Deterministic agentic AI.
Priced in joules. Signed at the edge.
A memoizing cascade with energy-tracked containers. Free models, paid frontier, and three native AIs through one API. Every inference returns a signed receipt with joules consumed and the datacenter that performed the work.
The Cascade
Repeat work is cheap. Novel work is rare.
The cascade learns the difference.
Most AI work is repeat work. The first occurrence is a full inference; every occurrence after that is retrieval. TX Science AI implements the exponential decay function the industry hides. Each inbound request enters at the cheapest gear. Frontier models are the last gear, not the first.
Memoizing Cascade
Every request enters at the lowest gear: deterministic lookup, semantic retrieval, compositional reuse, small model, custom AI, frontier model. Each gear is tried in order. Once work is performed, it is stored. Cost decays toward the deterministic floor as the cascade learns the workload.
Joule + Location Receipts
Every response carries a signed receipt with joules consumed, datacenter location, jurisdiction, and energy mix. First-party measurements from the underlying hyperscaler — not estimates. Auditable for CSRD, SEC climate, and regulated procurement.
Energy-Tracked Containers
Every gear, every tool, every customer agent runs in a container with a declared joule budget and expected consumption profile. Deviation flags potential compromise. Joules are physics — harder to spoof than text-level monitoring. Bad-actor detection comes free with the metering.
Imagine AI
World-model gear for generative simulation. Scenario synthesis, content generation, knowledge-graph-grounded imagination. Causal and structurally consistent. Used when the work needs a coherent world, not a probability distribution.
Semantic AI
Database-outside, search-first, query-based gear. Knowledge lives in the substrate, not in the weights. Update facts without retraining. Fast learning, cheap retrieval, no hallucinated provenance. The right gear when the work is a fact lookup, not a generation.
Math Ground AI
Edge physical AI. Deterministic mathematical computation on-device, sensor- and actuator-bound. Small, fast, grounded in math primitives. Runs where the network does not reach and where determinism matters. mathground.ai
AI is software on hardware. Energy is the cost. Repeat work is most work.
Per-token AI pricing treats every request as a first-occurrence random walk. Real workloads are heavy-tailed: a small set of patterns is asked again and again. Once a pattern has been computed, the answer is a lookup, not an inference. The honest cost curve for a workload is exponential decay toward the deterministic floor.
TX Science AI implements that decay function. The cascade enters at the cheapest gear. Memoization compounds. The customer's cost decreases as their workload is absorbed. The frontier model is the last gear, invoked only when nothing in the cascade can answer.
Every gear runs on the underlying Joule Cloud hyperscaler. Every inference is signed in joules. The cost of intelligence converges to the energy required to compute it.
Work Classes
Declare the work. The cascade picks the gear.
The customer says what kind of work this is. The cascade picks the cheapest gear that can do it. Routing is by work class first, cost tier second. The same API surface handles every class.
Conversation
Open-ended natural-language exchange. Cascade prefers memoized turns and small models; frontier providers used only when the request is genuinely novel.
Fact Lookup
Knowledge retrieval routed first to Semantic AI. Database-outside, source-cited, refreshed without retraining. The cheapest, most reliable gear when the answer exists in a corpus somewhere.
World Simulation
Scenario synthesis, content generation with causal consistency, knowledge-graph-grounded imagination. Routed to Imagine AI when the work needs a coherent world rather than a token distribution.
Edge Physical
Deterministic mathematical computation on-device. Sensor fusion, control loops, embedded inference. Routed to Math Ground AI. Small, fast, runs where the network does not reach.
Agentic Work
Multi-step iterative work with tools. The cascade orchestrates gears across the loop; the path of least resistance is engineered so stochastic gears converge on the correct action. Determinism of the system from gradient engineering of the gear.
Embeddings
Vector representations for retrieval, clustering, and similarity search. Memoized aggressively — the same input produces the same vector, every time.
Code
Code generation, transformation, and review. Cascade prefers structural pattern reuse and deterministic refactoring over re-generation.
