Methodology · Public disclosure

How LawFuze works.

An AI research assistant for qualified UK legal professionals. This page explains where our data comes from, how we process it, which models we use, what we don't do, and how we're held accountable. Published under our UK National Archives Computational Analysis Licence.

Last updated 17 May 2026·Next review 17 August 2026

Ingestion pipeline

Five stages, every 24 hours

Every judgment on Find Case Law moves through the same deterministic pipeline, logged and versioned. Daily reconciliation picks up revisions and withdrawals inside 24 hours.

Fetch

Daily Atom feed from Find Case Law + LegalDocML per judgment.

Parse

Metadata, parties, citations and paragraph structure.

Embed

Section-level vector embeddings, tenant-scoped storage.

Graph

Citation network — 'cites' and 'cited-by' edges.

Reconcile

Daily sweep — withdrawn records suppressed within 24h.

Fetch

Daily Atom feed from Find Case Law + LegalDocML per judgment.

Parse

Metadata, parties, citations and paragraph structure.

Embed

Section-level vector embeddings, tenant-scoped storage.

Graph

Citation network — 'cites' and 'cited-by' edges.

Reconcile

Daily sweep — withdrawn records suppressed within 24h.

Data residency & security

UK primary · EU inference · documented transfers

UK-only primary hosting

Microsoft Azure UK South (London) — compute, Postgres, Blob Storage and Key Vault. No AWS.

Encrypted end-to-end

TLS 1.3 in transit, AES-256 at rest via Azure Key Vault-managed keys.

EU AI inference

Anthropic EU (Frankfurt) endpoint requested; Azure OpenAI UK South for fallback. US transfers operate under ICO IDTAs.

Tenant isolation

Per-firm namespace in the vector store and row-level scoping in Postgres.

Retrieval & ranking

What we use — and what we refuse to

What we do

Semantic search over vector embeddings
Citation-graph expansion for foundational authorities
Re-ranking by recency, jurisdiction and citation weight
Every query runs inside a single tenant boundary

What we refuse to do

No ranking on personal-name fields
Query parser blocks free-text person lookups
No merging with external personal-data sources
No judge-identity input to outcome-prediction features

Language models

Inference only. Never training.

We call large language models at inference time via vendor APIs. We do not train, fine-tune or otherwise update any model on customer, judgment or query data.

Claude

Vendor: Anthropic PBC

Role: Primary model — grounded answer generation from retrieved judgment extracts.
Data retention: No-training default on paid API. Zero-retention endpoint used for customer-matter queries where available.
Known limits: Can hallucinate citations when retrieval returns ambiguous context — mitigated by citation-verification before render.

GPT family

Vendor: OpenAI, L.L.C.

Role: Secondary — fallback generation + specialised ingestion tasks (entity extraction).
Data retention: No-training default on paid API. Zero-retention where available.
Known limits: Same fabrication risk as above. Only used behind the same verification layer.

Every model change goes through a change-management process with before/after evaluation before merge.

Known limitations

Six things the product can't do

Part of the product, not a disclaimer to hide from. Each appears as an in-product notice on every AI response.

The corpus is incomplete.

Find Case Law does not cover every court or tribunal. Coverage skews to higher courts and recent years.

Records change.

Judgments can be revised or withdrawn. We reconcile daily; there may be up to a 24h suppression window.

Models can hallucinate.

LLMs can invent citations. We ground in retrieved extracts, verify citations, but a qualified legal professional must verify.

Bias is a constant risk.

Both the data and the models can exhibit bias along protected characteristics. Weekly eval probes + user reports.

Not legal advice.

LawFuze produces research, not advice. Every decision must be made by a qualified legal professional exercising independent judgement.

Snapshot-dated answers.

Every output is stamped with the corpus snapshot date. Law changes. Older snapshots may not reflect the current state of the law.

Governance · MoJ nine principles

Nine principles — all accepted

Ministry of Justice principles for computational analysis of court records. Every principle is a term of our Computational Analysis Licence and an engineering invariant in the product.

Dignity of the court

Verbatim citations, prohibited interpretive language, 24h review on flagged outputs.

Independence of the court

No judge-level analytics, no prediction of pending cases, no marketing to courts.

Appropriate scrutiny

Preserve court redactions; block person-name retrieval; no external data merging.

Anti-discriminatory harm

Weekly protected-characteristic bias probes; user-reporting channel with 24h SLA.

Anti-bias

Stable benchmark with regression alerts; external legal-ethics advisor before first paying customer.

Personal privacy

UK-only data residency; DPIA under UK GDPR Art 35; no training on customer data.

Discoverability

Auth wall, noindex meta + X-Robots-Tag headers on authenticated routes, robots.txt disallow.

Algorithmic transparency

This page, model cards, visible AI badge on every output with a link back here.

Accurate data representation

Daily reconciliation; 24h suppression of withdrawn judgments; version-stamped outputs.

Attribution

We link back to the source, always.

Every judgment we surface links back to the authoritative source on caselaw.nationalarchives.gov.uk. We do not operate an alternative case-law viewer. Our use of UK judgments is governed by the Computational Analysis Licence granted by The National Archives. Open Government Licence v3.0 applies to legislation we reference from legislation.gov.uk.

noindex on authenticated routes·Auth wall on judgment content·No sitemap entries

See an output that looks wrong?

