Insights and Research

A matching headline number can hide wrong subgroups, collapsed variance, and answers that flip when you reorder the questions. Two 2026 benchmarks on why an aggregate pass is not validation.

Ask a language model to pick a number from 1 to 100 and most of the time it says 42. A KAIST group named why: a model can describe a distribution and still not sample from it. What that breaks for synthetic research, and what it does not.

2026 was the year the field named its hard problem. A map of what synthetic population research now agrees is broken, what scale will not fix, and where the answer actually lies.

Half the synthetic research on the market is validated by a benchmark that grades its own homework. Preregistration is the bar that separates real evidence from a demonstration.

Public bodies must consult the hardest to reach, which is where traditional research is weakest. How a grounded synthetic panel lets a public body rehearse a policy or service change before it commits.

Several independent research groups proved in 2026 that the standard way of building synthetic populations does not reproduce real people. We reached that conclusion first, and have a pre registered trial on real humans to prove ours works.

The quiet bet behind most synthetic research is that a bigger model will fix it. A Stanford group tested that bet and it lost on the settings that matter. We never made it, and have a pre registered human trial to show why.

Every synthetic research vendor can now quote the same 2026 papers back to you. Here are five questions that separate a solved problem from a rehearsed answer, and why we can pass all five on real human evidence.

For twenty five years the web’s link graph gave up a single number: authority. We have now shown two more things propagate through it, the character of content and the geography of a site, by one mechanism. So what else is in there?

Pre-registered, hash-committed before voting. KPM-1 2.96pp vs YouGov 3.20pp, Opinium 3.20pp, Electoral Calculus 5.00pp. Honest miss on council control owned in paragraph 2. Updated 11 May with the full verifiable portfolio state: KPM-2.2 v15.1 hits 59.2% on 130 councils, by-election engine 75%, ward-level v17.10 pre-registered for 2027 county and 2028 metropolitan cycles. Every accuracy claim reproducible from github.com/Kronaxis/kpm.

The bet-able subset of the pre-registered predictions: 10 lean calls + 10 tightest toss-ups. Of the 136 contested, KPM-1's 4-seed ensemble has 10 races it'll bet on; the other 116 are too close to call. Specific, betable, falsifiable.

KPM-1's first public election stress-test. 136 councils. SHA-256 hash committed to GitHub before voting opened. 5 May 4-seed ensemble: 0 confident calls, 10 leans, 126 noted toss-ups. Council pluralities Labour 62, Reform UK 46, Liberal Democrat 15, Conservative 11, Green 2.

Traditional research takes weeks and costs thousands. Panel Studio builds synthetic consumer panels from personality science, runs multi-turn conversations, and delivers reasoning traces you can audit. Self-hosted or cloud. Comparison vs traditional panels, SurveyMonkey, and synthetic alternatives.

Reform UK is about to sweep English local government. We publish predictions for 20 councils with full methodology and validation data.

We tested synthetic persona panels against 10 real council by-elections. V9 predicted 6 of 8 winners with 7pp mean absolute error.

Full research paper: methodology, per-ward results tables, nine pipeline iterations, calibration discovery, and 20-council predictions for 7 May 2026.

Every agent session preloads 100 KB of project notes per turn. Kronaxis Fabric replaces it with on-demand hybrid search over a Postgres + pgvector store. 92K tokens saved per session on a real project, paired with Kronaxis Router for 100-600x per-session cost drops.

Small open-weight LLMs now handle 80% of production workloads identically to frontier APIs. Auto-classify prompts, route to the cheapest capable model, validate quality in a closed loop. Full comparison vs LiteLLM, OpenRouter, Portkey, and Martian.

A UK grocery retailer ran Panel Studio alongside traditional focus groups. Same winning design. Deeper explanation. Four hundred times the speed.

Big Five measures traits. DYNAMICS-8 predicts behaviour. The gap between personality measurement and behavioural prediction, and why it matters.

Reasoning traces are worth 10 to 100 times more than plain text for model training. The scarcity is structural.

From API key to exported results in under ten minutes. Step by step with Python code examples.

Discipline predicts price resistance. Mercuriality amplifies anxiety. Impulsivity triples conversion on scarcity framing. The data behind personality driven pricing.

Six years of building in ultra competitive digital markets. The pivot that turned persona simulation into a consumer behaviour prediction platform.

Groupthink, social desirability bias, and post-hoc rationalisation. The three traps that make focus group data unreliable.

How DYNAMICS-8 country norms adjust personality expression across cultures. Same framework, different baselines, comparable results.

Two women, same age, same income, same postcode. Completely different purchase decisions. Why demographic segmentation is not enough.

The 3 pass generation pipeline, 18 rule validation engine, and census weighted demographic sampling behind every Panel Studio panel.

How to build agents that behave consistently by personality type, not generically. From prompting to training on DYNAMICS-8 traces.

A proprietary framework nobody cites is worth nothing. An open framework that becomes a standard is worth everything.

Synthetic consumer panels for startup founders. Test pricing, messaging, and positioning against personality segments before writing code.

High Discipline visitors read every feature. High Impulsivity visitors click in 8 seconds or leave. One pricing page cannot serve both.

An honest assessment of when synthetic panels are not the right tool, and the hybrid approach that cuts research costs by 80%.

The definitive explainer. What synthetic panels are, how they work, when to use them, and how they compare to traditional research methods.

Demographics explain 10-15% of purchase behaviour. Personality conditioned models explain 50-60%. The three layer prediction hierarchy.

Run full factorial conjoint through a 300 person synthetic panel. Hours not weeks. Under £200 not £50,000. Personality segmented willingness to pay.