Cross-era prediction simulator
Prime Radiant
Most public reasoning about the long-term future fails for the same reason: it tries to forecast events without first modeling the machinery that produces events. Prime Radiant is an attempt to build that capacity — a retrieval-augmented reasoning stack over a hand-curated catalog of civilizational machines, capable of predicting which new institutions will emerge from the current substrate.
The system learns the Modernity Machine’s transition into the Divergence Machine, then applies that pattern forward: given the DM machines in Day now, what Liveness Machine forms should we expect in Dawn?
- Cards on main
- 182 / ~360
- Status
- Phase 1
- Era span
- MM → DM → LM
- Method
- RAG + primitives
The operational frame
Three world-machine eras
Civilizational machinery clusters into three overlapping operational logics, each running its own Dawn → Day → Dusk lifecycle. In 2026 all three are simultaneously active.
Modernity Machine
~1200–2150
Currently Dusk
Legibility, fiat progress, narrative coherence, bureaucratic states, mass institutions. The dominant logic of the last several centuries — still operational but losing the conditions that originally made it work.
Divergence Machine
~1650–2400
Currently Day
Pluralism, push, distributed information, post-narrative coordination, network-effect capture. Fully turned on as of roughly 2015 — the platform era, the open-source era, the surveillance-capital era.
Liveness Machine
~2000–2300 (speculative)
Currently Dawn
Self-organized criticality, recovered embodiment, generative cohabitation with AI, post-capture coordination forms. Barely visible, mostly extrapolated — the class of institutions the substrate is reorganizing to produce.
System architecture
How it works
A retrieval-augmented reasoning stack: research findings pin the schema; Sonnet generates Machine Cards; DuckDB + VSS stores and indexes them; HippoRAG-lite retrieves relevant subgraphs; Gemma 3 4B reasons inline with typed primitive tokens.
Conceptual core
Thinking with Machine Primitives
Analogous to how visual reasoning models cite spatial bounding boxes inline while processing images, Prime Radiant’s reasoning model cites civilizational machines inline while it reasons. Seven typed primitive tokens define the grammar.
| Token | Purpose | Example |
|---|---|---|
<|machine|> |
Cite an existing machine card | <|machine|>JSC-MERCANTILE<|/machine|> |
<|coupling|> |
Assert a relation between two machines | <|coupling|>VOC→DUTCH-STATE:supplies<|/coupling|> |
<|signal|> |
Cite a measurable trace from a machine | <|signal|>opp_strength@AWS:0.82@2024<|/signal|> |
<|stress|> |
Name a pressure on a machine | <|stress|>multicloud_threat@AWS<|/stress|> |
<|transition|> |
Assert a regime change between machines | <|transition|>MM-DAY→MM-DUSK@2000<|/transition|> |
<|trace|> |
Assert a phase / cross-era trajectory | <|trace|>JSC:mm-day@1700→mm-dusk@1850→..<|/trace|> |
<|new_machine|> |
Propose a machine stub not yet in the graph | <|new_machine|>{...JSON stub...}<|/new_machine|> |
Project phases
Build plan
Six sequential phases, gated on artifact quality. Phases −1 and 0 are complete. Phase 1 is in progress at 182 cards.
Phase −1
Research
Ten background research waves across ontology, causal substrates, cyclical dynamics, world-systems, network theory, and the machine atlas. 23 findings merged into a consolidated synthesis.
Phase 0
Schema
Machine Card schema v0.1: 50+ fields, closed enums, phase-snapshot trajectories, lineage links, cross-era coupling typology. Three hand-authored reference cards validated every field.
Phase 1
Machine Cards
Hand-curated corpus of ~360 civilizational machine classes, spanning MM, DM, and LM eras with phase-snapshot trajectories and cross-era couplings. 182 cards on main; Batches 1 and 2 complete.
Phase 2
Storage & Indexing
DuckDB + VSS with three HNSW indices over multi-view embeddings: semantic (768d), state (~50d), graph (~64d). Enables blended and staged retrieval modes.
Phase 3
Retrieval
HippoRAG-lite: vector seed over the three embedding spaces, expanded via PageRank over the NetworkX coupling graph. Different retrieval modes suit different prediction tasks.
Phase 4
Reasoning
Gemma 3 4B with the seven-token primitive grammar. Autoresearch loop (prompt-variant search, Opus judge) optimizes against four prediction tasks: generate stubs, score plausibility, classify era, predict transition timing.
Phase 5
Validation
Cross-era smoke tests: predict LM-Dawn stubs from DM-Day input sets; classify 10 cards on the live/zombie/ghost/necromancy spectrum; run a Calvino-layer narrative constrained by the retrieved subgraph.
What is a Machine Card?
Two examples from the corpus
Each card is a class description of a civilizational machine — not an instance ("the VOC") but a type ("the joint-stock company"). Cards carry phase-snapshot trajectories, typed state variables, cross-era coupling declarations, and per-field provenance flags.