A five-layer methodology. Deterministic where deterministic beats guesses. Multi-model where consensus beats hero-worship. Sourced against real authorities where the source exists. No black boxes, no vendor spin, no scores we can't defend.
The score isn't a vibe. It decomposes every response into three orthogonal reads on a 1–5 scale — because "is this AI answer good?" is three questions in a trench coat, and answering all three at once produces the vague meters other tools give you.
Does the substance match reality? Are the facts, figures, and citations grounded in something you can check — or invented? Fidelity is the anti-hallucination read.
Does the AI's confidence match its evidence — or is it overselling? A hedged uncertain answer with hedged uncertainty is fine. A definitive answer with no basis is dangerous.
Did the answer actually address the question — or dodge, deflect, redirect, or half-answer? Completeness matters. So does staying on topic.
F, P, and R combine into a single RI 0–100 that appears in the LiveScope overlay tile. Stable (≥75), Warning (55–74), Danger (<55). The number is anchored by median-of-3 sampling so the same input doesn't wobble across reloads — the variance problem that plagues one-shot LLM judging.
The weakest link in every AI-scoring tool is: "the judge model thinks it looks fine." LiveScope's Fidelity number is computed differently — every checkable claim in an AI answer is extracted, ranked for materiality, and verified against real sources. F_verified is the receipt, not the vibe.
Deterministic regex — no LLM. Pulls out entity-figure pairs ("Barclays 23%"), named studies with dates ("TSB First-Time Buyer Report May 2025"), named academics, named institutions, and explicit URLs. No LLM interpretation in this step, so the extraction itself is reproducible.
Every claim gets a materiality weight — is it load-bearing to the answer's conclusion, or an incidental aside? A wrong load-bearing figure sinks Fidelity. A wrong passing mention barely moves it. That's the difference between "the mortgage rate is 4.5%" (load-bearing) and "the FTSE closed near 8,200" (incidental) in an answer about mortgage strategy.
Each claim is checked against two independent verification stacks, in this order:
Materiality-weighted receipts roll up into F_verified. If no checkable claims exist (creative writing, pure opinion), the score returns "Reasoning-only" instead of inventing a number — because a fabricated fidelity read on an opinion is worse than none.
When an AI states "the ISA allowance is £20,000" or "the personal allowance is £12,570", that's a claim you can verify deterministically. LiveScope's Number Verifier matches each numeric claim in the answer to a curated UK-finance fact map, then live-fetches the current authoritative value from GOV.UK's Content API. Zero LLM in the verify step — zero opinion.
| Verdict | What it means |
|---|---|
| MATCH | The AI's figure exactly matches the current authoritative value on GOV.UK. |
| MATCH_WITH_CAVEAT | Matches, but with a date-sensitivity note (e.g. rate applies to a specific tax year). |
| MISMATCH | AI's figure disagrees with the source. The AI is wrong on the number. |
| UNCHECKED | No matching fact in our map, or fetch failed. Never a fabricated verdict. |
The fact map is curated. As of now it covers ~40 UK-finance figures (ISA + LISA allowances, income tax bands, personal savings allowance, dividend + CGT allowances, State Pension, National Insurance thresholds, some Bank of England base rate references). Anything not in the map returns UNCHECKED — not a made-up verdict.
The map expands weekly. Coming next: NHS + medical thresholds, employment law figures, education sector reference data. If there's a numeric category you rely on, we want to know — it goes in the queue.
One judge is a single point of failure. Every AI-scoring tool that uses "the LLM to grade the LLM" has the same in-group-bias risk: a Claude judge scoring a Claude answer might over-favour Claude's style. LiveScope solves this the honest way — Pro users get their answers scored by three tier-1 judges from three separate model families.
Primary judge, median-of-3 stability sampling for variance control. EU-hosted via AWS Bedrock.
Secondary judge, single-shot. Cross-checks Claude on cost-efficient tier.
Tertiary judge, single-shot. Independent of Anthropic and OpenAI. Fast inference.
The overlay shows all three scores + a consensus_ri (median) and a max_disagreement figure. When the three judges agree within a few points, you have three tier-1 brands endorsing the read. When they diverge by 30+ points, that's a real signal — Claude might be being over-harsh, or Gemini over-lenient, or the answer is genuinely on a knife-edge. Either way, you know before you act.
Beyond the score, LiveScope watches for named behavioural patterns that reliable AI answers should never exhibit — and flags them the moment they show up. These are the M-codes: our documented taxonomy of ways AI can go wrong even when the answer superficially looks fine.
