Page map
01 / 12Legal AI shouldnot only writelike a lawyer.It should learnto think like one.
Component relationships · not a release sequence
Generative models already produce convincing legal language. OLC Engine is being built for the step that matters next: a reviewable system that can work through a legal matter using the recurring cognitive architecture of expert legal practice. The method is intended to travel; its legal authority, language and source systems remain jurisdiction-specific.
The model supplies generative capacity. The engine supplies the legal cognitive architecture.
- Current build
- PRIVATE · ACTIVE DEVELOPMENTCurrent version · 0.0.7
- Implementation generation
- Third major iteration
- Next gate
- Production cutoff and migration under the OLC product identity
- Verified
- 2026-08-27
01 / THE NEW PARADIGM
The next legal instrument is not a faster typewriter.
The pen preserved legal thought.
The typewriter accelerated its transcription.
The computer made legal material searchable, reproducible and computable.
Generative AI can now produce convincing legal language in seconds.
01Threshold / generative capacity
Generative models have crossed an important threshold. They can explain a provision, summarise a decision, compare passages and produce legal-sounding text with a speed that changes the economics of knowledge work. That achievement is real. It is also insufficient.
02Failure state / legal cognition
Those are not merely writing defects. They are failures of legal cognition.
Legal language is the visible end of legal work, not its complete substance. A fluent answer can still rely on the wrong historical version of a statute; a persuasive memorandum can restate an allegation as though it were proved; and a confident recommendation can overlook an adverse decision, an exception, a procedural obstacle or a missing document that would change the result. Even a polished appeal can fail because no one first built the matter, identified the decisive issues or tested the argument from the other side.
03Engine premise / disciplined method
OLC Engine begins from a different premise: the central problem in legal AI is not how to make a model sound more like a lawyer. It is how to give generative capacity a disciplined legal method—one that can decide what work is required, use the right sources and tools, preserve the status of the material it receives, expose the basis of its work and stop when the record cannot support a responsible result.
From tools for lawyers to an architecture of legal work
The pen externalised thought. The typewriter standardised expression. The computer organised files, calculated dates and made large bodies of law searchable.
Each tool changed legal practice, but the trained practitioner still supplied the cognitive architecture: what to notice, what to doubt, what to verify, what to compare and what form the result should take.
Generative AI creates the possibility of moving part of that architecture into software.
Not the personality, consciousness or professional authority of a lawyer. Not a theatrical display of hidden model reasoning. What can be represented is the recurring discipline of expert practice:constructing a matter, preserving competing accounts, resolving authority, testing hypotheses, examining the opposing position, forming strategy, controlling expression and returning the result to a responsible person.
This is the paradigm OLC Engine is designed to pursue.
The ambition is not lawyer-like prose.It is expert-level legal cognition under inspectable constraints.
02 / THE TRAINED LEGAL MIND
Legal expertise is a way of directing attention.
That means more than retrieving a provision and more than drafting a plausible paragraph.
It means recognising the kind of problem presented, directing attention to what is material, separating an assertion from its proof, resolving the law that applied at the relevant time, testing competing legal positions, anticipating challenge, choosing a defensible course, producing the right form of legal work and revising the result when its basis does not hold.
- 01Perceive
Expert lawyers do not simply know more legal words. Their professional practice is shaped by a trained, recurring way of perceiving and testing a problem.
- 02Distinguish
They ask who is speaking and in what capacity.
They distinguish what happened from what someone claims happened.
They notice which date changes the applicable law, which fact changes the legal characterisation and which missing document prevents a conclusion.
They preserve alternative interpretations long enough to test them.
They anticipate what an opponent, court, authority or reviewer is likely to challenge.
They change strategy when the evidentiary basis changes.
They know that a short answer, a legal opinion, a contract, an objection and an appeal demand different structures, sources and standards of review.
- 03Contextualise
Judges, advocates, prosecutors, in-house counsel and public-sector lawyers work under different duties, procedures and genres.
Jurisdictions change the governing sources, legal language, institutions and rules by which authority is established.
The recurring discipline remains recognisable: identify what is material, establish what can be supported, resolve authority, test competing accounts, anticipate challenge, choose a defensible course and review the result.
- 04ApplyEngine rule
OLC Engine is designed around that common cognitive method while each role, task and jurisdiction supplies its own constraints.
