Health outcomes are reported as totals. The decisions that produced them (what was bought, when, for whom, on which evidence) stay invisible. Totals cannot be fixed; decisions can.

Procurement, supply and programme data reconciled into one narrative spine, then challenged by four AI models for gaps, bias and overclaiming before a single chart was drawn.

Which decisions matter is a judgement from the field, not a statistic. The piece names the decision points where small changes moved outcomes, and labels every figure by tier and confidence.

Scroll, and the argument builds itself. Stop anywhere and open the evidence. A funder or a ministry sees where the leverage is, and how much each claim rests on.

Hundreds of offices, thousands of suppliers, one annual report. By the time the pattern is visible, the year is over. Static spend reviews describe the past; they do not control anything.

The method I built by hand inside a global institution, rebuilt with AI as an agency-neutral engine: categorisation, clustering and anomaly flags, each step reviewed by competing models and kept only when they agreed on the evidence.

Which cluster deserves attention is a judgement, not a score. The system ranks; the controller decides. Every flag carries its rule and its threshold, so the decision can be defended later.

Pick a region, a category or a risk tier and watch the control picture reorganise itself. Built for the question an audit committee actually asks: where should we look first, and why.

Arrivals returned above the 2019 peak, so the story looked like recovery. The trap: volume came back, behaviour didn't. The real question isn't how much travel recovered, but what broke and stayed broken.

ChatGPT, Gemini and Grok produced independent versions for comparative review, structured disagreement, not one source of truth. Claude held the architecture and ran visual validation before deploy.

The thesis, tourism doesn't bounce back, it mutates, is a judgement, not an output. Every claim is tiered by hand: Official, Derived or Proxy, with a confidence score the viewer can inspect.

Eight linked modules let a decision-maker do, not just read, and open the Evidence Drawer on any claim to see its source, tier and confidence. Nothing asks for blind trust.

193 nations, 150 years, 25 indicators across incompatible sources. Most development stories cheat: one chart that confirms the author's prior, frozen and called a trend. The trap is the single axis.

Dozens of authoritative datasets ingested through an automated pipeline and reconciled into one model: 194 countries × 25 indicators × 150 years. Every weaker indicator is flagged, not hidden.

Twenty-five indicators is a curation, not a dump. Secondary sources are marked, and every projection to 2050 is labelled projections, not commitments. Showing where data is modelled is the judgement.

Ask your own question of 150 years of data, press play, and inspect every indicator's source and coverage in Methodology & Sources, and know exactly how much each answer rests on.

A 4.8× recommendation looks robust on paper. Then reality intervenes, and the biggest number is often the one whose model is most wrong in your favour. The optimizer's curse.

Each of four parameters is estimated from imperfect evidence. The tool runs a hundred scenarios and maps the entire parameter landscape, the disagreement between the optimistic number and the realistic spread is made visible.

Knowing where a model breaks in the field isn't in the model. Twenty years across major UN agencies, procurement, supply chain, public health, is what flags the fragile assumption first.

Drag four sliders and watch a recommendation flip between Holds and breaks in real time. A funder sees, before committing, exactly which assumptions their confident number rests on.

Ninety-six years, results, hosts, economics, governance scandals, carbon, scattered across primary documents and contested estimates. Most tellings quote one tidy cost that averages a stadium bill and a national infrastructure programme as if they were the same thing.

Here the orchestration is of sources, not models. Every figure checked against primary documents. FIFA accounts, the Parliamentary Budget Officer, DOJ indictments, Our World in Data, and carried at its true confidence. Four tiers: Official, Estimate, Modelled, Snapshot.

The discipline is refusing to average things that aren't the same. The 2026 carbon footprint is shown as a range, no official figure exists. The uncertainty is part of the story, not an embarrassment to smooth over.

The flagship, The Ledger, asks who really pays, every figure sourced and labelled. Alongside it: Global Atlas, Outlier Hunter, The Mirror, a 3D Penalty Shootout. Proof the method scales into a beloved subject, and still refuses fake precision.

Hundreds of sites across inconsistent registers, half of them ruins, a quarter of them private. Lists exist; a way to see the pattern did not.

Public registers and open geodata reconciled into one dataset, deduplicated and tiered by source quality. The same discipline as the institutional work, applied to stone.

What to show by default is the curation. Visitable sites lead; uncertain locations are marked as such rather than placed with false precision.

Proof that the method travels: an audience with no interest in procurement explores evidence the same way a board does, by changing the criteria and watching the map answer.