Image
Generation, editing, classification. Routed to specialist gears with diffusion or transformer-based image models depending on the task.
Audio
Speech recognition, synthesis, and audio analysis. Cascade routes by latency requirement and language coverage.
API
One endpoint. Energy- and compute-aware. Receipt-backed.
A single primary inference endpoint with rich policy. Declare the work, set the joule cap, pin the jurisdiction, get a signed receipt. Compatibility shims exist for OpenAI, Anthropic, and Gemini SDKs — change the base URL, the cascade benefits flow through.
{
"input": [
{"role": "user", "content": "..."}
],
"work_class": "auto",
"policy": {
"prefer_gear": "lowest_cost",
"max_joules": 1000,
"fallback_chain": [
"semantic_ai",
"frontier_open",
"frontier_paid"
],
"jurisdiction": ["us", "eu"],
"energy_mix_min_renewable": 0.7,
"deterministic_only": false,
"container_profile": "standard"
},
"tools": [],
"memory_ref": "session_xyz",
"stream": false
}{
"id": "inf_01HX...",
"output": [...],
"cascade": {
"gear": "semantic_lut",
"trace": ["lut_miss", "semantic_hit"],
"memoized": true,
"decay_factor": 0.42
},
"joules": {
"consumed": 12.4,
"predicted": 14.0,
"deterministic_floor": 0.8,
"saved_vs_frontier": 89.2
},
"location": {
"datacenter": "joule-cloud-us-west-2",
"jurisdiction": "us",
"energy_mix": {
"renewable": 0.84,
"g_co2_per_kwh": 110
}
},
"provenance": {
"model": "semantic-ai-v1",
"data_sources": [...],
"signed_receipt": "<JWP>"
}
}Primary Inference
One endpoint for every work class. Streaming via SSE with cascade updates and joule ticks.
Joule Prediction
Get a predicted joule cost before committing. Set hard caps based on the prediction.
Embeddings
Vector representations. Aggressively memoized — same input, same vector, signed receipt.
Rerank
Relevance scoring as a first-class operation. Cheap, deterministic where possible.
Cascade Status
Live gear distribution, decay metrics, memoization hit rate for the caller's workload.
Signed Receipts
Fetch the JWP receipt for any past inference. Auditable, exportable to compliance systems.
Compatibility
Existing SDKs work unchanged through compatibility shims at /compat/openai, /compat/anthropic, and /compat/gemini. Change the base URL; cascade memoization, joule receipts, and energy-tracked containers flow through invisibly. The native contract exposes the full surface — receipts, cascade trace, joule prediction, work-class routing.
Family
Seventeen siblings. One architecture.
TX Science AI runs on Joule Cloud and ships joule receipts in the same JWP format every Transaction Science platform uses. Every sibling can compose with the cascade for inference; every receipt is auditable across the family.
TradingOS
Matching engine, pre-trade risk, settlement, surveillance.
ComplianceOS
Continuous compliance across 50+ frameworks and 195 jurisdictions.
Veritas
Tax computation with IRC citation provenance.
LegalOS
Global legal corpus structured as a programmable system.
InformationOS
Library science for the AI era. Knowledge, provenance, citations.
ExpertOS
Measurement as resolution. The distance function as authority.
TrustOS
Identity substrate. WebAuthn, OAuth, SAML, DID, post-quantum crypto.
CyberSecurityOS
Detection, response, threat intelligence. Every anomaly a distance.
Insights
Energy-aware cloud intelligence and CSRD carbon reporting.
InsuranceOS
Underwriting, claims, actuarial science, and reinsurance.
TerraOS
Property search, financing, closing, and tokenization.
DestinationOS
DMO as a service — inventory, arbitrage, visitor routing.
EducationOS
The global university and school system as infrastructure.
MedicineOS
Global medical curriculum encoded for the AI doctor economy.
Settlement
Stocks, crypto, and cash in one consumer account.
CommunicationOS
Multi-channel messaging, voice, video, and PBX with signed receipts.
TX Science TV
24/7 live broadcast across every Transaction Science vertical.