Universal cognitive architecture. Jurisdiction-specific legal authority.
What universality means here
Common method · local authority · reviewable application
Universality is a system-level engineering thesis, not a claim that one current implementation already knows every legal order. It means translating the recurring operations of legal work into bounded stages, typed records, source-specific authority contracts, explicit revision paths, reviewable outputs and stopping conditions.
Application remains local. Legislation, case law, procedure, institutions, language and professional duties must be supplied through governed jurisdictional and task-specific contracts. The method can be shared without pretending that legal authority can be copied from one country to another.
Inspectable emulation does not mean consciousness, professional authority or publication of private chain of thought. It means a reviewer can see which task was accepted, which sources were resolved, which evidence was admitted, which assumptions constrained the analysis, which checks were performed, what remained uncertain and why the process completed, narrowed its answer or stopped.
ResolutionA universal legal method is credible only when every jurisdiction-specific application remains explicit and reviewable.
03 / SYSTEM CORE
One engine. Independently governed components.One reviewable legal method.
01 / Identity
OLC Engine is being developed as a general-purpose legal cognition engine and as the private system core at the centre of OpenLegalCore. Slovenia is its first reference jurisdiction: the present legal-source paths are Slovenian, while the target architecture is designed to connect independently governed jurisdictional sources, matter evidence, specialist components, analysis, strategy, drafting and review through one inspectable legal method.
OpenLegalCore is not a single model wrapped in a chat window. It is a system of independently governed components.
02 / Governed components
Legislation and case-law pipelines maintain legal records under their own source, identity, temporal and release contracts. OCR and document preparation turn difficult files into page-aware material. Extraction components propose evidence-bound legal facts. Specialist tools address tasks that require their own calculations or decision traces. Client surfaces own interaction, authorised files and the human review path.
03 / Composition
A source pipeline can maintain a statute without knowing why one version matters to a dispute. An extractor can identify a proposition without deciding whether it is contested, material or safe to carry into a draft. A client can preserve a conversation without turning it into evidence. A model can generate text without knowing whether the matter has been adequately constructed.
OLC Engine is the core because it composes those capabilities into legal work.
The Engine supplies task interpretation, sequence, context, admissibility, policy, stopping and return. It determines which bounded capability should participate in a legal task, what contract it must satisfy and what may proceed to the next stage.
Two forms of memory. One legal method.
01 / Long-term legal memory
Maintained legal sources form the long-term legal memory available to the Engine. Legislation, case law, legal language, source identity and temporal rules remain jurisdiction-specific and independently governed.
02 / Working matter memory
Case Memory is the authorised working memory of one legal matter in the target architecture. It preserves participants, events, claims, evidence, conflicts, issues and corrections without turning a conversation into proof or a model summary into fact.
03 / Selective cognition
The Engine provides the cognitive movement between them: interpret the task, decide what to retrieve, construct the matter, test evidence against authority, challenge alternatives, choose the responsible course and control expression. Selective orchestration is designed for speed because it invokes the work the task requires rather than every available branch.
04 / Explicit separation
Explicit separation is designed for accuracy: discovery is not authority, allegation is not evidence, matter memory is not law and generated language is not a legal conclusion. The method is designed to travel. Its legal memory must be rebuilt, governed and verified for each jurisdiction.
Component relationship and authority boundaries
| Layer | Owns | The Engine does not absorb | Participation |
|---|---|---|---|
| Maintained legal sources | Identity, authority, temporal record and release boundary | Source authority | Resolves and returns admissible legal material. |
| Documents and evidence | Original-file control, page continuity, locators and evidence preparation | Custody of every original file | Supplies reviewable matter material. |
| Specialist functions | Their own legal inputs, calculations, traces and limitations | Their specialist rules | Performs bounded tasks when the matter requires them. |
| Client surfaces and processes | Interaction, authorisation, original files, review and final action | Human and organisational responsibility | Receives work and returns it to the responsible process. |
TARGET COGNITIVE ARCHITECTURE / CONCEPTUAL SPINE
The recurring movements of reviewable legal work.
The target architecture follows eight connected movements. They are a public conceptual model, not a claim that all eight run in 0.0.7 and not a disclosure of hidden model reasoning.
