What is running today

This chapter reports what is live, not what is planned. Every capability below has been in production long enough to have a ship date, a source, and — where it matters — a recorded trace you can follow end to end.

What's already in production

• Knowledge graph at production scale. 86,522 nodes and 133,404 relationships across STJ, TJPR, and TJSP, running on Neo4j Aura Professional. Verified live April 2026.
• MCP remote server with OAuth 2.0 and 45 tools. Silo runs a production Model Context Protocol server with OAuth 2.0 authorization, tier enforcement (free/pro), and daily rate limiting. Any MCP-capable client — Claude, ChatGPT, or a custom agent — can use Silo today by adding the remote endpoint as a custom tool. Shipped April 4, 2026.
• Autonomous case analysis pipeline. The case analysis agent ingests raw legal PDFs, extracts structural signals through document intelligence, queries the graph, and produces a structured case dossier with provenance at every step. Real cases under production tracing today, with structured iteration on each run. Latest version shipped April 8, 2026.
• Document intelligence end-to-end. Raw Brazilian legal PDFs — text, scanned, .doc, .docx — pass through per-page quality detection, Portuguese OCR, deterministic semantic chunking over nineteen legal landmark patterns, and LLM-based extraction of claims, arguments, rulings, and citations. Semantic chunker and per-page OCR shipped April 4, 2026.
• Graph analytics REST surface. 57 endpoints covering case similarity, citation chains, minister profiles, argument divergence, gap detection, contested criteria, evidence strength, and PageRank. Redis-cached with graceful degradation. Most recent additions shipped April 2, 2026.

A recorded trace

Two pipelines, one query. This is a multi-PDF case analysis run from March 2026 — a Brazilian appellate case submitted to the production pipeline as three document sources totaling 837 pages and roughly 964,000 characters. Behind the trace sit two distinct pipelines. The batch ingestion pipeline built the graph this case was analyzed against. The case analysis pipeline runs per case.

Batch ingestion (F0→F6). The knowledge graph is built by a six-phase pipeline that runs offline over curated batches of decisions: F0 ingests and validates the raw corpus, F1 segments, F2-F4 run LLM extraction for facts, criteria, defensive theses, and dispositivos, F5 canonicalizes and embeds, and F6 writes the final node-and-edge load into Neo4j Aura. This is what makes the 86,522 nodes and 133,404 relationships exist in the first place.

Case analysis, Phase 1 — structuring. The three PDFs pass through the document intelligence layer — per-page quality detection, OCR where needed, and a deterministic regex chunker over nineteen legal landmark patterns — producing 54 chunks. Each chunk is then run through Qwen3.5-flash in a three-pass extraction that pulls claims, arguments, facts, rulings, and cited proceedings. (An open-source alternative, GPT-OSS 120B via OpenRouter, is currently under A/B evaluation as a candidate replacement, gated by automated prompt-versioning metrics.) Phase 1 deliberately uses open-source models rather than frontier models because the per-chunk cost has to stay linear as cases grow.

Case analysis, Phase 2 — enrichment. Sonnet 4.6 runs in an agentic loop, calling Valter's MCP tools to enrich the structured output with precedents, citation chains, minister profiles, and similar cases from the graph.

Case analysis, Phase 3 — reasoning. Sonnet 4.6 again, still in the agentic loop, now reasoning over the enriched output to produce the final case dossier.

Totals. 1.73 million input tokens. 23,000 output tokens. 32 minutes 51 seconds wall time. US$5.58. Every chunk, every tool call, every token count is preserved in the canonical run directory for audit.

What is not yet production

• Marketplace distribution. Silo's MCP server runs, but pre-built integrations with Claude and ChatGPT are not yet live. Applications to both directories are in review. Today, any user still has to manually add the remote MCP endpoint as a custom tool in their client.
• Legislative grounding and doctrine layers. Two of the five architecture layers — the canonical resolution of statutes and provisions, and the structured ingestion of legal doctrine — are functional in isolation but not yet integrated into the case analysis pipeline. The current dossier output reasons primarily over case law.
• TRF4 coverage. The federal court TRF4 has 1.43 million decisions already classified in the corpus pool, but only a token load (roughly twenty nodes) has been ingested into the graph so far. Expanding TRF4 coverage is part of the next ingestion wave.
• Case analysis stability. The pipeline is functional end to end on real production cases, but it has not yet been stress-tested at volume — concurrent jobs, long-tail PDFs, or cases outside the initial tribunal mix have not been exercised.
• Prompt evaluation framework. A lightweight A/B harness is in use today for iterating on the structuring prompt. The full evaluation framework — automated rubric, gold standard corpus, regression gates across more than one dimension — is in design and not yet built.