Fusion

Multi-model decision infrastructure for production AI.

Route one request through a panel of models, compare their reasoning, synthesize the best answer, enforce provider and data policy, and settle the full cost through Aurona credits.

OpenAI-compatible

Token ledger

GPU route layer

aurona/fusion

multi-model route

panel + judge

Panel models

3-7

Judge output

Structured

Fallback

Automatic

Classify

Detect task type, risk level, required modality, region, and budget before any model call.

Panel

Run approved frontier, value, open-weight, or private GPU models in parallel.

Judge

Compare consensus, contradictions, evidence gaps, unique insights, and confidence.

Synthesize

Return one final answer, route trace, token cost, provider health, and fallback notes.

Fusion control layer

Multi-model quality needs visible controls.

Fusion should expose the knobs developers actually need: panel membership, judge behavior, provider routing, credit ceiling, fallback policy, and trace output.

Panel control

3-7 models

Choose frontier, value, code, multimodal, or private GPU models for each Fusion run.

Judge schema

Structured

Return consensus, contradictions, confidence, coverage gaps, unique insights, and final recommendation.

Provider policy

sort + order

Tune provider selection by price, throughput, latency, fallback permissions, and data collection policy.

Endpoint policy

queue + cache

Route panels through provider pools, serverless endpoints, or dedicated lanes using queue depth, warm-cache, and quota signals.

Cache affinity

traceable reuse

Keep repeated system prompts, tool schemas, and agent sessions on a compatible warm route when policy and health still match.

Credit guardrail

budget ceiling

Set max tokens, max credit, trace depth, and fail-closed rules before a panel starts.

Route modes

Call the outcome, not a single model.

Model choice and provider routing are only the base layer. Aurona Fusion turns them into a full decision router with panel models, judge models, private supply, and token settlement.

Route

Primary signal

Best for

Execution

aurona/fusion

multi-model deliberation

research, agents, high-risk answers

panel + judge + final

aurona/fusion-fast

small panel, low latency

chat, copilots, realtime UX

2 panel models + fast judge

aurona/fusion-code

coding model panel

repository tasks, PR review, migration

code score + tool use

aurona/fusion-private

approved private supply

enterprise data, regulated workloads

ZDR + region + GPU lane

aurona/fusion-endpoint

serverless or dedicated endpoint panel

app bursts, evals, private tenants

endpoint health + judge trace

aurona/auto

quality, price, latency, health

default production traffic

single best route

aurona/exact

pinned model and provider

stable behavior, audits, certification

no automatic substitution

Fusion simulator

Run a real front-end route preview

Choose a route preset and run it. The panel, chart, and trace update immediately.

quality

cost control

latency

Trace #18

aurona/fusion ran a panel, checked zero-retention policy, wrote token ledger events, and returned a judge confidence score of 96.

classify

task, risk, modality, budget

panel

4 approved model calls

judge

96% confidence

ledger

1.73 credit ceiling

{
  "model": "aurona/fusion",
  "panel_size": 4,
  "judge": "aurona/judge-premium",
  "provider": {
    "sort": "throughput",
    "allow_fallbacks": true,
    "cost_quality_tradeoff": 7
  },
  "endpoint": {
    "type": "serverless",
    "health_policy": ["queue_depth", "warm_cache", "saturation"]
  },
  "policy": "zero_retention_optional",
  "session_id": "fusion-preview-18",
  "observability": { "destination": "route-trace-stream" },
  "budget": { "max_credit": "1.73" }
}

Auto Route decision lab

simulated planning data

See why a route won before traffic runs

Classify the workload, score only approved routes, choose a cost-quality objective, and return a compact receipt with fallback and capacity evidence.

Workload prompt

Objective

Allowed pool

Decision receipt #2048

aurona/code-balanced

fallback · aurona/private-code

workload

Code

route score

109

credit estimate

0.048

capacity

Shared + burst GPU

Code workload detected before generation.

Balanced objective scored quality, latency, and simulated credit exposure.

5 routes remained after the all approved pool check.

Shared + burst GPU matched the selected route's current planning profile.

