--- title: "Anyone can demo AI. We prove it. | AimRank" url: https://www.aimrank.io/methodology/ description: "The AimRank Standard: every build carries the Four Guarantees (Measured value, Defensible, Self-improving, Yours & everywhere) or we say which we scoped down." language: en updated: 2026-11-12 --- [← Back to the climb](https://www.aimrank.io/#climb) The AimRank Standard # Anyone can demo AI. We prove it. Individual AI is incredible, until a business depends on it. Then it breaks on reliability, calibration, drift, audit, and scale. The AimRank Standard is the layer that turns a clever demo into a system you can trust. Every solution we ship meets four guarantees, or we tell you plainly that it can't. ## The Four Guarantees™ 01 ### Measured value Cuts cost, grows revenue or builds an advantage rivals cannot copy. We prove the number, before launch and after. How we achieve it A conservative ROI case is agreed up front, then the build runs against a versioned eval suite scored by an independent judge (itself validated against human labels), with task, trajectory and groundedness checks. It ships only if the suite passes, and after deploy, delivered value is tracked against that agreed number. Where a classical model is in the loop, we add fairness checks on top. **What you get as proof:** The eval report with the pass-or-fail gate, plus the value report against the agreed number. 02 ### Defensible You will defend any decision to a client, a board or a regulator, with a readable reason, a tamper-evident audit trail and the EU AI Act paperwork already in hand. How we achieve it Each decision carries its reasoning, the full tool-call trace (what it did, with what inputs and outputs) and the sources it relied on, recorded as it runs so a human can review, approve or override. A classical model in the loop adds per-decision SHAP reason codes. Annex IV documentation and a tamper-evident, hash-chained audit log (Article 12) are generated from the run itself. Documented, evidenced and defensible against EU AI Act Articles 13 and 14, as of the dossier date. **What you get as proof:** The decision trace and citations for any single decision, plus the Annex IV dossier and the replayable audit chain. 03 ### Self-improving It never fails silently: it tells you the moment it starts slipping. Where there is a feedback signal, it gets better the longer it runs. How we achieve it Watched by default (the floor): every output carries calibrated confidence and the system escalates when unsure instead of bluffing; we re-run the eval suite on a schedule and on every model upgrade, watch behaviour (refusal rate, tool mix, cost) and incoming topics, and pin the model version so a provider update can't silently shift it. A classical model in the loop adds probability calibration (Brier, ECE) and input-drift monitoring (PSI / KS). Closing the loop (the escalator): human corrections, outcomes and signals feed a living eval set, and a challenger version only promotes if it beats the incumbent on that frozen gate, every promotion human-approved, versioned and reversible. A build with no feedback signal still clears the floor; one with a signal compounds instead of decaying. **What you get as proof:** The live drift dashboard, the eval-regression history and alert log, plus the versioned eval set and the champion-or-challenger promotion record. 04 ### Yours & everywhere Any cloud or your own infrastructure, with your data kept in your region and access you control. An MCP endpoint for your agents. Yours to own, no lock-in. How we achieve it The whole system ships as one Terraform-managed container in your own region (serverless container host by default, Kubernetes where you run one), so your data never leaves. Every build exposes its capability as MCP tools (predict, explain, evidence, healthcheck, plus act for agents) with tiered access, so it's a composable tool in your agent toolbelt, and autonomy is consequence-gated with human-in-the-loop and a kill-switch. **What you get as proof:** The repo, the Terraform, and a live MCP endpoint your agents can call: all yours to keep. Proof, not description ## What "we prove it" looks like: a real reliability diagram Dot size = appointments in the bin. Dotted stem = the gap between what was predicted and what happened. The ship gate, same run Calibration (ECE, overall) 0.0228 gate ≤ 0.05 PASS Discrimination (ROC-AUC) 0.696 gate > 0.6 PASS Fairness (worst within-group ECE, age 55-74) 0.0710 gate ≤ 0.05 FAIL Reliability diagram from the #11 No-Show Prediction backtest: 30,730 appointments from the latest week of a public dataset (OpenML 43439, Brazil, 2016), held out in time. ECE 0.0228 against a gate of ≤ 0.05\. The fairness check fails on this run, so this build does not ship; the bar stays where it is. Model 0.0.1+5fa3292b, 8 Oct 2026\. Every build regenerates this from its own run; nothing here is drawn by hand. The two bins above 50% hold 12 appointments between them. ECE weights every bin by its count, which is why they are shown and why they barely move the number. ## How a build actually runs Six steps, with a human checkpoint at each one. 1. 1 Scope & risk-class Define the buyer, the wedge, and the risk class. Legal sign-off where it carries weight. 2. 2 Data intake We point at your data in your own store. No personal data leaves your infrastructure. 3. 3 Model, calibration & bias audit Baseline first (classical often beats deep), validate calibration, review disparate impact. 4. 4 Drift baseline & feedback loop A real reference window, a nightly drift job and alerting before launch, plus feedback capture wired into a living eval set and champion-or-challenger promotion. 5. 5 Evidence & dossier Complete the validation docs and verify the audit chain. 6. 6 Deploy & handover Cloud-agnostic deploy, runbooks, and training so you own the system. ## For agents: Agent Assurance For an agent we tell you three things: how often it fails, when it hands off to a person, and whether it paid for itself against the number agreed before the build. Before we build one, we test whether a simpler system passes the same test; if it does, that is what ships. Where an agent would have to judge its own certainty, it asks a calibrated model instead, and an unsure answer goes to a person. Reliability is counted over repeated runs of the same task, next to cost and speed. The four guarantees secure each tool an agent calls. When agents start to act, we lift assurance to the decision and action level: we evaluate the decision and the whole trajectory (not one output), watch for behavioral drift, trace every action into a tamper-evident log you can replay, and gate autonomy with guardrails, human-in-the-loop, and a kill-switch. Autonomy without assurance is recklessness; agents make assurance more valuable, not less. ## Built for agents, not just screens Every system ships as an MCP server (predict, explain, evidence, healthcheck), so your agents can use it directly, not just your people. This page describes the methodology and the evidence each step produces. The implementations live inside the AimRank solutions we build for you. ## Want this level of rigor on your AI? [Not sure where you are? Take the 2-min assessment →](https://www.aimrank.io/assessment/)