Core Concepts
Timepoints
A Timepoint is a verified, confidence-scored node in a temporal causal graph. It’s a structured record of a moment — who was there, what they said, why it mattered, what happened next. Every timepoint has a canonical spatiotemporal URL — 8 segments encoding when and where: The 8 segments are, from left to right: year (negative for BCE), month, day, time (24-hour), country, region, city, and slug.SNAG (Social Network Augmented Generation)
SNAG synthesizes and maintains structured social graphs to ground LLM generation in complex group dynamics. Where RAG answers questions from what was written down, SNAG reasons about what people did, felt, and caused. 19 composable mechanisms handle entity states, knowledge provenance, dialog steering, emotional dynamics, relationship tracking, and more. Each mechanism can be enabled/disabled independently.Temporal Modes
Timepoint supports five temporal reasoning modes:Content Layers
The Clockchain stores moments at increasing levels of detail:Edge Types
Moments are connected by typed causal edges:
Each edge tightens the Bayesian prior — fewer plausible things could have happened in the gaps.
Temporal Navigation
Because edges are typed and directed, the graph can be traversed through time, not just across it. Temporal navigation walks multiple hops from an anchor moment toward its causes (past) or consequences (future), or finds the shortest chain of connections linking two moments. Edge direction maps to temporal direction like this:
Symmetric edge types (
contemporaneous, same_location, thematic, and friends) have no temporal direction — they traverse both ways and are only followed when explicitly requested. This powers the graph traversal API, the traverse_moments and find_path MCP tools, and the Temporal Navigator in the web app.
TDF (Timepoint Data Format)
JSON-LD interchange format connecting all services. Every TDF record includes:- id — canonical URL or service UUID
- source — which service produced it (flash, clockchain, proteus, snag-bench)
- provenance — generator, run ID, confidence score
- payload — source-specific content
- tdf_hash — SHA-256 of canonicalized payload for content addressing
Gateway
The API Gateway (api.timepointai.com) is the entry point for all client traffic. It owns authentication (Apple, Google, and GitHub OAuth), JWT signing and verification, credit management, and user accounts. Generation requests are proxied to Flash with an X-User-ID header — Flash itself runs with authentication disabled as a pure generation engine.
MCP (Model Context Protocol)
The Clockchain exposes a Model Context Protocol endpoint atclockchain.timepointai.com/mcp/ (Streamable HTTP, v1.26.0). This allows MCP-compatible clients — Claude Desktop, Cursor, and other LLM tools — to query and traverse the temporal causal graph directly as a tool call.
Conductor
The Conductor is the unified natural-language interface to the whole Timepoint stack. Rather than calling each service’s API directly, you (or an agent) tell the Conductor what you want in plain language — find money, run a simulation, render a moment, explore the graph — and it routes to the right services, runs and monitors long jobs, recovers from failures, and reports the credits it spends as it goes. It is the authenticated home ofapp.timepointai.com and is available over both REST and an MCP converse tool at conductor.timepointai.com/mcp, so a single endpoint can drive the entire suite. Like everything else, it routes through the Gateway, so auth and credit metering stay centralized. See Conductor.
The Flywheel
The system forms a self-reinforcing loop:- Gateway authenticates users and routes generation requests
- Flash renders historical moments into grounded scenes
- Clockchain stores them as graph nodes with causal edges
- Expander (LLM) discovers related moments and grows the graph
- SNAG-Bench scores quality across all outputs
- More data → stronger Bayesian prior → better renderings → more data