Novacore

The API: divergence inside your own stack.

Novacore is the Novamine API (Application Programming Interface) and chat: a drop-in step at the point where an existing workflow needs non-consensus breadth.

Ask one question and get a diverse set of responses, all highly usable, where a standard AI model repeats the same two or three. Ranked first on Carnegie Mellon’s NoveltyBench for both novelty and utility.

Your data is never used for training. Every response is scored for novelty and utility before you see it.

Try Novacore →Talk to us →

NoveltyBench · Carnegie Mellon University · merged September 2026
9.30/10
Distinct answers in ten
+39% above the next best entry on the leaderboard
6.65
Utility
+25% above the next best entry: more ideas without losing quality
3.8×
More distinct ideas
Than a standard AI model asked ten times
1st
On both measures
The only system to hold the novelty and utility crowns together

Independently reproduced and merged by the Carnegie Mellon NoveltyBench maintainers (10 September 2026). Every benchmark figure on this page was scored by the benchmark's own scoring.

Novelty and rigour, simultaneously the trade-off other approaches accept, and Novamine removes

Until now you had to choose. Push an AI model for variety and the answers get stranger but less dependable. Hold it to quality and it falls back to the answer everyone already has.

Novamine removes that trade-off. On the independent benchmark it produced the most different answers of any system, and those answers were more useful, not less.

A standard AI model
One question, asked ten times
Two or three ideas, repeated
Novelty
Utility
The rest are the same ideas in different words.
Other ways to force variety
The best published alternative
More ideas, at a growing cost
Novelty
Utility
Each extra answer costs more than the last.
Novacore
One question, ten answers
Nine different ideas in ten
Novelty
Utility
The most novel and the most useful on the benchmark
The cost per answer stays flat.
Illustration drawn from Carnegie Mellon's NoveltyBench results (September 2026). Each square is one of ten answers: a coloured square is a new idea, a grey square repeats an earlier one.
The Evidence an independent benchmark and an independent researcher

First on Carnegie Mellon's NoveltyBench, on both measures

NoveltyBench is an independent test of whether an AI system can give answers that are both different and good. Novamine's entry was reproduced and merged by the benchmark's maintainers on 10 September 2026. It ranks first for novelty and first for utility, and it is the only system to hold both. See the leaderboard.

The “Artificial Hivemind”, all but removed

The NeurIPS 2025 best paper showed that AI models give near-identical answers to open questions. We repeated the study using the author's own test, then applied Novamine. The share of questions stuck in the “hivemind” fell from 96% to 1%, and the effect carried over to a model from a second AI company.

The reply from the study's author:

"Thank you so much for reaching out and for sharing these results. I'm really glad to see your … system working so well on Infinity-Chat, especially in addressing a challenge that LLMs tend to handle poorly by default. The magnitude of the reduction is particularly striking, and it's also very interesting to see the effect transfer across model providers without modifying the system." Dr Liwei Jiang, author of the Artificial Hivemind study
In Your Workflow one call, at the point where you need breadth

Novacore is a drop-in step: 3 to 60 minutes, Rigour 0 to 3, at the point where an existing workflow needs non-consensus breadth. Your system stays as it is before and after the call. Six places it is used: three everyday workflows and three high-throughput uses of the same API.

In your organisation
Enterprise AI assistantYour system
Your R&D function
Discovery workflowYour system
Your planning process
Strategy workflowYour system
Your agent platform
Multi-agent and forecasting stacksYour system
Your model team
RLAIF and synthetic training dataYour system
Your knowledge base
RAG over your own dataYour system

Analyst query

A consultant puts a client problem to your internal LLM assistant.

?

Target and constraints

A mechanism, material or process problem, with its design constraints fixed.

Strategic mandate

Market entry, growth, pricing or capital allocation.

Question set

A forecast, decision or debate question goes to your agent ensemble.

Training objective

A model to tune, or an evaluation set to build.

User question

A question put to your retrieval-augmented assistant.

Baseline completion

One high-probability answer: competent, and the one every firm on the same model receives.

Consensus hypotheses

In-house models and literature-led ideation converge on the known mechanisms.

Consensus option set

Analyst work and standard AI return the plays your competitors are already modelling.

Correlated agents

Agents built on one model reason alike, so their errors add up instead of cancelling.

Collapsed samples

Synthetic data from one model clusters around the same few answers.

Retrieval

Your retrieval layer pulls the context from your knowledge base, data lake or warehouse.

Divergence request

The user asks for alternatives; the assistant routes the brief to Novacore.

Hypothesis generation call

High rigour for a few defensible mechanisms, or a wide pool for your own models to triage.

Option-space expansion

The team calls for non-consensus routes, tested to the rigour the decision needs.

Reasoning-path call

Your orchestrator asks Novacore for independent lines of reasoning.

Generation call

Your pipeline requests scored, distinct responses for each prompt.

Reading call

Question and context go to Novacore, in place of the single consensus answer.

Our system · the Novacore API
Novacore APIThe divergence step
One call returns distinct, high-utility answers from across the solution space.
Ranked #1 on CMU NoveltyBench (9.30/10 distinctness · 6.65 utility)

Ranked candidate set

Non-overlapping answers returned in-thread, each scored for novelty and utility.

Screened, distinct hypotheses

Mechanistically distinct, screened against prior art and scored.

Stress-tested options

Distinct strategies, each challenged and scored against the brief.

Independent answer pool

Distinct, scored reasoning paths: one for each agent, debater or forecaster.

Scored, distinct set

Responses from across the range, each scored for novelty and utility.

Distinct readings

Distinct, scored readings of the same evidence.

Session continues

The user selects a direction and iterates in your assistant, context intact.

Downstream screening

Your in-silico triage and wet-lab validation proceed unchanged.

Modelling and decision

Your team runs the financials on the strongest and takes one to the board.

Aggregation

Your mixture-of-agents, debate or forecast step runs unchanged. crowd error = average error − diversityDiversity Prediction Theorem: Page, The Difference, 2007

Tuning and evaluation

Your next model learns from the full range, not the collapsed middle.

Answer in your interface

Mode collapse does not follow you into your own data.