Milestones

Milestones

Detailed breakdown of each milestone — what is included, what success looks like, and what must come first.

v0.2 — Stable Pipeline

Objective: Eliminate critical technical debt that prevents the pipeline from running reliably on any machine.

Features included:

#	Feature	Priority	Status
F22	Standardize all paths via os.environ.get()	P0	Planned
F23	Extract pipeline/utils.py (parse_frontmatter, slugify, etc.)	P1	Planned
F24	Pin dependency versions in requirements.txt	P1	Planned
F20	Complete env var standardization (2 of 5 scripts done)	P1	In progress

Acceptance criteria:

• python3 pipeline/process_books.py --dry-run runs without error on a clean machine (with env vars set)
• No hardcoded absolute paths in any script
• parse_frontmatter() and slugify() exist in a single shared module (utils.py)
• pip install -r pipeline/requirements.txt installs deterministic versions

Estimate: 1 work session

---

v0.2.5 — Data Validation (Proposed)

Objective: Validate metadata quality before building innovation layer on top. This milestone was proposed after the PREMORTEM analysis identified unvalidated metadata (PF01) as an existential risk.

Actions included:

#	Action	Risk mitigated	Effort
M06	Validate 200 random chunks — verify instituto, tipo_conteudo, ramo against actual content	PF01, RT11	2-3 hours
M07	Create evaluation set: 30+ queries with expected results, measure recall@5 and nDCG	RE04	2 hours
M10	Schema validation in enrichment — validate LLM output fields against known value sets	RT11, PF01	1 session

Quality gate:

Hard gate — no exceptions

Metadata accuracy must be >= 85% on the 200-chunk sample. If accuracy is below 85%, all downstream work stops and the corpus must be re-enriched with a better model or prompt before proceeding to v0.3.

Rationale: Building a Synthesis Engine (v0.3.5) or Ontology (v0.6) on top of 60% accurate metadata produces sophisticated-looking but unreliable output. In legal technology, unreliable output is worse than no output.

Estimate: 1 work session

---

v0.3 — Quality & Coverage

Objective: Tests, linting, complete MOCs, documentation — the minimum for the project to be contributable by someone other than the creator.

Features included:

#	Feature	Priority	Status
F25	Create missing MOCs (Tributario, Constitucional, Compliance, Sucessoes)	P1	Planned
F26	Tests for rechunk_v3.py (pytest + fixtures)	P1	Planned
F27	Tests for utility functions	P2	Planned
F28	Complete README	P2	Planned
F31	Makefile (make pipeline, make test, make lint)	P2	Planned
F32	Linting with ruff	P2	Planned
F42	Version and commit enrich_prompt.md	P1	Planned
F19	Complete MOC_CONSUMIDOR	P1	In progress

Acceptance criteria:

• 8/8 MOCs exist as files (even if some are structural placeholders)
• make test passes with at least 1 test for rechunk_v3.py
• make lint passes without errors
• README has sections: Setup, Usage, Architecture, Corpus, Env Vars
• enrich_prompt.md is present in the repository and version-controlled

Prerequisites: v0.2 completed Estimate: 2-3 work sessions

---

v0.3.5 — Doctrine Synthesis (Proposed)

Objective: Transform Douto from a “book search engine” into a “doctrine reasoning engine” that synthesizes multiple authors’ positions on a single legal concept.

Planned Feature — The Doctrine Synthesis Engine is proposed but not yet implemented or approved.

Proposed features:

#	Feature	Description
F43	Synthesis Engine	Given an instituto juridico, collect all chunks, synthesize a multi-author brief
F44	Synthesis prompt	Versioned prompt template for generating Doctrine Briefs
F45	Doctrine Brief template	Structured output format: consensus, divergence, evolution, practical implications

Conditional: Only proceeds if v0.2.5 quality gate passes (metadata accuracy >= 85%).

Prerequisites: v0.3 completed (tests exist), v0.2.5 quality gate passed Estimate: 2-3 work sessions

---

v0.4 — sens.legal Integration

Objective: Douto connects to the ecosystem — Valter can query doctrine programmatically.

