R1
Additional intelligence has to be captured
Another source may expand available capability, but only if its marginal information can be exposed, evaluated, and preserved.
IQ256 — AI Fusion Research
We believe the future of high-value AI is Fusion.
Fusion explores how different forms of intelligence — models, tools, research, and people — can be selected, given the right context, and combined for problems where additional reasoning effort is worth its cost.
multi-source
Models, tools, research, and people where relevant
adaptive allocation
Intelligence added where it can change the result
internal use
Running today in selected workflows
Current status — as of 3 September 2026
- IQ256 uses Fusion internally, in selected analysis and decision-support workflows.
- External requests for scoped work are handled case by case through iq256.com.
- There is no public API and no public access to the technology.
- No IQ256 benchmark or customer result has been published yet.
- Published IQ256 evidence is being developed.
External field evidence — not IQ256 performancethird-party sources
Published result
Mixture-of-Agents reports 65.1 vs 57.5 LC win rate on AlpacaEval 2.0 (as reported, 2024).
In production
OpenRouter ships Fusion as a documented production feature, not a research demo.
Active research field
Test-time compute, debate, and multi-agent composition are actively researched, with mixed and setting-specific results.
These are third-party results, cited as evidence that the field is real. None of them is an IQ256 result, and none describes IQ256's architecture.
Early-stage context
IQ256.ai is the research and engineering annex of IQ256; iq256.com remains the primary service. This is a founder-led effort. Internal evaluations are encouraging but unpublished, so we make no public performance claim from them. Non-confidential technical discussion and investor conversations are welcome through iq256.com.
01Fusion
Fusion is already practical. The design space is still being mapped.
Multi-model and multi-agent systems have demonstrated useful results in research and in production: cross-checking, debate, decomposition, tool use, and verification stages that catch what a single pass does not.
Those results validate the field, not one universal recipe. Solo AI is sometimes correct and sufficient, and then it is the right answer. The open engineering problem is how to allocate intelligence under cost, latency, reliability, context, and objective constraints — and when to stop.
External field evidence — not IQ256 performancepublished third-party benchmark
Mixture-of-Agents65.1
GPT-4o (as reported)57.5
0AlpacaEval 2.0 — LC win rate (%)80
Published third-party benchmark (Mixture-of-Agents, 2024). Not an IQ256 result: it shows only that selected multi-model architectures can outperform strong Solo AI in specific evaluations.
Vendor / platform observation — not IQ256 performance
“use it when the cost of being wrong outweighs the cost of a few extra completions.”
Fusion is a documented production feature elsewhere in the industry. IQ256 does not imply a shared architecture.
Maturity of the fieldworking assessment
| Area | Status | What is still unclear |
|---|---|---|
| Cross-model verification | Established in practice | Which claims are worth the verification budget. |
| Decomposition into sub-problems | Established in practice | How decomposition interacts with synthesis quality. |
| Debate and structured critique | Demonstrated in research | When critique adds signal rather than confident noise. |
| Inference-time compute allocation | Actively researched | How to predict, ahead of time, where compute pays. |
| General rules for combining intelligence | Open | No general theory; results remain setting-specific. |
Cross-model verification
- Status
- Established in practice
- What is still unclear
- Which claims are worth the verification budget.
Decomposition into sub-problems
- Status
- Established in practice
- What is still unclear
- How decomposition interacts with synthesis quality.
Debate and structured critique
- Status
- Demonstrated in research
- What is still unclear
- When critique adds signal rather than confident noise.
Inference-time compute allocation
- Status
- Actively researched
- What is still unclear
- How to predict, ahead of time, where compute pays.
General rules for combining intelligence
- Status
- Open
- What is still unclear
- No general theory; results remain setting-specific.
A weaker source can still be the right source for one subproblem.
A cheaper model, a tool, an external expert, or an employee of the customer may be globally weaker or much narrower than a frontier model, and still hold the information a specific subproblem depends on.
So the question is not “which source scores highest on average?” but “what information does this source contribute that is not already available, and what does it cost to obtain it?”
Fusion is not about averaging intelligence. It is about capturing complementary capability.
More intelligence may be available. Capturing it is the hard part.
noteThe implication is architectural: additional reasoning sources are useful when the system can preserve their independent signal and capture the information they add.
02Technology
Fusion is a spectrum, not a single architecture.
A simple Fusion implementation may look like this: run several models, aggregate their outputs, return an answer. That is a valid implementation, and for some problems the simplest useful configuration is a single model.
Both are points in a much larger design space. The decisions that matter — which source, for which subproblem, with which context, in which sequence, and when to stop — are made before and around any aggregation step.
Public system modelconceptual
01input + constraints
Objective, budget, latency, risk, and what evidence is available.
02decomposition
The problem is split into subproblems with different requirements.
