System and method for automated generation of invention disclosures with deterministic canonicalization and evidence-anchored provenance
The system addresses inconsistencies in invention disclosure generation by using Deterministic Canonicalization and an evidence ledger to ensure reliable cryptographic hashing and automated licence enforcement, resolving authorship and provenance uncertainties introduced by generative AI.
Patent Information
- Application Number
- AU2026201324
- Authority / Receiving Office
- AU · AU
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2025-05-23
- Filing Date
- 2026-02-22
- Publication Date
- 2026-11-05
AI Technical Summary
Existing invention disclosure generation processes suffer from inconsistencies in document formatting, field ordering, and encoding, leading to unreliable cryptographic hashing, provenance uncertainty, and manual licence enforcement, especially with the integration of generative AI models, which complicates inventorship and authorship attribution.
A computer-implemented method employing Deterministic Canonicalization, cryptographic hashing, and an evidence ledger to ensure reproducible and legally binding hash values, detect AI-generated content, and enforce licence terms programmatically.
Provides reliable cryptographic proof of authorship and content integrity, automates licence enforcement, and ensures auditable provenance for invention disclosures.
Smart Images
Figure 00000002_0000 
Figure 00000011_0000
Abstract
Description
TECHNICAL FIELD
[0001] The present invention relates to systems and methods for generating invention disclosure documents with cryptographically verifiable provenance. More particularly, the invention relates to Deterministic Canonicalization of invention disclosures, AI-content detection, licence enforcement, and evidence-anchored authorship binding. BACKGROUND
[0002] The generation of invention disclosure documents is a critical step in the intellectual property lifecycle. Existing disclosure generation processes suffer from several deficiencies. First, inconsistencies in document formatting, field ordering, and encoding introduce variance that prevents reliable cryptographic hashing --- the same substantive disclosure may produce different hash values depending on the software environment used to generate it.
[0003] Second, the increasing use of generative AI models to draft patent text creates provenance uncertainty. Patent offices and courts are grappling with questions of inventorship and authorship when AI-generated content is incorporated into invention disclosures without attribution. Existing plagiarism detection tools are not designed to detect the specific statistical patterns characteristic of generative model output.
[0004] Third, existing invention management systems do not provide cryptographic proof linking a specific disclosure document to a specific author at a specific point in time, creating vulnerability to disputes over priority, authorship, and originality.
[0005] Fourth, licence enforcement for derivative works based on disclosed inventions relies on manual tracking and contractual compliance rather than automated detection and enforcement. There is a need for systems that can detect derivative works through content affinity analysis and enforce licence terms programmatically. SUMMARY OF THE INVENTION
[0006] In accordance with the present invention, there is provided a computer-implemented method for automated generation of an invention disclosure document with evidence-anchored provenance. The method comprises receiving structured invention input data, applying a Deterministic Canonicalization transformation that removes generative variance to produce a legally binding, reproducible output hash, binding the canonicalized disclosure to an authorship identity via cryptographic hashing and digital signature, and recording a provenance event in a hash-chained evidence ledger.
[0007] As used herein, "Deterministic Canonicalization" means a transformation applied to invention input data that removes generative variance --- including whitespace irregularities, field ordering ambiguity, character encoding inconsistencies, and non-content metadata --- to produce a legally binding, reproducible output hash. Specifically, given identical invention input data I, the Deterministic Canonicalization function C produces identical canonical output C(I) on every invocation, such that SHA-3-512(C(I)) is a fixed, deterministic value. This property enables independent verification that a disclosure document has not been modified after canonicalization, and provides a cryptographically anchored basis for authorship binding, licence enforcement, and regulatory compliance.
[0008] In a further aspect, there is provided a system for automated generation of invention disclosures comprising one or more processors and memory storing instructions that, when executed, cause the system to perform the method described herein.
[0009] In a still further aspect, the system comprises an AI-content detection module that analyses submitted text for statistical patterns indicative of generative model output, including token distribution anomalies and hallucination markers.