Features included:

#	Feature	Priority
F29	Douto -> Valter integration protocol	P1
F30	MCP server for doctrine	P1

Acceptance criteria:

• Valter can query doctrine via a defined protocol (MCP or REST)
• At least 3 MCP tools are functional: search_doutrina, get_chunk, list_areas
• Doctrine search is accessible via Claude Desktop/Code

Prerequisites: v0.2 completed, Decisions D01 and D02 resolved Blocking decisions:

Decision	Question	Options
D01	Integration protocol	MCP stdio / MCP HTTP/SSE / REST / JSON files
D02	Douto as independent service or Valter module	Separate repo / valter/stores/doutrina/ / Proxy through Valter

Estimate: 3-5 work sessions

---

v0.5 — Knowledge Graph & Automation

Objective: Atomic notes, quality evaluation, ingestion automation, and CI/CD.

Features included:

#	Feature	Priority
F36	Automatic atomic notes (1 note per instituto juridico)	P2
F40	Embedding quality eval set (30+ queries)	P2
F41	Unified ingestion CLI (douto ingest livro.pdf)	P3
F39	CI/CD: GitHub Actions with ruff + pytest	P3
F21	Complete knowledge nodes	P2 (in progress)

Acceptance criteria:

• nodes/ contains atomic notes generated from enriched chunks
• Evaluation set with at least 20 queries and expected results exists
• make ingest livro.pdf runs the complete pipeline
• GitHub Actions runs lint + test on PRs

Prerequisites: v0.3 completed (tests exist) Pending decisions: D03 (atomic notes: auto vs. curated), D04 (issue tracking) Estimate: 3-4 work sessions

---

v0.6 — Legal Ontology (Proposed)

Objective: Build a concept graph that maps relationships between legal institutes, creating a structured ontology of Brazilian law.

Planned Feature — The Legal Ontology is a proposed innovation that depends on validated metadata and the synthesis engine.

Proposed features:

#	Feature	Description
F46	Instituto relationship extraction	Map how institutos relate (prerequisite, exception, subtype, etc.)
F47	Ontology graph structure	Data model for the concept graph
F48	Ontology navigation API	Query the graph (e.g., “what are the subtypes of responsabilidade civil?”)

Prerequisites: v0.3.5 completed (synthesis engine), validated metadata Estimate: 5-8 work sessions

---

v1.0 — Integrated Platform

Objective: Douto fully integrated into sens.legal — doctrine as a first-class source alongside case law and legislation.

Features included:

#	Feature	Priority
F33	Doctrine in Neo4j knowledge graph (Valter)	P2
F34	Cross-reference: doctrine <-> case law (STJ)	P2
F35	Cross-reference: doctrine <-> legislation (Leci)	P3
F37	Progressive Briefing support (Juca’s 4 phases)	P2
F38	Docker containerization of pipeline	P3

Acceptance criteria:

• Doctrine nodes in Neo4j with CITA_DOUTRINA and FUNDAMENTA relationships
• Juca briefing includes doctrinal sources automatically
• Pipeline executable in Docker container

Prerequisites: v0.4 completed (MCP functional) Pending decisions: D05 (Neo4j schema) Estimate: Multiple sessions, depends on Valter and Juca evolution

---

Gantt Chart

gantt
    title Douto -- Roadmap
    dateFormat YYYY-MM
    axisFormat %b %Y

    section Foundation
    v0.2 Pipeline Estavel           :a1, 2026-03, 1M
    v0.2.5 Data Validation          :a15, after a1, 0.5M
    v0.3 Quality & Coverage         :a2, after a15, 2M

    section Innovation
    v0.3.5 Doctrine Synthesis       :a25, after a2, 1M

    section Integration
    v0.4 sens.legal Integration     :a3, after a25, 2M
    v0.5 Knowledge Graph            :a4, after a3, 2M

    section Platform
    v0.6 Legal Ontology             :a55, after a4, 2M
    v1.0 Integrated Platform        :a5, after a55, 3M

Milestone Comparison: Original vs. Post-PREMORTEM

The PREMORTEM analysis recommended inserting validation checkpoints before building on unverified foundations:

ORIGINAL:                       POST-PREMORTEM (current):
v0.2 Pipeline Estavel           v0.2 Pipeline Estavel
  |                               |
  v                               v
v0.3 Quality                    v0.2.5 Data Validation  <-- NEW
  |                               |  Quality gate: >= 85%
  v                               v
v0.4 Integration                v0.3 Quality (tests, MOCs, docs)
  |                               |
  v                               v
v0.5 Knowledge Graph            v0.3.5 Doctrine Synthesis  <-- NEW
  |                               |
  v                               v
v1.0 Platform                   v0.4+ (continues as before)

Impact: Adds ~1 work session (v0.2.5) but prevents building innovation features on unreliable data. If the quality gate fails (accuracy < 85%), re-enrichment happens before v0.3, not after.