Where Silo is going

The next 4-6 months are about turning a technically convincing system into a workflow an outside lawyer prefers over an LLM-only baseline. The roadmap below names that sequence, and the kill criteria at the end name the conditions under which Silo stops.

Phases

1. 01

   Foundation

   Deepen the graph where the first wedge is real: corporate and M&A matters, cleaner signals, and more reliable structure at ingestion time.

   Status: in_progress

2. 02

   Momento 360

   Ship the full case loop: diagnosis, assisted argumentation, and side-by-side comparison against a frontier-model-only workflow.

   Status: planned

3. 03

   Validation

   Put the 360 workflow in front of an external lawyer on live matters and measure whether it changes speed, confidence, and output quality.

   Status: planned

4. 04

   Flywheel

   Use real matters to learn where the graph is thin, where the agent stalls, and which additional structure improves the next case.

   Status: planned

5. 05

   Pitch

   Turn the technical and workflow evidence into a compact diligence pack for investors and partners.

   Status: planned

Risks

• External validation does not convert into preference. A technically impressive workflow can still fail if outside lawyers do not prefer it over a frontier-model-only baseline in live matters. Mitigation: run narrow comparative cases early, measure the delta, and tighten the workflow before widening the audience.
• Corpus growth becomes expensive before it becomes useful. The graph can absorb large amounts of legal structure without each added layer changing the quality of real case analysis. Mitigation: keep densification focused on the corporate and M&A wedge, and only deepen the corpus where it improves the live 360 workflow.
• Vendor concentration slows the product. Today Silo depends on external model providers and on MCP clients as the distribution surface. Mitigation: keep the model layer swappable, keep the core value in the structured corpus and tool surface, and avoid product assumptions that require a single client or provider.
• Legal AI adoption tightens before the workflow is proven. A shift in client expectations, bar guidance, or procurement posture could slow usage even if the system works technically. Mitigation: keep the workflow human-reviewed, provenance-first, and narrow enough that each step can be explained in professional terms rather than AI terms.

Kill criteria

• Silo stops if external lawyers do not consistently prefer the 360 workflow to an LLM-only baseline after structured validation on live matters.
• Silo stops if unit economics at the case-analysis layer remain structurally bad even after narrowing scope and keeping corpus growth disciplined.
• Silo stops if frontier models absorb enough of the structural advantage that the graph and tool layer no longer create a material, inspectable delta in legal work.