[0010] In a yet further aspect, the system comprises a licence enforcement module that tracks derivative works through affinity analysis and enforces licence terms programmatically. BRIEF DESCRIPTION OF THE DRAWINGS
[0011] FIG. 1 shows a system architecture diagram illustrating the disclosure generation system with input reception module, Deterministic Canonicalization engine, authorship binding module, AI-content detection module, licence enforcement module, and evidence ledger components. DETAILED DESCRIPTION OF PREFERRED EMBODIMENTS
[0012] Deterministic Canonicalization comprises the following ordered steps applied to invention input data: (a) whitespace normalisation --- converting all line endings to line-feed (LF) characters, collapsing multiple consecutive whitespace characters to single space characters, and removing trailing whitespace from each line; (b) deterministic field ordering --- sorting all key-value pairs in structured data by lexicographic ordering of keys, applied recursively to nested structures; (c) terminology standardisation --- replacing variant spellings and synonyms with controlled vocabulary terms from a configurable terminology dictionary; (d) character encoding normalisation --- converting all text to UTF-8 NFC (Canonical Decomposition followed by Canonical Composition) form as specified in Unicode Standard Annex 15; and (e) metadata removal --- stripping non-content metadata including author tags, creation timestamps, revision identifiers, and application-specific markup. The ordered application of these steps ensures that the mathematical property of Deterministic Canonicalization is maintained: for any input I, the function C always produces identical output C(I), such that SHA-3-512(C(I)) is a fixed, deterministic value regardless of the software platform or execution environment.
[0013] Referring to FIG. 1, the system comprises an input reception module (201) that receives structured invention input data. The input data comprises one or more of: a technical disclosure narrative describing the inventive concept; structured claim elements specifying the components or steps of the invention; prior art references; drawings or diagrams; inventor identity data; and supporting evidence such as experimental data or simulation results.
[0014] The input reception module (201) validates the structure and completeness of the input data against a configurable disclosure template. Missing required fields generate validation warnings. The input reception module also performs initial format normalisation, converting input from various document formats into a canonical internal representation.
[0015] A Deterministic Canonicalization engine (202) applies the canonicalization transformation to the validated input data. The engine processes each field of the internal representation through the five ordered canonicalization steps described in paragraph
[0012] . The output of the Deterministic Canonicalization engine is a canonical disclosure document.
[0016] The Deterministic Canonicalization engine computes a content hash of the canonical disclosure document using SHA-3-512. The content hash serves as the primary identifier for the disclosure and is used in all subsequent provenance, authorship binding, and licence enforcement operations.
[0017] An authorship binding module (203) binds the canonical disclosure to an inventor identity. The binding comprises: computing a binding message from the concatenation of the content hash, the inventor identity (represented as a public key), and a timestamp; and signing the binding message using the inventor's private key. In the preferred embodiment, the signature uses Ed25519. In an alternative embodiment conforming to FIPS 204 (ML-DSA-65), the signature uses the CRYSTALS-Dilithium post-quantum signature scheme. In a further alternative embodiment conforming to FIPS 205 (SLH-DSA), the signature uses the SPHINCS+ hash-based post-quantum signature scheme.
[0018] The authorship binding module (203) supports multi-inventor disclosures wherein each inventor signs the binding message independently. The disclosure provenance record includes the ordered list of inventor signatures.
[0019] An AI-content detection module (204) analyses text within the invention input data for indicators of generative model output. The AI-content detection module performs the following analyses:
[0020] (a) Token distribution analysis: the module computes statistical properties of the token distribution in the submitted text, including token frequency distributions, bigram and trigram transition probabilities, and entropy measures. Generative model outputs characteristically exhibit lower perplexity and more uniform token distributions compared to human-authored text. The module compares computed statistics against reference distributions derived from known human-authored and known model-generated corpora using a calibrated classifier. The classifier produces a confidence score indicating the probability that the submitted text was generated by a generative model.
