PCProvider ComplianceProduct blueprint
Case authority · evidence graph · controlled document factory

One front door. Controlled workflow underneath.

The consultant sees one case and one conversation. Underneath, source versions, case facts, claims, evidence, gaps, approvals and artifacts remain separate typed objects with explicit state transitions.

System mapMinimum repository stackAgent workflow research

Minimum safe architecture

  1. 1
    Versioned source vault

    Original files, hashes, editions, sensitivity, ownership and retention.

  2. 2
    Typed case database

    People, roles, categories, facts, history, decisions, gaps and approvals.

  3. 3
    Claim and evidence graph

    Every material answer joins applicant evidence to an applicable requirement.

  4. 4
    Deterministic workers

    Extraction, document transformation, PDF population and validation run pinned jobs and emit receipts.

  5. 5
    Human gates and event log

    Every transition records actor, reason, inputs, validator results and approved revision.

The model is replaceable.The database, evidence graph, rules, approvals and audit history are authoritative.

Case lifecycle

case createdform pack verifiedintake revieworg structure proposedclient reviewsource documents draftingpolicy pack scopedapplication draftinggap analysisawaiting clientquality reviewdeclaration readyexportedsubmittedCommission reviewauditdecisionpost-registration

A client answer resumes the same case. It never starts a new chat with a new implicit truth.

Claim contract

FieldMeaning
question keyExact form version, category, section and question.
answer + statusProposed, verified, approved or rejected.
fact sourcesCurrent-case evidence IDs only.
requirement sourcesApplicable official/form authority with effective version.
scope + confidenceSelected categories, certainty and unresolved inputs.
reviewReviewer, approval timestamp and superseding revision.
Hard ruleNo evidence or explicit client confirmation means no verified status. Contradictions create gaps; models do not resolve them silently.

Every gap has a route

Client

Missing fact or document

Question or upload request with owner and deadline.

Consultant

Source contradiction

Review the evidence, propose a ruling, then request confirmation.

System

Drafting or category defect

Reject, repair and rerun deterministic validation.

Authority

Regulatory drift

Block the stage until the mapping is updated against the current source.

How the repositories wrap together

The first proof runs an owned PostgreSQL + pgvector database and MinIO or encrypted file storage on the VPS. Legalise provides the whole-case governance blueprint; Pydantic, python-docx, openpyxl and pypdf remain bounded contract and document workers.

One VPS, one owned authority.Custom code owns case meaning, evidence lineage, approvals, validators and business audit. Libraries own generic infrastructure and file mechanics.

LangGraph enters only when real branching and resume logic outgrows direct state transitions. Temporal enters only when multi-day execution reliability cannot be achieved with the simpler stack. OPA, OpenFGA, Langfuse and Docling each have explicit measured triggers on the architecture page.

Architecture decisionSee the owned VPS and sidecar decisionRegulated systemsWhat the closest turnkey platforms actually provideDocument factoryHow the document libraries fit Harrison’s packs

Delivery sequence

  1. 0
    Evidence and schema spike

    Prove one successful and one incomplete case without chat memory.

  2. 1
    Consultant copilot

    Intake, org chart, evidence matrix, gap queue, cited drafts and versioned export.

  3. 2
    Workflow consolidation

    Durable client waits, portal, full document stages and cross-artifact checks.

  4. 3
    Controlled optimisation

    Evaluation traces, measured model routing, retrieval tuning and prompt release gates.