ATTENTION → MATTER MODEL → AUTHORITY → HYPOTHESES → ADVERSARIAL TEST → STRATEGY → EXPRESSION → REVIEW
Attention
Recognise the task, the parties and their roles, the procedural posture, the requested outcome, the material inputs and the information that is missing. Direct work only where the matter requires it. Clarify, narrow or refuse when the task cannot be responsibly framed.
Matter model
Construct a bounded representation of the legal matter. Preserve events, attributed assertions, denials, admissions, documents, evidence, official findings, conflicts, corrections and unresolved questions as distinct records. Do not let conversation become proof merely because it is available.
Authority
Resolve the correct source, identity, locator and temporal version through the contract of the relevant source family. Keep discovery separate from authority. A relevant passage becomes usable only after the authoritative record and its legal role are established.
Hypotheses
Form and compare plausible legal characterisations. Test each against the material facts, admitted evidence, applicable law, interpretive alternatives and procedural position. Preserve more than one account when the record does not yet resolve the conflict.
Adversarial test
Examine adverse authority, likely opposing arguments, exceptions, evidentiary gaps and the conditions under which a conclusion would fail. A legal position becomes stronger when it survives challenge, not when the first generated explanation sounds confident.
Strategy
Compare available actions, consequences, risks, fallback routes and next evidence. State the objective, assumptions and failure conditions of each option. Strategy is a reviewable work product, not a confidence score attached to one prediction.
Expression
Produce the form of legal work required by the task from an approved analytical basis and admitted sources. A short explanation, research memorandum, contract, objection and appeal do not share one structure or authority burden. Genre is an explicit policy.
Review
Check grounding, citations, source coverage, evidentiary support, allowed conclusions, genre, missing information and the reason to complete, narrow or stop. Return the result, its limits and the material next action to a responsible human process.
04 / CURRENT RECORD
A working foundation. A larger architecture. A strict boundary between them.
OLC Engine is an active private development project. Version 0.0.7 is a working Slovenian legal-retrieval and source-bound answer foundation. It is not yet the complete legal cognition system described by the target architecture on this page.
The distinction is deliberate. OpenLegalCore records what works in the present core, what was built or materially prototyped in a retired predecessor, and what remains behind a future release gate. The ambition can be broad without turning planned capability into shipped capability.
- Status
- PRIVATE · ACTIVE DEVELOPMENT
- Current version
- 0.0.7
- Code / artefact
- GITHUB / PRIVATE
- Role
- General-purpose legal cognition engine · system core
- Current foundation
- Slovenian legal retrieval and source-bound answers
- Implementation generation
- Third major iteration
- Licence
- NOT YET DECIDED
- Next gate
- Production cutoff and migration under the OLC product identity
- Public access
- No source, download, licence or access-on-request is implied
CURRENT PRIVATE BUILD · v0.0.7
What is available now
Version 0.0.7 can interpret bounded Slovenian legal-source requests, plan the required source work, resolve legislation and case law through their maintained data services, select admissible material, produce cited text or structured legal output, validate source use and return an explicit completion state. It exposes a private OpenAI-compatible text interface with a reference profile verified in Open WebUI 0.9.5 and 0.11.3.
Persistent Case Memory, production Document Intelligence, general analysis of complex matters, procedural strategy, specialised legal drafting, deadline and jurisdiction tools, additional European source families, durable Analyst findings and controlled whole-file review are not part of 0.0.7.
They appear on this page because they define where the Engine is going. They do not appear as if it has already arrived.
05 / CURRENT PRIVATE BUILD · v0.0.7
The present core proves the discipline at the source boundary.
Version 0.0.7 is deliberately compact. Its present path is:
This is narrower than the target cognitive architecture. It establishes several of the contracts on which the larger system depends.
- Frame the request
- Plan source work
- Discover candidates
- Hydrate authority
- Select admissible material
- Produce + validate a cited result
Bounded text interaction
The current private service accepts text-based, multi-turn work through an OpenAI-compatible interface. The client supplies the messages required for the current task; the Engine does not silently turn chat history into a legal case file or treat an earlier user statement as proof.
Input is validated rather than silently shortened or accepted under unsupported parameters. That boundary matters in legal work: if a system discards part of a request without notice, later claims about completeness become unreliable.
Typed legal-source planning
The Engine interprets whether a request requires an exact legislative source, exact case law, semantic discovery across legislation or case law, or a bibliographic literature path. It can proceed, request clarification or report that the request is unsupported under the current contract.
The model may interpret meaning, but deterministic contracts decide whether a proposed source path is admissible and how its result may be used.
Maintained Slovenian legislation
The legislation path supports exact and semantic discovery, structured provision text, current status and selection of the formal historical version relevant to a specified date. The discovered candidate is hydrated from the maintained authoritative data service before it is admitted to an answer.
If the historical chain is missing, ambiguous or inconsistent, the Engine does not silently replace it with the current provision and present that text as historical law.
A formal validity interval establishes the temporal record. It does not by itself resolve every transitional provision, exception or substantive question of legal application. Historical retrieval is a prerequisite for temporal legal analysis, not a substitute for it.
Maintained Slovenian case law
The case-law path supports exact identity by recognised case identifiers, semantic discovery, deduplication and hydration of complete decisions from the maintained legal data service. A missing, conflicting or ambiguous exact identity fails closed rather than becoming a plausible-looking citation.
The current selector can decide which retrieved decisions are admissible for the immediate answer. It does not yet perform the full legal work of establishing a line of authority, comparing favourable and adverse decisions, distinguishing material facts or building procedural strategy. Those capabilities belong to later analysis gates.
Legal literature
The literature path is bibliographic only. It may discover and verify the identity, author, title, publication and other admitted bibliographic facts of a potentially relevant work.
Without admitted full text, the Engine does not infer doctrine, reasoning or conclusions from a title, metadata or apparent prestige. The distinction prevents a citation from appearing more authoritative than the material actually available to the system.
Source-bound results
The current Writer receives only material that was actually retrieved, hydrated from its authoritative service, admitted under the contract of its source family and supplied with a stable identity and locator. Ordinary answers use visible source labels and conclude with a source record. Structured legal output can preserve status, claims, source references, notice and sources in a fixed evidence-bound form.
An exact-source path may report completion when the identified material and answer contract have been satisfied. A path using semantic retrieval remains conservatively partial because similarity search cannot prove exhaustive coverage of a corpus. Fluency never upgrades coverage.
Controlled delivery and reference client
The current service supports streaming and non-streaming text delivery, public workflow-status signals, usage reporting and cancellation when a client disconnects. Public status describes the stage of work; it is not private chain-of-thought or raw model reasoning.
Open WebUI 0.9.5 and 0.11.3 are the current verified client builds for the recorded private text-only, multi-turn profile. That verification does not extend to every other Open WebUI version, every OpenAI API surface, file processing, multimodal work, persistent Case Memory or every named client under evaluation.
Inspect the bounded compatibility record06 / PREDECESSOR EVIDENCE · NOT IN CURRENT BUILD
The current core is deliberately smaller than what came before.
OLC Engine is now in its third major implementation iteration.
A broader retired private predecessor explored much more of the intended legal-work model. It implemented or materially prototyped broad legal intake, conditional execution, richer source assessment, document analysis, matter-ledger foundations, structured legal analysis, procedural strategy, multiple forms of legal writing and explicit quality control.
That work matters. It demonstrates that the product direction did not begin with a diagram. It also revealed where breadth had outgrown architectural control.
The current core is a deliberate reconstruction: a smaller foundation with clearer contracts, request-bounded state, independent source authority, explicit lifecycle control and a release path that can be tested one capability at a time.
The predecessor contributes product knowledge. Its architecture is not the migration target.
What the predecessor proved
- 01
Broad legal intake
The system could distinguish different kinds of legal work, identify fields, issues, procedural position, roles, relevant dates, cited sources, missing information, risk and the required form of output. It could decide that a task needed clarification, legislation, case law, document analysis or a later analytical stage.
- 02
Conditional legal work
The task could determine which stages were required rather than sending every request through one fixed sequence. Short answers, source explanations, matter analysis, procedural strategy and formal submissions did not need to pretend to be the same job.
- 03
Documents and matter continuity
The predecessor developed page- and span-conscious document structures, attributed propositions, dates, conflicts, missing information and early matter-ledger controls. It explored corrections, revisions, bounded recall and the separation of user assertions from material safe for later work.
- 04
Analysis and strategy
Structured work products represented material facts, assumptions, legal issues, applicable rules, interpretive alternatives, source-bound application, conclusions, evidence gaps, opposing arguments, options, next actions and risk.
- 05
Controlled expression
Different output modes were designed for different legal tasks, from short explanations and research to opinions, correspondence, agreements, objections, appeals and other submissions. The Writer was intended to receive an analytical basis rather than invent the matter and the method while drafting.
- 06
Quality control
Checks addressed source grounding, citations, allowed conclusions, genre, truncation, missing sources, missing information and adherence to the analytical packet. The preferred correction path was explicit: expose the defect and decide what happens next instead of allowing an invisible model rewrite to erase the record.
07 / TARGET COGNITIVE ARCHITECTURE · RELEASE-GATED · MATTER + DOCUMENTS
The matter must exist before the answer can be trusted.
The target Engine will not confuse a conversation with a case file.
A conversation is transient, compressed and rhetorically loose. A legal matter contains people and roles, events and dates, allegations and denials, documents and evidence, claims and objections, procedural acts, corrections, conflicts and questions that remain open. Those categories do not become interchangeable merely because a model can place them in one context window.
- 01PHYSICAL PAGE
- 02REVIEWABLE PROPOSITION
- 03CANONICAL MATTER ITEM
Document Intelligence: from the physical page to a reviewable proposition
Every material fact derived from a document should retain a route back to the physical page and exact supporting span.
The target architecture therefore separates five things that ordinary document chat often merges:
PROVENANCE SEQUENCE
- 01the original file controlled by the authorised client process
- 02prepared page-aware text
- 03extracted candidate propositions and their exact locators
- 04selected canonical matter items
- 05later analysis or drafting that relies upon them
A retrieved chunk or model summary is never canonical evidence merely because it is convenient. If only part of a document was processed, omitted scope must remain visible. If a quotation cannot be matched to the source span, it must be rejected or quarantined rather than polished into apparent certainty.
Production Document Intelligence remains gated by page continuity, exact locators, coverage accounting, controlled canonicalisation and evidence from long, difficult legal files.
Case Memory: an auditable matter ledger
In the target architecture, Case Memory supplies the authorised working memory of a particular legal matter: a concise, canonical and versioned ledger that gives the Engine continuity across authorised work.
It does not replace the long-term legal memory held by governed source systems. It records which authority was resolved and how it bears on the matter while preserving the route back to the legislation, case law or other source contract.
It is not a transcript, a model summary, a collection of embeddings, an automatically stored answer or an undifferentiated long document. It is intended to preserve participants and roles; attributed assertions, admissions and denials; events and chronology; claims and objections; documents and evidence; procedural events and significant dates; legal issues; conflicts; missing information; corrections; and the relationships among them.
MATTER REGISTER
- 01Participants, capacities and authorisation
- 02Attributed assertions, admissions and denials
- 03Events, chronology and procedural acts
- 04Documents, evidence and exact provenance
- 05Claims, objections and legal issues
- 06Conflicts, corrections and unresolved questions
Canonical does not mean unquestionable. The ledger must show who asserted a proposition, what a document states, what an authority found, what remains disputed, what was corrected and which evidence supports each version. Conflicting items may coexist. A later statement must not silently rewrite the historical record.
Every item must retain a stable identity, matter and authorisation boundary, type, epistemic status, provenance, evidence references, revision history, conflict or supersession path and an explicit decision about whether it is safe for analysis or drafting.
08 / ANALYSIS · STRATEGY · EXPRESSION
Retrieval gathers material. Legal work begins when that material is tested.
A source-grounded answer is an essential foundation. The governing objective is material truth—not rhetorical confidence.
Legal work must establish what evidence and real circumstances support, resolve applicable legislation and relevant case law, test competing accounts and expose the uncertainty that remains. Complex work—an appeal, a response to a claim, a procedural decision, a legal opinion or the review of an extensive file—therefore requires an explicit analytical layer between evidence and prose. The Writer should not have to invent the matter, the legal method, the strategy and the final genre in one opaque act of generation.
Reviewable transformation / target architecture
Build an explicit analytical basis.
The target analytical work product connects material facts, unresolved assertions, document evidence and applicable legal authority. It should identify legal issues, assumptions, relevant rules, interpretive alternatives, source-bound application, provisional conclusions, evidentiary gaps, missing information, risk and the limits of permissible conclusions.
This is not a final legal opinion and not hidden model reasoning. It is the reviewable basis supplied to later stages.
- Receives
- Matter state + admitted authority
- Tests
- Issues, alternatives, gaps + permissible conclusions
- Returns
- Explicit basis for review
Compare action under visible constraints.