[0021] (b) Hallucination marker detection: the module identifies patterns characteristic of generative model hallucination, including: factual inconsistencies within the text; citations to non-existent publications or patent documents; technical descriptions containing plausible but incorrect technical details; and stylistic patterns characteristic of model-generated text including repetitive sentence structures, hedging language patterns, and characteristic transitional phrases associated with specific model families. The module maintains a configurable database of hallucination markers that is updated as generative model characteristics evolve.
[0022] (c) AI model output fingerprint detection: the module analyses the submitted text for statistical fingerprints associated with specific generative model families or architectures. These fingerprints include characteristic vocabulary preferences, syntactic patterns, and response formatting conventions. Detection is performed against a reference library of model output fingerprints derived from controlled sampling of tier-1 generative model frameworks.
[0023] The AI-content detection module (204) produces a report comprising: an overall AIcontent probability score; per-paragraph AI-content probability scores; identified hallucination markers with location references; identified model fingerprints; and a recommendation regarding 4 the suitability of the text for use in an invention disclosure. The report is recorded as a provenance event in the evidence ledger.
[0024] A licence enforcement module (205), also referred to as a LicenseChain module, tracks derivative works based on disclosed inventions. The licence enforcement module maintains a registry of canonical disclosure hashes and their associated licence terms. The module performs affinity analysis on new documents submitted to the system, computing content similarity scores against the registry using an Affinity Vault component. When a new document exceeds a configurable similarity threshold against a registered disclosure, the module flags the document as a potential derivative work and triggers licence compliance verification.
[0025] The affinity analysis comprises: (a) semantic similarity analysis using text embedding vectors computed by a pre-trained language model, comparing embeddings via cosine similarity; (b) structural similarity analysis comparing claim structures, component hierarchies, and technical architectures; and (c) terminology overlap analysis measuring the degree of shared technical vocabulary after normalisation. The combined similarity score is a weighted average of these three components. The Affinity Vault maintains an index of embeddings and structural features optimised for efficient similarity search across the registered disclosure corpus.
[0026] When a potential derivative work is detected, the licence enforcement module retrieves the licence terms associated with the matched disclosure and generates a compliance report. Licence terms may specify: attribution requirements; royalty obligations; scope limitations (e.g., field-of-use restrictions); geographic restrictions; and sublicensing conditions. The compliance report identifies which licence terms apply to the detected derivative work and the current compliance status.
[0027] An evidence ledger (206) records all provenance events in an append-only, tamperevident log. Provenance events include: disclosure creation events recording the content hash, inventor identity, and authorship binding signature; AI-content detection events recording analysis results; licence events recording derivative work detections and compliance determinations; and modification events recording any subsequent changes to disclosures. Each evidence record is hash-chained to its predecessor using SHA-3-512 and signed using Ed25519 or ML-DSA-65. In a preferred embodiment, the evidence ledger is anchored to a blockchainbased timestamping service to establish independently-verifiable timestamps.
[0028] The system maintains version control for canonical disclosures. When a disclosure is modified, the Deterministic Canonicalization engine produces a new canonical version with a new content hash. The evidence ledger records the relationship between the prior version and the new version, including a change delta describing the modifications. The authorship binding for the new version references both the current inventor signature and the content hash of the prior version, creating a cryptographically verifiable modification history. BEST METHOD
[0029] The best method known to the applicant as of the priority date for implementing the invention is as follows: The Deterministic Canonicalization engine is implemented as a pipeline of five sequential transformation stages as described in paragraph
