Approach — Reeco Horizon

First principles

Four principles, applied without exception.

1 · Provenance for every metric

If we cannot show you the underlying documents, telemetry or filings behind a number, we will not publish that number. Every output in every framework links back to its source documents, their version, and the parse that produced the structured fact. We treat this as a non-negotiable. The cost of one bad citation, in a market where the next billion-pound decision rests on it, is too high to amortise.

2 · Uncertainty as a first-class output

Every probabilistic claim ships with a calibrated interval, not a single number. "We expect connection in Q3 2027" is not an output of this system. "P(connection by Q3 2027) = 0.42, 90% interval [0.31, 0.55], calibration class 'UK distribution-connected solar > 50 MW'" is. If we cannot calibrate, we say so — and we tell you in which direction the bias is most likely to run.

3 · Counterfactual reasoning, built in

An IC memo is not "here is the answer" — it is "here is the answer, and here is what the answer would have been under three reasonable alternative assumptions." We bake counterfactuals into every framework output. The reader sees the central estimate, the policy-prior sensitivity, the document-coverage sensitivity, and the model-class sensitivity. If the answer flips under any of them, that is the headline.

4 · