Individual observations capture point-in-time snapshots. But the most useful knowledge often emerges from synthesizing across multiple observations to derive higher-order facts that weren’t explicitly stated in any single source.
Consider three source facts:
- “Paula works at Google” (validAt: 2020-01-15)
- “Paula is a Senior Engineer at Google” (validAt: 2022-06-01)
- “Paula joined Microsoft as Principal Engineer” (validAt: 2024-03-15)
From these, you can synthesize:
- “Paula worked at Google from January 2020 to March 2024” (duration fact)
- “Paula was promoted to Senior Engineer at Google in June 2022” (progression fact)
These synthesized facts have different properties than directly observed facts:
- They derive meaning from temporal relationships between observations
- They capture patterns (career progression, organizational changes) that no single document states
- They make implicit timelines explicit and queryable
- They enable reasoning about duration, causality, and evolution
Types of Synthesized Facts
Timeline synthesis: Combining hire and departure facts to understand employment duration, project timelines, or configuration lifespans.
Corroboration synthesis: When multiple sources assert the same fact, confidence increases. A fact mentioned in an email, confirmed in a meeting, and recorded in CRM becomes more reliable than any single observation.
Pattern synthesis: Recognizing that similar decisions made under similar conditions create precedent. “Whenever healthcare companies face procurement delays, we extend payment terms.”
Causal synthesis: Linking events that occurred in sequence to infer relationships. “The deployment on June 15 preceded the P1 incident on June 16; deployments to this service are high-risk.”
Why Synthesis Is Hard
Determining what’s true from historical assertions requires judgment:
- Reconciling conflicts: Two sources claim different facts—which is authoritative?
- Identifying supersession: A new fact makes an old fact invalid—when did the transition occur?
- Inferring gaps: Between “joined company” and “left company” is a gap—can we infer continuous employment?
- Maintaining evidence chains: Each synthesized fact must trace back to source facts, which trace to original content
This is where LLM-powered resolution becomes essential. The model can:
- Cluster similar facts
- Identify supersession relationships
- Synthesize timeline facts from scattered observations
- Judge confidence levels based on source quality and corroboration
Synthesized Facts Enable Simulation
Context graphs function as organizational world models that enable simulation depends on synthesized knowledge. Point-in-time observations alone can’t answer:
- “How long do typical deployments take?” (requires duration synthesis)
- “What’s the blast radius of changes to this service?” (requires causal synthesis)
- “Is this precedent still applicable?” (requires temporal synthesis with policy evolution)
Simulation requires understanding not just what happened, but patterns in how things happen
Status Hierarchy
Facts carry status indicating their provenance:
- Directly observed: Extracted from a single source document
- Corroborated: Same fact appears in multiple sources
- Superseded: A newer fact has replaced this one
- Synthesized: Derived from multiple observations
- Canonical: Currently accepted as ground truth
This hierarchy preserves epistemic uncertainty. An agent reasoning over facts can weight them appropriately: corroborated canonical facts are more reliable than superseded observations.
The Evidence Chain
Critically, synthesized facts maintain audit trails:
- Each synthesized fact points back to source facts
- Source facts point to original content
- The reasoning for synthesis is captured
This enables:
- Debugging why an agent reached a conclusion
- Validating that synthesized knowledge reflects reality
- Updating synthesis when source facts change
- Explaining decisions to stakeholders
Without evidence chains, synthesis becomes ungrounded—the system “hallucinates” facts without being able to justify them.
Relation to Entity Resolution
Synthesis depends on Enterprise context requires resolving entities across disparate systems. You can’t synthesize Paula’s employment timeline without first knowing that “Paula,” “P. Chen,” and “paula@google.com” refer to the same person across different sources.
Entity resolution must happen before fact synthesis. The progression is:
- Observe raw content
- Resolve entities (people, organizations, products)
- Extract facts with entity mentions
- Synthesize higher-order facts across resolved entities
- Enable agents to reason over synthesized knowledge