03source selection
Models, tools, research, or humans — chosen per subproblem.
04context allocation
Each source receives only the context that changes its answer.
05adaptive execution
Planned branches may be dropped; findings can open or close others.
06verification
Claims that matter are contested before they are accepted.
07synthesis
Partial results become one coherent position.
08output design
Form is designed for the decision, not merely formatted.
09outcome observation
What happened afterwards informs later orchestration research.
The stages are not a fixed pipeline: input itself can be improved, branches can be abandoned, and one discovery can create or close another subproblem. Routing rules, thresholds, and participant counts are deliberately not shown.
A simple implementationvalid, one point
Model A + Model B + Model C → aggregate → answer. Easy to describe, easy to operate, and bounded in what it can improve.
lab noteAggregation assumes the useful information is already present in the outputs. Much of the interesting design work happens before aggregation, not at it.
Dimensions of the design spacenon-exhaustive
01objective and constraints
Budget, latency, risk tolerance, and what the answer is for.
02input and evidence acquisition
What is already known, and what must be obtained first.
03problem decomposition
How a problem is split before intelligence is applied.
04intelligence-source selection
Which model, tool, research path, or person for which subproblem.
05capability calibration
Under which conditions a given source has been useful before.
06source diversity
How genuinely different the participants are, not how many there are.
07context allocation
What each source is given, compressed, or deliberately not given.
08memory management
What is carried across steps, and what is allowed to decay.
09information flow and asymmetry
Who sees what, when, and whether independence is preserved.
10interaction topology
Sequential, parallel, adversarial, hierarchical — or mixed.
11sequence and parallelism
Whether reasoning develops simultaneously or builds in stages.
12external research and tools
When to look outward instead of reasoning further inward.
13verification
How claims are challenged before they are accepted.
14adaptive compute / effort
How much effort a given subproblem actually deserves.
15synthesis
How partial results become one coherent position.
16output architecture
The form an answer takes so it can actually be used.
17stopping conditions
When further reasoning stops being worth its cost.
18outcome observation
What happened afterwards, and what it implies for orchestration.
19human participation
Where a person holds information the machines do not.
These are dimensions of the general design space. They do not describe IQ256's implementation, which remains proprietary.
The important problem is not how many models to run. It is which intelligence to invoke, for which subproblem, with which context, in which sequence — and when to stop.
External research finding — not IQ256 performance
“how to harness the collective expertise of multiple LLMs is an exciting open direction.”
Overlap, complement, extractionconceptual
Reasoning sources overlap substantially. The engineering question is how to reach the parts that do not overlap, test them, and fold them into one answer — not how to collect more votes.
conceptual pseudo-codeillustrative
# conceptual only — not production logic
state = frame(problem, constraints)
while expected_value(next_action(state)) > cost(next_action):
subproblem = select_uncertainty(state)
source = select_intelligence(subproblem, state)
evidence = invoke(source, context_for(source, state))
state = update(state, evidence)
return design_outputs(synthesize(state))Conceptual pseudo-code, not production source. It illustrates adaptive selection of the next uncertainty and the next intelligence source, and stopping once the expected value of further work falls below its cost. Routing rules, prompts, participant counts, scoring, and thresholds are deliberately not shown.
External research finding — not IQ256 performance
“the effectiveness of different approaches to scaling test-time compute critically varies depending on the difficulty of the prompt.”
Same ingredients, different topologyconceptual
parallel → aggregate
Independent passes, combined once.
challenge / verify
Claims are contested before acceptance.
sequential / refine
Each stage rewrites the prior state.
Changing interaction topology changes what information becomes visible and how errors propagate.
The same sources, given different context and a different sequence, can produce materially different results. Composition is a design decision, not a formatting one.
03Research
Fusion is an active engineering and research space.
The positions below shape how we allocate effort. They are stated as working positions, not settled results.
R2
Sources differ in practical value
Capability, cost, latency, availability, correlation, and context needs differ. Those differences are the design material.
R3
Capability is conditional
A single leaderboard score does not say when a source is useful. Usefulness is conditional on problem, context, and task.
R4
Context transfer is part of the cost
Bringing a source up to speed consumes budget, attention, and time — and can degrade as much as it adds.
R5
Composition and sequence matter
The same sources in a different order, topology, or context can produce materially different results.
R6
Intelligence avoided is a result
Not invoking an unnecessary model, tool, search, or person is a successful outcome, not a missed opportunity.
R7
Output architecture changes usefulness
The same reasoning, shaped differently, can be decisive or unusable for the person receiving it.
R8
Outcomes validate adequacy, not optimality
A good result shows the configuration was sufficient. It does not show that a cheaper or better one did not exist.
