TIMEPOINT PRO
Open Source
Apache 2.0
Social Network Augmented Generation (SNAG)
The problem: LLMs collapse when simulating more than ~10 characters over ~20 interactions. Context explodes, personalities blur, causality drifts.
The solution: SNAG. You know RAG — retrieve documents to ground generation. SNAG retrieves and maintains a structured social graph: relationships, secrets, power dynamics, emotional states, knowledge flow, and memory. Every agent maintains lifelong internal state — beliefs, goals, secrets, emotional valence — that survives across hundreds of simulated time-steps. Conversations reference events from 40 simulated days ago with perfect recall.
20 pluggable behavioral mechanisms (M1–M20) — reciprocity, status-seeking, temporal discounting, coalition formation, narrative self-consistency — each versioned, testable, and documented. Swap or weight them at runtime: run a high-trust society or a zero-sum cutthroat one. Scales to dozens of entities across hundreds of timepoints. $0.15–$1.00 per run.
Training data and Rendered Futures are the killer apps. Every run produces richly structured JSONL/SQLite in the Timepoint Data Format (TDF) — the unified interchange format across Flash, Pro, Clockchain, and SNAG Bench — with full causal ancestry, provenance chains, counterfactual branches, quantitative entity states, and ADPRS shadow reports. Purpose-built for fine-tuning on causal reasoning, temporal consistency, and multi-agent dynamics. The synthetic data pipeline is the moat.
Every simulation run is also an appreciating data asset. TDF output includes full causal ancestry, provenance chains, counterfactual branches, and quantitative entity states — purpose-built for fine-tuning causal reasoning, training roleplay models, and as an end-state goal, conditioning diffusion models on temporal causal graphs.
- FORWARD — strict forward causality
- PORTAL — backward-reason from any target outcome to map the critical path
- BRANCHING — counterfactual "what if" analysis
- CYCLICAL — future constrains past (generational sagas)
- DIRECTORIAL — dramatic tension drives events
PORTAL mode stands out: reason backward from any target outcome to identify the minimum set of present conditions required for that future. This isn't forecasting — it's causal decomposition. What must be true today for X to happen tomorrow?
Rendered Futures
Every Pro simulation is a Rendered Future — a causal subgraph stored in TDF, scored for convergence, and eligible for ingestion into the Clockchain. As reality unfolds and validates predicted paths, the rendering appreciates: it becomes training data, grounding data, and a stronger Bayesian prior for all future simulations. The more we render, the better we predict. The better we predict, the more valuable the graph.
In SNAG Bench
Pro templates are the test harness. Run the same scenario with different models. Score entity consistency, causal validity, knowledge boundaries, voice distinctiveness, mechanism activation coverage, graph coverage quality, and convergence across repeated runs. Pro ships with built-in SNAG Bench compatibility — its quality gates are the rubric.
5 temporal modes · 20 mechanisms (M1–M20) · 21 verified templates · $0.15–$1.00/run
github.com/timepoint-ai/timepoint-pro
Run Pro simulations in the Lab →