Strategy compares available actions under factual, evidentiary, legal and procedural constraints. It should make the party's objective visible, identify strengths and vulnerabilities, anticipate likely opposing arguments, expose missing evidence, compare procedural options and present a recommended route, fallback route, next actions and risk.
A strategy is not responsible because it carries a high confidence score. It is responsible when its basis, assumptions, alternatives and possible failure conditions can be examined.
- Receives
- Approved analytical basis + objective
- Tests
- Actions, consequences, opposition + failure conditions
- Returns
- Inspectable options, not invisible certainty
Make genre an explicit policy.
The target Writer produces the form of legal work required by the task from an approved analytical basis and admitted sources. Its policy may vary the model, reasoning profile, budget, structure, style and validation requirements, but those choices remain subordinate to the legal task.
Planned forms include concise answers, explanations of provisions, source-grounded research, long-form legal research and academic-grade analysis, matter and case-law analyses, document analyses, legal opinions, correspondence, agreements, demands, objections, appeals, criminal complaints, court and administrative submissions, assessments of prospects and risk, summaries and requests for missing information.
A short explanation and an appeal do not share one structure, authority burden or review path. Genre is a contract, not a prompt variation.
- Receives
- Approved basis + task-specific genre contract
- Controls
- Structure, style, source burden + budget
- Returns
- Required form of legal work
Report the defect. Decide the route.
Quality control checks whether the result remains grounded in admitted material, whether its citations support the claims made, whether coverage and truncation are visible, whether the genre and analytical basis were respected and whether the conclusion exceeds the available record.
The preferred response to a defect is a report and an explicit workflow decision. An invisible rewrite would make the correction harder to inspect than the original error.
- Receives
- Work product + execution trace
- Checks
- Grounding, support, coverage, genre + limits
- Returns
- Pass, defect report, revision or stop
The capability horizon
One discipline. Different burdens of work.
The target Engine is intended to support a spectrum of legal work:
The same cognitive discipline connects the spectrum and different legal roles. Judges, advocates, prosecutors, in-house and public-sector lawyers apply it under different duties; responsible deployments can make the same structured assistance available to people, businesses, institutions and civil-society organisations. The amount of source work, matter state, analysis, challenge, strategy, expression and review changes with the task. The obligation to keep authority, evidence, uncertainty and human responsibility visible does not.
- 01Legal question
- 02Research memorandum
- 03Complex matter
- 04Pleading or appeal
- 05Procedural strategy
- 06Whole-file review
09 / SPECIALIST FUNCTIONS
Some legal decisions need their own trace.
Not every legal problem should be delegated to a general Writer. Some tasks require a dedicated calculation, source contract or durable state that can be tested independently.
- 01Matter signal
- 02Typed input
- 03Specialist trace
- 04Reviewable return
01Calculation
Deadlines
A deadline function should distinguish calendar arithmetic from the legal characterisation that controls it. A limitation period, appeal deadline or procedural time limit may depend on the triggering event, service or receipt, the temporal version of the applicable rule, counting rules, non-working days, exceptions and disputed events that may interrupt, suspend or extend the period.
The result should expose the inputs used, legal basis, calculation steps, resulting date, assumptions, missing information and degree of certainty. It should warn when expiry may have occurred or may be approaching. It must not treat a disputed event as established merely because doing so produces a date.
- Receives
- Triggering event + applicable temporal law
- Tests
- Counting rules + exceptions
- Returns
- Date + basis, assumptions and certainty
02Decision path
Territorial and subject-matter jurisdiction
A jurisdiction function should identify a candidate court or authority from the proceeding and claim type, value in dispute, party status and location, place of the event, location of property, special or exclusive rules, jurisdiction agreements, cross-border dimensions and applicable temporal law.
“Candidate” is deliberate. Jurisdiction can depend on disputed facts and legal characterisation. The tool should expose the path, conflicts and missing inputs rather than disguise uncertainty as a database lookup.
- Receives
- Proceeding, claim, parties + location
- Tests
- Competing rules + disputed facts
- Returns
- Candidate forum + conflicts and missing inputs
04Controlled state
Durable findings
Some risks should outlive one generated answer.
The target Analyst layer can preserve unresolved legal uncertainty, missing facts or evidence, limited source coverage, procedural risks, approaching deadlines, conflicts and recommended next actions as typed findings with severity, evidence references, revision history and a controlled lifecycle.