{
  "simulated": true,
  "objective": "Balanced",
  "scope": "All approved",
  "task": "Code",
  "poolSize": 5,
  "selectedRoute": "aurona/code-balanced",
  "fallbackRoute": "aurona/private-code",
  "capacity": "Shared + burst GPU",
  "routeScore": 109,
  "estimatedCredits": "0.048"
}

Session continuity lab

simulated planning data

Keep context stable without hiding route changes

Preview whether a multi-turn session can reuse its route, should switch models or GPU lanes, or must pause for policy and AI Token review. This lab does not pin production traffic.

Next task

Route health

Workspace policy

AI Token gate

GPU capacity preference

Continuity receipt #731

reuse route

aurona/code-balanced

prior route

aurona/code-balanced

next task

Code

capacity

Serverless approved pool

Code remained compatible with the prior session route.

Route health remained eligible for continuity planning.

Workspace policy remained unchanged and eligible.

The AI Token budget gate remained open.

Tool Route Evidence Lab

simulated · planning only

Rank tool-capable endpoints before a route runs

Check the required tool schema and capacity scope first, then compare simulated validity, evaluation, throughput, and AI Token evidence with a visible fallback.

Workload

Objective

Capacity scope

Preview #417

No traffic, billing, or capacity change

{
  "route": "aurona/agent-balanced",
  "workload": "Support actions",
  "tools": {
    "schema_contract": "strict",
    "parallel": true
  },
  "objective": "balanced",
  "capacity_scope": "workspace-approved",
  "mode": "preview"
}

Selected endpoint

aurona/endpoint-cobalt

EvidencestrongSchema valid99%Throughput index92CapacityDedicated GPU

Reviewed fallback

aurona/endpoint-cedar

EvidencestrongSchema valid99%Throughput index76CapacityServerless

Endpoint	Score	Schema valid	Eval confidence	Throughput	AI Tokens
aurona/endpoint-cobalt	91.2	99%	95%	92	0.038
aurona/endpoint-cedar	87.9	99%	92%	76	0.052
aurona/endpoint-onyx	87.8	98%	90%	88	0.044
aurona/endpoint-amber	86.9	97%	84%	112	0.031

strict tool schema and parallel execution checked

4 approved endpoints remained after capability and capacity policy

balanced evidence ranked schema validity, evaluation confidence, throughput, and AI Token exposure

aurona/endpoint-cobalt selected with aurona/endpoint-cedar retained as the reviewed fallback

Panel presets

Fusion should feel configurable, not mysterious.

Each preset defines which models join the panel, which judge evaluates them, what budget limits apply, and how much route trace should be returned.

Frontier Research

quality first

GPT-5.5 Pro, Claude Opus 4.8, Gemini 3.1 Pro Preview

Premium reasoning judge

Best for source-heavy analysis, executive decisions, strategy memos, and answers where disagreement matters.

Code Agent

tests + reasoning

Claude Opus 4.8, Kimi K2.7 Code, DeepSeek R1, Qwen3.7 Plus

Code-aware judge

Best for complex coding, migrations, debugging, architecture review, and tool-heavy developer agents.

Market Intelligence

recency + coverage

GPT-5.5 Pro, Grok 4.3, Gemini 3.1 Pro Preview

Evidence judge

Best for fresh market reads, competitive analysis, policy monitoring, and multi-source synthesis.

Private GPU Panel

ZDR + capacity

Llama 4 Maverick, Nemotron Ultra, DeepSeek V4 Pro

Private enterprise judge

Best for customers that need controlled data paths, predictable throughput, and dedicated inference capacity.

Endpoint Panel

queue + cache

DeepSeek V4 Flash, Llama 4 Maverick, Qwen3.7 Plus

Endpoint-aware judge

Best for apps that need burstable serverless capacity first, then a dedicated lane after traffic becomes predictable.

API configuration

Make advanced routing obvious to developers.

A Fusion request should expose the important controls directly: panel models, judge behavior, fallback policy, provider sorting, zero-retention requirements, and credit budget.