Telling you what you want to hear. "Great question!" opening, agreement without evidence, softening a legitimate correction to please the user. Flagged so you know when the AI's warmth is priced in.
Fabricated citations, invented experts, made-up study names. Detected as a distinct signal so it doesn't get lost in "Fidelity: 2".
Ungrounded claims delivered with certainty — the classic hallucination pattern. Deterministically triggered from the judge's flag output.
A weasel-word disclaimer at the end that doesn't actually reduce the specificity of the claims above it. "This is general information, not advice" tacked onto a detailed prescription. Detected separately so you spot the illusion of hedging.
Pretending to answer both sides while quietly favouring one. The "on the one hand… on the other hand…" that isn't balanced.
Patterns that only emerge across multiple turns — escalating confidence, register drift, warmth compounding. Tracked at session level, not single-response.
The most common way AI conversations fail isn't a single wrong answer — it's ten turns later, where the AI has quietly wandered off the question you actually asked, and you've been following along. Coach exists to catch that.
Coach reads up to 10 turns of your conversation and identifies your active goal — usually your first ask. That's the reference point for every intervention.
If you asked about UK pensions in turn 1 and by turn 6 the AI is discussing crypto tangents — without you asking it to — Coach fires a back_on_track intervention. The suggested prompt pulls the conversation back to your original ask.
If you explicitly redirect ("actually, forget pensions — help me with ISAs instead"), Coach recognises the goal shift and follows the new anchor. Only user-initiated pivots move the anchor. AI-initiated topic drift never does.
Coach only fires when its confidence is above 65% and it can name a concrete next action. If the best it can produce is a generic nudge, it holds silent. Every intervention must MOVE THE USER FORWARD — never demand the AI prove itself, never stall progress.
When the tile isn't enough — when you need to defend a decision, escalate a finding, or run a formal review — Deep Audit produces a five-card structured breakdown for any AI response. Sector-aware: legal, financial, medical, coding, education, and creative all get sector-tuned scoring rules and reference data.
Every quote we couldn't verify + why. Severity-graded (high alerts are load-bearing to the answer; warns are secondary; ok = nothing flagged).
Points in the answer where the AI's stated confidence exceeded its evidence. Overselling → flagged.
Up to 8 URLs live-checked. Each returns one of: trusted (allowlisted publisher), ok (2xx/3xx response), broken (does not resolve), blocked (couldn't check). Broken citations are the fabrication tell — legitimate answers cite real pages.
Session-level early-signal detection — "something is forming" across the recent turns. Includes a confidence % and a suggested check.
How your session's RI has moved across recent responses. Declining / stable / improving, with the actual trend line. Catches the slow-drift-across-a-long-chat pattern.
The methodology only means something if the machine running it is trustworthy. Here's exactly how we set that up.
Names, emails, phone numbers, postcodes, NHS numbers, card numbers, IP addresses, dates of birth — all stripped in the browser BEFORE the AI's answer leaves your extension for scoring. The judge never sees who you are.
All scoring runs on AWS Bedrock in the EU cross-region inference profile (Ireland, London, Paris, Frankfurt, Stockholm, Milan, Spain). Data never leaves the EU/UK perimeter during model processing.
Independent research. Not funded by any AI company. Not affiliated with any AI company. We score them all the same way, and we don't answer to any of them.
EverythingThreads is an ICO-registered data controller (C1896585). Single controller for UK + EU subjects. Sub-processor list published in the Privacy Notice.
The calibration methodology anchor was submitted to a preregistered analysis plan on 24 April 2026 (currently embargoed pending publication). This means the way we score can't be silently changed after the fact — the ruleset is time-stamped.
The judge models are named: Claude Sonnet 4.6 (primary), GPT-4o-mini + Gemini 2.5 Flash (Pro consensus). If a model changes, it's in a changelog. No silent A/B on live users.
The methodology is the same. Access to layers deepens with the paid tier. Free is genuinely useful — not a paywalled tease.
A Free-tier score costs us roughly £0.007–£0.010 to compute (LLM + detector calls). A Pro-tier score with Source Checker + 3-judge panel costs roughly £0.019–£0.024. A Deep Audit costs about £0.028. We tell you this because "free" AI tools that don't have a business model become the business model. Ours is Pro subscriptions.
Install the Chrome extension. Ask any AI anything. See the RI + F/P/R + flags the moment the answer lands. No signup required for the free tier.
Get LiveScope →