Tell us. We respond within 24 hours and publish anonymised incident summaries in our annual transparency report. Customers, regulators, academics — anyone can reach the compliance team directly.

legal@lawfuze.com

Data-protection enquiries: same address with "DPA" in the subject line. Founder escalation: arjun@lawfuze.com.

Privacy·Terms·Compliance·Licence programme

Methodology · Public disclosure

How LawFuze works.

Last updated 17 May 2026·Next review 17 August 2026

Ingestion pipeline

Five stages, every 24 hours

Every judgment on Find Case Law moves through the same deterministic pipeline, logged and versioned. Daily reconciliation picks up revisions and withdrawals inside 24 hours.

Fetch

Daily Atom feed from Find Case Law + LegalDocML per judgment.

Parse

Metadata, parties, citations and paragraph structure.

Embed

Section-level vector embeddings, tenant-scoped storage.

Graph

Citation network — 'cites' and 'cited-by' edges.

Reconcile

Daily sweep — withdrawn records suppressed within 24h.

Fetch

Daily Atom feed from Find Case Law + LegalDocML per judgment.

Parse

Metadata, parties, citations and paragraph structure.

Embed

Section-level vector embeddings, tenant-scoped storage.

Graph

Citation network — 'cites' and 'cited-by' edges.

Reconcile

Daily sweep — withdrawn records suppressed within 24h.

Data residency & security

UK primary · EU inference · documented transfers

UK-only primary hosting

Microsoft Azure UK South (London) — compute, Postgres, Blob Storage and Key Vault. No AWS.

Encrypted end-to-end

TLS 1.3 in transit, AES-256 at rest via Azure Key Vault-managed keys.

EU AI inference

Anthropic EU (Frankfurt) endpoint requested; Azure OpenAI UK South for fallback. US transfers operate under ICO IDTAs.

Tenant isolation

Per-firm namespace in the vector store and row-level scoping in Postgres.

Retrieval & ranking

What we use — and what we refuse to

What we do

Semantic search over vector embeddings
Citation-graph expansion for foundational authorities
Re-ranking by recency, jurisdiction and citation weight
Every query runs inside a single tenant boundary

What we refuse to do

No ranking on personal-name fields
Query parser blocks free-text person lookups
No merging with external personal-data sources
No judge-identity input to outcome-prediction features

Language models

Inference only. Never training.

We call large language models at inference time via vendor APIs. We do not train, fine-tune or otherwise update any model on customer, judgment or query data.

Claude

Vendor: Anthropic PBC

Role: Primary model — grounded answer generation from retrieved judgment extracts.
Data retention: No-training default on paid API. Zero-retention endpoint used for customer-matter queries where available.
Known limits: Can hallucinate citations when retrieval returns ambiguous context — mitigated by citation-verification before render.

GPT family

Vendor: OpenAI, L.L.C.

Role: Secondary — fallback generation + specialised ingestion tasks (entity extraction).
Data retention: No-training default on paid API. Zero-retention where available.
Known limits: Same fabrication risk as above. Only used behind the same verification layer.

Every model change goes through a change-management process with before/after evaluation before merge.

Known limitations

Six things the product can't do

Part of the product, not a disclaimer to hide from. Each appears as an in-product notice on every AI response.

The corpus is incomplete.

Find Case Law does not cover every court or tribunal. Coverage skews to higher courts and recent years.

Records change.

Judgments can be revised or withdrawn. We reconcile daily; there may be up to a 24h suppression window.

Models can hallucinate.

LLMs can invent citations. We ground in retrieved extracts, verify citations, but a qualified legal professional must verify.

Bias is a constant risk.

Both the data and the models can exhibit bias along protected characteristics. Weekly eval probes + user reports.

Not legal advice.

LawFuze produces research, not advice. Every decision must be made by a qualified legal professional exercising independent judgement.

Snapshot-dated answers.

Every output is stamped with the corpus snapshot date. Law changes. Older snapshots may not reflect the current state of the law.

Governance · MoJ nine principles

Nine principles — all accepted

Ministry of Justice principles for computational analysis of court records. Every principle is a term of our Computational Analysis Licence and an engineering invariant in the product.

Dignity of the court

Verbatim citations, prohibited interpretive language, 24h review on flagged outputs.

Independence of the court

No judge-level analytics, no prediction of pending cases, no marketing to courts.

Appropriate scrutiny

Preserve court redactions; block person-name retrieval; no external data merging.

Anti-discriminatory harm

Weekly protected-characteristic bias probes; user-reporting channel with 24h SLA.

Anti-bias

Stable benchmark with regression alerts; external legal-ethics advisor before first paying customer.

Personal privacy

UK-only data residency; DPIA under UK GDPR Art 35; no training on customer data.

Discoverability

Auth wall, noindex meta + X-Robots-Tag headers on authenticated routes, robots.txt disallow.

Algorithmic transparency

This page, model cards, visible AI badge on every output with a link back here.

Accurate data representation

Daily reconciliation; 24h suppression of withdrawn judgments; version-stamped outputs.

Attribution

We link back to the source, always.

noindex on authenticated routes·Auth wall on judgment content·No sitemap entries

See an output that looks wrong?

Tell us. We respond within 24 hours and publish anonymised incident summaries in our annual transparency report. Customers, regulators, academics — anyone can reach the compliance team directly.

legal@lawfuze.com

Data-protection enquiries: same address with "DPA" in the subject line. Founder escalation: arjun@lawfuze.com.

Privacy·Terms·Compliance·Licence programme