Fusion request

{
  "model": "aurona/fusion",
  "messages": [{ "role": "user", "content": "Evaluate this acquisition memo." }],
  "fusion": {
    "panel": ["openai/gpt-5.5-pro", "anthropic/claude-sonnet-4.6", "google/gemini-3.5-flash"],
    "judge": "aurona/judge-premium",
    "return_analysis": true,
    "budget": { "max_tokens": 180000, "max_credit": "12.00" }
  },
  "provider": {
    "sort": "throughput",
    "allow_fallbacks": true,
    "cost_quality_tradeoff": 7,
    "cache_affinity": "best_effort",
    "cache_key": "acquisition-memo-tools-v2",
    "data_collection": "deny",
    "require_parameters": true
  },
  "endpoint": {
    "type": "serverless",
    "health_policy": ["queue_depth", "warm_cache", "quota"]
  },
  "session_id": "acquisition-memo-review",
  "observability": { "destination": "workspace-usage-warehouse" }
}

Panel selection

Choose a fixed panel, a cost-aware panel, or an adaptive panel based on task risk, context size, modality, and customer tier.

Provider routing

Sort by price, latency, throughput, or provider order while respecting region, retention, parameter support, and vendor approval.

Fallback chains

Define exactly when Fusion may retry, downgrade, switch providers, use private GPU capacity, or fail closed for regulated traffic.

Structured judge

Ask the judge to return consensus, contradictions, confidence, evidence gaps, policy flags, and recommended next actions.

Token settlement

Track model tokens, judge tokens, final tokens, orchestration fees, and customer credits in one route ledger.

Enterprise trace

Return model IDs, providers, route decisions, fallback reasons, policy checks, and cost events for audit and procurement teams.

Judge analysis

The best part of Fusion is the comparison layer.

Instead of hiding the panel, Aurona can expose a structured analysis object that lets developers inspect quality, disagreement, confidence, and cost.

Field

What it means

Product use

Consensus

Where panel models agree

Returned as the safest shared answer

Contradictions

Where models disagree

Flagged before final synthesis

Coverage gaps

What the panel missed

Used to trigger retry or retrieval

Unique insights

Distinct high-value observations

Merged into the final answer

Confidence

Evidence, model agreement, and policy fit

Shown in the route trace

Cost trace

Input, output, judge, and final tokens

Settled through Aurona credits

Provider controls

Use provider controls, then add enterprise policy.

Fusion should support provider sorting, fallbacks, price ceilings, parameter requirements, and data collection controls while also adding Aurona's region, vendor approval, and GPU capacity logic.

Control

Example value

Why it matters

sort

price, throughput, latency

Choose the cheapest, fastest, or highest-throughput healthy provider

order

approved provider list

Prefer enterprise-contracted providers before public fallback

allow_fallbacks

true or false

Control whether Fusion can substitute when a provider fails

cost_quality_tradeoff

0-10

Tune the route between cost control and answer quality

session_id

string

Keep related agent turns pinned to a compatible provider when cache and policy allow

cache_affinity

off, best_effort, required

Prefer a compatible warm provider without bypassing policy, price, or health gates

cache_key

string

Group reusable system prompts, tool schemas, and agent turns into one traceable cache scope

max_price

per-token ceiling

Block accidental premium spend on large or repeated calls

data_collection

allow or deny

Respect zero-retention and sensitive-data policy

endpoint.type

provider_pool, serverless, dedicated

Choose which capacity class can run a panel member

endpoint.health_policy

queue_depth, warm_cache, quota

Avoid cold or saturated endpoints during panel execution

require_parameters

true or false

Only route to providers that support required parameters

Token economics

Fusion is where Aurona's token layer becomes valuable.

A multi-model route creates a new commercial object: panel cost, judge cost, final synthesis, orchestration fee, and GPU capacity can all be priced, settled, and optimized.

Cost layer

Measured from

Business value

Panel calls

Input and output tokens for each selected model

Makes multi-model quality measurable

Cached input

Provider-reported cache reads and writes

Separates reusable context from fresh prompt input in the credit trace

Judge pass

Structured comparison tokens

Turns disagreement into product signal

Final synthesis

One final model call or deterministic merge

Keeps the user experience simple

Fusion fee

Aurona route orchestration and settlement

Creates the AI Token revenue layer

GPU lane

Dedicated or partner inference capacity

Improves margin once volume is predictable

Aurona.ai

Build on the token network for AI applications, models, and compute.

API

OpenAI-compatible

Billing

AI Token credits

Capacity

GPU-backed routes

Start building Read docs