────────────▶│  NormalisationProfile│
   │  (any format) │               │  (version-bound)     │
   └───────────────┘               └──────────┬──────────┘
                                              │ parsing rules
                                   ┌──────────▼──────────┐
                                   │   Processing Engine  │
                                   │  • encoding rules    │
                                   │  • ordering rules    │
                                   │  • lossless check    │
                                   │  • fail-closed gate  │
                                   └──────────┬──────────┘
                              ┌───────────────┴──────────────┐
                     ┌────────▼────────┐           ┌─────────▼────────┐
                     │  CanonicalForm  │           │  Rejection Record │
                     │  H(CF‖ProfileID)│           │  (violated rule)  │
                     └────────┬────────┘           └──────────────────┘
                              │
                     ┌────────▼────────┐
                     │ Canonical Ledger│
                     │  (append-only)  │
                     └─────────────────┘

Definitions

[0011] QueryHash: A cryptographic hash computed over the canonical serialisation of an evidence query.

[0012] QueryOptimisationRegistry: A versioned registry defining deterministic ordering rules for evidence query execution.

[0013] QueryExecutionRecord: An append-only record comprising QueryHash, registry version identifier, canonical execution sequence, and timestamp.

[0014] Indeterminate State: A condition in which query optimisation cannot produce a single deterministic ordering under applicable registry rules.

Brief Description of the Drawings

Figure 1 illustrates query canonicalisation and QueryHash computation.
Figure 2 illustrates registry rule retrieval and deterministic ordering derivation.
Figure 3 illustrates fail-closed enforcement upon indeterminate state detection.

Detailed Description of Preferred Embodiments

1. Query Canonicalisation

[0015] Upon receipt of an evidence query, the system applies schema-defined canonicalisation comprising: stable key ordering; elimination of non-semantic variation; schema version binding; and timestamp assignment.

[0016] QueryHash is computed over the canonical query representation.

[0017] Identical queries produce identical QueryHash values across distributed nodes.

2. Registry-Governed Ordering

[0018] The QueryOptimisationRegistry maintains versioned ordering rules specifying the deterministic sequence in which candidate artefact retrievals are evaluated and executed.

[0019] Ordering rules are schema-bound and version-pinned.

[0020] Registry rules are applied to the canonicalised query to produce a single canonical execution sequence.

[0021] Heuristic ordering is prohibited. No ranking function dependent on runtime state, model weights, or probabilistic scoring is permitted.

3. Fail-Closed Enforcement

[0022] Where application of registry ordering rules cannot produce a single deterministic execution sequence, the system detects an indeterminate state.

[0023] Upon indeterminate state detection, query execution is suppressed in a fail-closed manner.

[0024] A failure record is appended to append-only lineage documenting the QueryHash and indeterminate condition.

4. QueryExecutionRecord

[0025] Upon successful deterministic query execution, a QueryExecutionRecord is generated and appended to append-only lineage.

[0026] The record enables replay certification of query behaviour.

Technical Effect

[0027] The invention ensures reproducible evidence retrieval ordering across distributed AIEP nodes.

[0028] The invention prevents heuristic or probabilistic ranking from introducing non-determinism into governed evidence pipelines.

[0029] The invention provides fail-closed enforcement when deterministic ordering cannot be achieved.

[0030] The invention enables replay-certifiable query execution traces.

Claims

1. A deterministic query optimisation system for AI agent evidence retrieval within an AIEP substrate, the system comprising: canonicalising a received evidence query; computing a QueryHash; retrieving ordering rules from a versioned QueryOptimisationRegistry; applying registry-defined rules to produce a canonical query execution sequence; executing in canonical order; generating a QueryExecutionRecord; and suppressing query execution fail-closed upon detection of an indeterminate state.

2. The system of claim 1 wherein heuristic ordering is prohibited and no probabilistic ranking function is permitted.

3. The system of claim 1 wherein identical queries against identical registry versions produce identical execution sequences across distributed nodes.

4. The system of claim 1 wherein a failure record is appended to append-only lineage upon indeterminate state detection.

5. The system of claim 1 wherein the QueryExecutionRecord enables replay certification of query behaviour.

6. A distributed computing system configured to perform the method of any preceding claim.

7. A non-transitory computer-readable medium storing instructions which, when executed, perform the method of any of claims 1–5.

Abstract

A deterministic query optimisation system for AI agents retrieving evidence artefacts within an AIEP substrate. Evidence queries are canonicalised and hashed via QueryHash. Deterministic ordering rules are retrieved from a versioned QueryOptimisationRegistry and applied to produce a canonical execution sequence. Heuristic and probabilistic ranking are prohibited. Query execution is suppressed fail-closed upon detection of any indeterminate optimisation state. QueryExecutionRecords enable replay-certifiable query behaviour. The invention ensures reproducible evidence retrieval ordering across all distributed nodes.