A live notification is only a presentation signal. The authorised record must remain in the system that owns the finding, its history and its resolution. A model must not write arbitrary warnings directly into an uncontrolled database.
- Receives
- Unresolved risk + supporting evidence
- Controls
- Severity, revision + lifecycle
- Returns
- Authorised finding + next action
10 / EMBEDDED LEGAL WORK
A service core,not only a chat backend.
OLC Engine does not require one proprietary interface. A legal workspace, a compatible general client or an existing institutional process can provide the interaction layer, provided that task, data, source, authority and approval boundaries remain explicit.
In the current Slovenian reference jurisdiction, the Engine is designed to use maintained Slovenian legislation and case law, authorised matter evidence, analysis, specialist functions and controlled drafting. The reusable unit is the legal method, not the Slovenian source bundle.
A new jurisdiction cannot be added through translation or a database swap. It requires governed legal memory, local language and institutions, explicit authority and temporal rules, real legal workflows and acceptance evidence maintained with people who understand that legal order.
The cognitive discipline is intended to support judges, advocates, prosecutors, in-house and public-sector lawyers in the different forms their duties require. Through responsible deployments, the same architecture can extend structured legal assistance to individuals, businesses, institutions and civil-society organisations.
That is partnership work. A jurisdiction partner brings local legal expertise, source knowledge or stewardship, representative workflows and the people who can test the result. OpenLegalCore brings the reusable Engine architecture, component contracts and integration method.
A partnership can begin with one bounded process—such as compensation assessment or administrative decision support—then expand only after its sources, evidence path and outputs pass agreed review. These are architectural use patterns, not claims of existing partners, foreign deployments or autonomous decision-making.
- 01Reviewable return
- 02Documented boundary
- 03Authorised action
What remains under organisational control
The responsible deployment can retain authentication, user and matter authorisation, original files, uploads, document display, interaction, approval and the final act. The Engine can own legal planning, invocation of source and specialist components, bounded matter context, analysis and Writer policy, validation and execution records.
Responsible deployment retains
- Authentication and user authorisation
- Matter authorisation
- Original files, uploads and document display
- User interaction and review
- Approval and final act
OLC Engine can own
- Legal-task planning
- Invocation of governed source and specialist Components
- Bounded matter context
- Analysis and Writer policy
- Validation and execution records
A core that can be examined from more than one side
- 01 / Technical evaluator
The technical evaluator should be able to distinguish the working 0.0.7 foundation from the target architecture, follow public satellite components to their repositories and challenge the contracts proposed for private work.
- 02 / Legal specialist
The legal specialist should be able to inspect how the system treats legal authority, temporal law, evidence, dispute, adverse positions, uncertainty and human responsibility—and contribute legal method or source expertise without being forced into a code-only path.
- 03 / Organisation
The organisation considering defined work should be able to identify a concrete process, the data and authority boundary, the review path, the intended outcome and the conditions that must be agreed before private access, integration, evaluation or a pilot begins.
Access follows the work. It is not granted merely because an enquiry was sent.
The exact division may vary. Its documentation may not.
11 / DEVELOPMENT GATES
A roadmap of dependencies, not promises of dates.
The Engine roadmap records the order in which capabilities can become trustworthy. A target date may organise internal work. Only implementation, tests, acceptance evidence and explicit review can change a public capability status.
Release ruleA capability changes status only through implementation, evidence and review.
Phase 01 Foundation
R0–R1
R0 CURRENT PRIVATE BUILD · v0.0.70.0.7 FOUNDATION
The present Slovenian legal-retrieval and source-bound answer backbone: bounded planning, authoritative source hydration, temporal legislation, exact and semantic source paths, conservative selection, citations, structured output and controlled delivery.
R1 REQUIRED NEXT GATECUTOFF AND OLC MIGRATION
Establish the compact current core as a release-controlled foundation under the OLC product identity. Complete production hardening, lock regression and reference-client acceptance, align the recorded version and migrate without importing predecessor architectural debt.
Phase 02 Cognitive substrate
R2–R4
R2 TARGET COGNITIVE ARCHITECTURE · RELEASE-GATEDGENERAL PLANNING, CASE CONTEXT AND WRITER POLICY
Expand from source-query planning to broad legal-task intake. Add clarification-first behaviour, a typed execution plan, request-local case context and explicit policy for the model, reasoning profile, budget, genre and validation required by the task.