Open questionsunresolved
- —capability calibration over time
- —context-transfer economics
- —intelligence-source routing
- —composition and interaction effects
- —counterfactual workflow evaluation
- —cost / quality / latency prediction
- —outcome feedback and orchestration learning
- —future Solo AI capability
We are confident because we hold a strong thesis, not because we consider these questions closed.
External research finding — not IQ256 performance
“multiple language model instances propose and debate their individual responses and reasoning processes over multiple rounds”
Debate is one interaction topology among many. It is evidence that structure between reasoning sources matters — not a description of IQ256's architecture.
04Human + AI
Human + AI is the logical extension of Fusion.
This is a practical engineering position, not philanthropy and not an argument against automation. Human + AI is not universally superior; it is one configuration among several.
A person is the right source when relevant experience, tacit knowledge, organizational context, an unusual perspective, or accountability contributes information that is unavailable elsewhere. Human attention is scarce, so it should be selected for marginal value — not for panel size.
Solo AI
AI + AI
AI + tools
AI + humans
Configurations, compared by what they add and what they costconceptual
| Configuration | Adds | Cost to consider |
|---|---|---|
| Solo AI | Speed and simplicity; often sufficient. | Single perspective, single failure mode. |
| AI + tools | Facts and computation the model does not hold. | Latency, tool reliability, and context handling. |
| AI + external human | Domain experience and accountability from outside. | Scheduling, briefing, rates, and conflicts. |
| AI + internal context | Organizational facts that were never written down. | Employee attention, and the cost of transferring context. |
Solo AI
- Adds
- Speed and simplicity; often sufficient.
- Cost to consider
- Single perspective, single failure mode.
AI + tools
- Adds
- Facts and computation the model does not hold.
- Cost to consider
- Latency, tool reliability, and context handling.
AI + external human
- Adds
- Domain experience and accountability from outside.
- Cost to consider
- Scheduling, briefing, rates, and conflicts.
AI + internal context
- Adds
- Organizational facts that were never written down.
- Cost to consider
- Employee attention, and the cost of transferring context.
This describes the broader Fusion research space. It does not describe an IQ256.ai product that orchestrates customer employees.
The question is not “human or AI?” It is which intelligence source has the most relevant information for this subproblem.
Context
Facts about the situation that were never written down.
Experience
Judgement built from outcomes, not from text.
Perspective
Framings a model is statistically unlikely to produce.
Accountability
A decision owner, which no model can be.
05About
Fusion is currently used internally across selected analysis and decision-support workflows.
IQ256 is a founder-led research and engineering effort focused on practical AI Fusion. We build this because we use it. Internal evaluations are encouraging but unpublished, so we make no public performance claim from them.
We deliberately do not generalise from that. The broader objective is learning when to invoke which intelligence, and how to compose it — including when the right answer is to invoke less.
What is demonstrated, observed, believed, and openposition
| Category | Position | Basis |
|---|---|---|
| Demonstrated externally | Selected multi-model and multi-agent architectures can outperform strong individual models in particular settings. | Published research and production systems. |
| Observed internally | Internal evaluations are encouraging, but they are unpublished and not independently verified, so we make no public performance claim from them. | Our own internal evaluations. No published figures. |
| Our thesis | The engineering problem is allocating and composing intelligence — which source, which subproblem, which context, which sequence, when to stop — not running more models. | Working conviction, actively tested. |
| What we do not claim | Not universally better. Not hallucination-free. Not an optimal or finished architecture. | Deliberate limit on our claims. |
Demonstrated externally
- Position
- Selected multi-model and multi-agent architectures can outperform strong individual models in particular settings.
- Basis
- Published research and production systems.
Observed internally
- Position
- Internal evaluations are encouraging, but they are unpublished and not independently verified, so we make no public performance claim from them.
- Basis
- Our own internal evaluations. No published figures.
Our thesis
- Position
- The engineering problem is allocating and composing intelligence — which source, which subproblem, which context, which sequence, when to stop — not running more models.
- Basis
- Working conviction, actively tested.
What we do not claim
- Position
- Not universally better. Not hallucination-free. Not an optimal or finished architecture.
- Basis
- Deliberate limit on our claims.
lab noteWe add intelligence where an error is more expensive than the extra effort — and we treat intelligence avoided as a result, not a shortfall.
Fusion is used internally today. For scoped assignments, visit iq256.com.
Fusion is used internally in selected analysis and decision-support workflows, and it may be used inside selected case-by-case decision reviews handled through iq256.com.
The technology itself has no public product access and no public API. We are not announcing a future API or public launch date.
06Contact
Get in touch
We welcome non-confidential technical discussion, collaboration, and investor conversations. There is no public Fusion product or API to sign up for: both routes open IQ256's contact form on iq256.com, and legitimate enquiries are reviewed personally. Replies are usually sent within 2–48 hours. Treat the first message as non-confidential: a short summary of the decision or question is enough — there is no need to attach documents or sensitive detail at that stage.