[0012] . The SHA-3-512 hash of the canonical output is computed using a NIST-compliant SHA-3 implementation. Authorship binding uses Ed25519 for classical signatures and ML-DSA-65 (FIPS 204) for post-quantum signatures, with key pairs stored in hardware security modules or platform-local key stores. The AI-content detection module uses a multi-layer classifier architecture: a statistical layer computing token distribution metrics, a pattern-matching layer for hallucination marker detection, and a machine learning classifier trained on labelled human-authored and modelgenerated patent text corpora. The evidence ledger is implemented as an append-only database with SHA-3-512 hash-chaining, anchored via a blockchain-based timestamping protocol. The LicenseChain module uses text embedding vectors from a pre-trained language model with cosine similarity for semantic matching and a configurable threshold (default 0.85) for derivative work detection. The system is deployed as a set of containerised microservices communicating over mutual-TLS connections. WORKING EXAMPLE
[0030] Example 1 --- Disclosure Generation and Provenance: An inventor submits a technical disclosure comprising a narrative description, three claim elements, and two prior art references. The input reception module validates the submission against the disclosure template and identifies one missing recommended field (experimental data). The Deterministic Canonicalization engine processes the input through all five canonicalization steps, producing a canonical disclosure with SHA-3-512 content hash. The authorship binding module creates a binding signed with the inventor's Ed25519 key. The AI-content detection module analyses the narrative and reports an AI-content probability of 0.12 (below the configurable threshold of 0.50), with no hallucination markers detected. The provenance record is written to the evidence ledger and anchored to the blockchain-based timestamp service. The complete provenance chain --- from input reception through canonicalization, authorship binding, and AI-content analysis --- is recorded in the evidence ledger with SHA-3-512 hash-chaining. INDUSTRIAL APPLICABILITY
[0031] The system is applicable to any organisation that generates, manages, or enforces intellectual property, including patent filing entities, research institutions, technology transfer offices, licence management organisations, and regulatory bodies requiring auditable provenance for invention disclosures. REFERENCE SIGNS LIST
[0032] 201 --- Input reception module; 202 --- Deterministic Canonicalization engine; 203 -- Authorship binding module; 204 --- AI-content detection module; 205 --- Licence enforcement 6 module (LicenseChain); 206 --- Evidence ledger; Affinity Vault --- similarity search index within the licence enforcement module.
Claims
1. A computer-implemented method for automated generation of an invention disclosure document with evidence-anchored provenance, the method comprising: receiving structured invention input data comprising at least a technical disclosure narrative and one or more claim elements; applying a Deterministic Canonicalization transformation to the structured invention input data, the Deterministic Canonicalization transformation removing generative variance to produce a legally binding, reproducible output hash, the transformation comprising at least: whitespace normalisation, deterministic field ordering by lexicographic key sorting, terminology standardisation using a controlled vocabulary, character encoding normalisation to UTF-8 NFC form, and non-content metadata removal; computing a content hash of the canonicalized disclosure using SHA-3-512; binding the canonicalized disclosure to an authorship identity by signing a binding message comprising the content hash, the authorship identity, and a timestamp with a digital signature using a private key of the authorship identity; and recording a provenance event in an append-only, hash-chained evidence ledger, the provenance event comprising at least the content hash, the authorship identity, and the digital signature.
2. The method of claim 1, wherein the Deterministic Canonicalization transformation is deterministic such that for identical invention input data I, the transformation produces identical canonical output C(I) on every invocation, and SHA-3-512(C(I)) is a fixed value.
3. The method of claim 1, wherein the digital signature uses Ed25519.
4. The method of claim 1, wherein the digital signature uses a post-quantum signature scheme conforming to FIPS 204 (ML-DSA-65).
5. The method of claim 1, wherein the digital signature uses a post-quantum signature scheme conforming to FIPS 205 (SLH-DSA).
6. The method of claim 1, further comprising anchoring the evidence ledger to a blockchainbased timestamping service to establish independently-verifiable timestamps for provenance events.
7. The method of claim 1, wherein the evidence ledger is tamper-evident by virtue of hashchaining wherein each provenance event includes a SHA-3-512 hash reference to a preceding event.
8. The method of claim 1, further comprising analysing text within the invention input data for indicators of generative model output using an AI-content detection module.
9. The method of claim 8, wherein the AI-content detection module performs token distribution analysis comprising computing token frequency distributions, bigram and trigram transition probabilities, and entropy measures, and comparing computed statistics against reference distributions derived from known human-authored and known model-generated corpora.
10. The method of claim 8, wherein the AI-content detection module performs hallucination marker detection comprising identifying factual inconsistencies, citations to non-existent publications, and stylistic patterns characteristic of specific generative model families.
