How Arc works, honestly
Arc is deterministic at its core, with a probabilistic way of explaining itself. Everything that computes here is versioned tables built from published research – and the results come back as distributions, because customers are individuals, not labels.
It exists to help marketers and business people think harder about their individual customers and their customer groups – so they can design better propositions, products, and campaigns, and execute them more effectively.
Underneath sits a selection of published models – see The models – tweaked with industry knowledge and a little smarts to bring them to life, much aligned to the marketing theory in Green But Not Heard by Mat Moakes.
Arc is for strategy and communication design – never pricing, credit, eligibility, or the exclusion of any group. The changelog, cross-examination record, sources, and limitations are below.
- v1.0.0 – 26 Aug 2026 – tables frozen after five-model cross-examination and written adjudication; engine stamp off draft
- v2.1.0-draft – 26 Aug 2026 – two lenses live with pre-panel draft tables; five-model cross-examination scheduled before v1.0
Before freezing, every table went to five independent AI models – ChatGPT, Gemini, Grok, Perplexity, and DeepSeek – with one fixed challenge prompt and every source document attached. Each was asked to audit anchors, attack judgement cells, hunt register leakage, break the discriminability of similar pairs, test for circular reasoning, probe the blend’s edge behaviour, and challenge the message layer. Every challenge was adjudicated in writing: accept, reject, or modify, with rationale.
What changed at v1.0: cells were removed and trimmed, nothing was added. Every cell that scored a motivational archetype from a circumstance – renting, news avoidance, purchasing capacity – was struck; the Practical Realist is defined by decision economics, assuming neither abundance nor constraint, and the Idealist has no age. Two cells that re-imported segment evidence through the crosswalk were struck, and a firewall now applies: the crosswalk may consume frozen lens outputs, and no lens cell may cite, inherit from, or be reasoned backwards from the crosswalk.
The engine says less when it knows less. Points in the confidence blend mean post-cap effective points. Zero evidence renders insufficient information, never a ranked result. One signal renders a single-signal label. Exact ties stay ties, and where population base rates order a close result, the output says so. Falsifier precedence is deterministic and published: when footprint and ripple both fire, the ripple clause decides; when evidence and control both fire, evidence outranks control; a stated motivation carries its full weight only as a confirmed picker selection.
The message layer downgraded itself. A resistant grade now requires a named page or figure citation; every uncited cell dropped to wary. Two phrases that were simultaneously scored and flagged are now flag-only. One housekeeping note for transparency: the Lens 2 structural family cap is 6 by convention, but the maximum attainable under v1.0 weights is 2.
Arc's third lens is in build – the Value Lens. It reads transition headroom, what remains of a household's journey step by step, and it will read it three ways. The commercial view is about fairness in the oldest sense: value is created for the customer first, and the business that serves them captures a fair share as profit. The policy view reads the same ladder for carbon, adoption, and fair distribution of the transition. A vulnerability view reads it for who needs protecting along the way. Same facts, different stakeholder.
The rules that protect this are already in the tables. Wealth and financial capacity never score the archetype lens – that is rule one of the freeze, and it stays. Attitude to finance, which is an attitude rather than a capacity, is authored as a candidate attribute for the v1.1 panel alongside cited £ ranges per headroom step; nothing joins the scoring tables without cross-examination, and no £ figure ships without a citation.
The Arc Index itself – a published benchmark built on all three lenses – is a stated ambition for the future, not a shipped product. What is live today is exactly what the pages show: two scored lenses, a message layer, and the qualitative headroom ladder. Arc never prices, scores credit, or excludes.
- More in Common – the seven British segments (2025)Source →
- Britain Talks Climate & Nature – Climate Outreach & More in Common (2025)Source →
- GBNH Archetypes – our own lens, formal validation pilot scheduledSource →
- Defra – Pro-Environmental Behaviour Framework (2008)Source →
- Rogers – Diffusion of Innovations (1962)Source →
- Yale – Global Warming’s Six AmericasSource →
- Ofgem – Energy Consumer Archetypes (2020, updated 2024)Source →
- Nesta – heat pump consumer segmentation, eight audiences (2025)Source →
The v1.0 tables are frozen but not finished. Every cell graded DJ→PILOT awaits the n=200 pilot crosstab export, which upgrades or strikes it; the GBNH lens runs on a uniform prior until a representative sizing study exists; and the receptivity matrix carries judgement grades pending a cell-by-cell citation pass. An auditable wrong number is still a wrong number – which is why the working is always on the page.