How Elite Intel works
Every verdict has a number behind it.
Elite Intel publishes BUY, AVOID and WATCH calls on UK performance cars. Each call is anchored to evidence we collect, score and grade — and to a threshold we won't publish a strong verdict below. This page explains how the system works, what the numbers behind every pick mean, and how editors can override the system when they have reason to.
How the data gets here
Every tracked car has a continuous evidence pipeline behind it. We pull listings, auction results, dealer stock and specialist commentary from a vetted source registry — AutoTrader UK, PistonHeads, Collecting Cars, CAP HPI, Tom Hartley Jnr, Joe Macari, Romans International and the rest. Each row goes through normalisation (so AutoTrader and Pistonheads call a Cayman GT4 the same thing), confidence scoring (so an unverified asking price doesn't carry the same weight as a recorded sale), and a recency check (so a six-month-old price isn't pretending to be current).
Sources we don't recognise are flagged for review rather than silently accepted. Sources that disagree with each other generate a data-quality issue our analysts work through. Manually-added evidence is protected from being overwritten by an AI refresh — operator input outranks the scraper.
Why some picks say "INSUFFICIENT EVIDENCE"
A verdict on a car you might spend £100,000 on shouldn't be louder than the evidence behind it. Every pick runs through a confidence gate: if the underlying score is below the segment's threshold, the BUY / AVOID / WATCH badge degrades to "INSUFFICIENT EVIDENCE" — even if our editors privately lean a direction. That keeps the public verdicts honest about what we actually know.
The thresholds aren't a single global number. A hypercar segment with two transactions a quarter can't reasonably be held to the same evidence floor as a Cayman GT4 with 200 listings live at any moment. So we tune the gate per segment.
Current thresholds
Thresholds are reviewed each quarter. A score below the segment threshold means we have a working hypothesis, but not yet enough verified evidence to underwrite a strong call to the kind of money the segment trades at.
What the 0–100 score actually measures
Each car's confidence score is a composite of six signals captured over the last thirty days: verified price points, supply (listings live), demand (search interest, days-on-market trends), known-issue coverage, ownership-cost data, and pre-purchase inspection rules. A car with sixty verified prices, four independent supply sources and three independent dealer signals scores high; a car with one asking price screenshot scores low.
We also track how stale the most recent market snapshot is. A model that hasn't been refreshed in eight weeks loses points even if the underlying evidence was strong — because the market doesn't sit still, and a stale signal can be worse than no signal.
On every public pick you'll see the score alongside the verdict: BUY · 73/100. The number is the input; the badge is the output the gate produced.
When editors override the system
Editors have the final word. They can override a system-suggested verdict — but they have to log a rationale, and that rationale is preserved in the audit trail. We don't publish a strong verdict on a weak-signal car without a named editor putting their name to the call.
In practice that means a couple of things you can rely on. We won't pretend to a confidence we don't have. And when we do publish a verdict against a thin-data segment, you'll be able to see in our reasoning why we backed it anyway.
Where AI helps and where humans decide
AI does the heavy lifting on the boring parts: pulling fresh evidence, matching listing names to canonical models, scoring confidence, drafting the long-form sections, and flagging copy that drifts from our editorial voice. Several AI models are used for different jobs — fast ones for matching, web-grounded ones for live news and auction roundups, an editorial model for the prose. The current mapping is documented internally and reviewed every release.
Humans decide on every verdict. Every monthly Report passes through a three-step approval pipeline before publish — automated content QA, an editor's review, and a final sign-off — and a Report can't be published with a section that has zero substantive body content or a pick whose evidence is weak and unsigned-off. Those guards are enforced in code, not policy.
Sources, citations, and what we'll never do
Every section of every monthly Report ends with a Sources block listing the data we drew on — typed once, by the system, from the actual evidence rows that fed the section. No padding. If a source isn't in our registry, it doesn't feed a verdict. If a verdict is the result of an editor override, the override is logged.
What we won't do: take dealer commissions on cars we cover, suppress an AVOID verdict because an advertiser wouldn't like it, or auto-publish a strong verdict on weak evidence to fill a slot. Editorial decisions are independent of commercial relationships and that boundary is non-negotiable.