11. The method of claim 8, wherein the AI-content detection module performs AI model output fingerprint detection comprising analysing the text for statistical fingerprints associated with specific generative model architectures, the fingerprints including vocabulary preferences, syntactic patterns, and formatting conventions.
12. The method of claim 8, wherein the AI-content detection module produces a report comprising an overall AI-content probability score, per-paragraph AI-content probability scores, identified hallucination markers, and identified model fingerprints, and the report is recorded as a provenance event in the evidence ledger.
13. The method of claim 1, further comprising tracking derivative works through a licence enforcement module that maintains a registry of canonical disclosure hashes and associated licence terms.
14. The method of claim 13, wherein the licence enforcement module performs affinity analysis comprising semantic similarity analysis using text embedding vectors, structural similarity analysis comparing claim structures, and terminology overlap analysis, and wherein a combined similarity score exceeding a configurable threshold triggers licence compliance verification.
15. A system for automated generation of invention disclosures with evidence-anchored provenance, the system comprising: one or more processors; and a memory storing instructions that, when executed by the one or more processors, cause the system to: receive structured invention input data; apply a Deterministic Canonicalization transformation that removes generative variance to produce a legally binding, reproducible output hash; compute a content hash of the canonicalized disclosure using SHA-3-512; bind the canonicalized disclosure to an authorship identity via digital signature; and record a provenance event in an append-only, hash-chained evidence ledger.
16. The system of claim 15, further comprising an AI-content detection module configured to analyse submitted text for statistical patterns indicative of generative model output and to produce AI-content probability scores.
17. The system of claim 15, further comprising a licence enforcement module comprising a LicenseChain component that tracks derivative works through content affinity analysis and an Affinity Vault component that maintains an index of text embeddings and structural features for similarity search.
18. A non-transitory computer-readable medium storing instructions that, when executed by one or more processors, cause the one or more processors to: receive structured invention input data;apply a Deterministic Canonicalization transformation that removes generative variance to produce a legally binding, reproducible output hash; compute a content hash of the canonicalized disclosure using SHA-3-512; bind the canonicalized disclosure to an authorship identity via digital signature; and record a provenance event in an append-only, hash-chained evidence ledger.
19. The method of claim 1, wherein the structured invention input data further comprises prior art references, drawings, and supporting evidence, and wherein the input reception module validates the structure and completeness of the input data against a configurable disclosure template.
20. The method of claim 1, further comprising maintaining version control for canonical disclosures, wherein modification of a disclosure produces a new canonical version with a new content hash and the evidence ledger records a relationship between the prior version and the new version including a change delta.
21. The method of claim 1, wherein the authorship binding supports multi-inventor disclosures wherein each inventor independently signs the binding message and the provenance record includes an ordered list of inventor signatures.
22. The system of claim 15, wherein the Deterministic Canonicalization transformation comprises five sequential stages: whitespace normalisation converting all line endings to LF characters, deterministic field ordering by lexicographic key sorting applied recursively, terminology standardisation using a controlled vocabulary dictionary, character encoding normalisation to UTF-8 NFC form, and non-content metadata removal.
23. The method of claim 8, wherein the AI-content detection module maintains a configurable database of hallucination markers that is updated as generative model characteristics evolve.
24. The method of claim 13, wherein licence terms specify one or more of: attribution requirements, royalty obligations, field-of-use restrictions, geographic restrictions, and sublicensing conditions.
25. The method of claim 14, wherein the affinity analysis uses text embedding vectors computed by a pre-trained language model with cosine similarity comparison and a configurable similarity threshold.
26. The system of claim 15, wherein the digital signature uses Ed25519 for classical signatures and ML-DSA-65 conforming to FIPS 204 for post-quantum signatures.
27. The system of claim 15, wherein the evidence ledger is anchored to a blockchain-based timestamping service.
28. The method of claim 1, wherein the Deterministic Canonicalization engine is implemented as a pipeline of five sequential transformation stages and the SHA-3-512 hash of the canonical output is computed using a NIST-compliant SHA-3 implementation.