R3 TARGET COGNITIVE ARCHITECTURE · RELEASE-GATEDCASE MEMORY v1
Create the canonical matter ledger and bounded continuity across authorised work. The gate requires identity and isolation, controlled matter creation, candidate-to-canonical workflows, immutable revisions, correction, supersession, conflict handling, source manifests, an active case view and authorised recall across sessions and clients.
R4 TARGET COGNITIVE ARCHITECTURE · RELEASE-GATEDFILES AND DOCUMENT INTELLIGENCE
Connect prepared legal documents to exact page-aware evidence and selected matter items. Record coverage, retain exact locators, separate evidence candidates from canonical matter state and make omitted or unprocessed scope explicit.
Phase 03 Legal method
R5–R7
R5 PREDECESSOR EVIDENCE · NOT IN CURRENT BUILDTARGET COGNITIVE ARCHITECTURE · RELEASE-GATEDANALYSIS, STRATEGY AND SPECIALISED DRAFTING
Rebuild the strongest predecessor capabilities as reviewable, source-bound work products. The gate requires continuity of evidence and authority through analysis, competing positions, strategy and specialised legal genres, with task-specific validation and visible quality reports.
R6 TARGET COGNITIVE ARCHITECTURE · RELEASE-GATEDSPECIALIST LEGAL TOOLS
Provide reviewable deadline and jurisdiction determinations using typed legal inputs, temporal sources, deterministic calculations or decision traces, exception tests, uncertainty states and an explicit human-review boundary.
R7 TARGET COGNITIVE ARCHITECTURE · RELEASE-GATEDEXTENDED LEGAL SOURCES
Add independently governed European and additional Slovenian legal-source families without erasing their distinct legal roles. Each family must pass its own identity, authority, temporal, provenance, locator, citation and acceptance gate.
Phase 04 Operating surface
R8–R9
R8 TARGET COGNITIVE ARCHITECTURE · RELEASE-GATEDANALYST AND WORKSPACE
Turn unresolved legal, factual, evidentiary and procedural conditions into durable, authorised findings. Preserve evidence links, revisions, lifecycle transitions and safe presentation through an authorised client process.
R9 TARGET COGNITIVE ARCHITECTURE · RELEASE-GATEDWHOLE-FILE REVIEW AND BOUNDED MULTI-STAGE AGENTS
Support controlled work across extensive matters and large files. The gate measures decomposition, bounded context, evidence continuity, coverage, stopping conditions and human review—not the appearance of greater autonomy.
The target product path
- Authoritative source
- Physical page
- Matter ledger
- Analysis
- Drafting
- Durable findings
From an authoritative legal source and the physical page of an input document, through a canonical matter ledger, to reviewable analysis, strategy, drafting and durable warnings.
Its defining property is the separation of conversation, original documents, extracted candidates, canonical matter state, legal authority, analytical work, durable findings and the final human-reviewed result.
That separation allows the Engine to emulate the structure of legal work without claiming the professional authority of the person who performs it.
12 / FINAL INVITATION · BUILD THE CORE
Build the method, not only the model.
Jurisdictional work begins with scrutiny
A universal legal method becomes credible one governed jurisdiction at a time.
A new jurisdiction begins with local legal expertise, governed sources, a real workflow and acceptance evidence—not a language switch.
Scrutiny protocol
Architecture claim boundary authority + evidence
- 01 / Review
Review the architecture.
- 02 / Challenge
Challenge the claim boundaries.
- 03 / Test
Test whether authority and evidence can survive the journey from source and physical page to matter state, analysis, strategy and draft.
Participation register
Three independent routes into the work
Developers can strengthen typed components, interfaces, tests and lifecycle controls.
Lawyers, researchers and source specialists can challenge the legal method, authority model, evidence rules and output genres.
A jurisdiction partner brings the legal order, source landscape, language, representative workflows and people able to test the result. OpenLegalCore brings the reusable Engine architecture and component contracts.
A jurisdiction partnership defines purpose, roles, governed sources, acceptance evidence, access, licensing and publication boundaries before work begins.
12 / 12 · Public principle
- 01Source
- 02Record
- 03Reason
- 04Review
The law is too importantfor black-box systems.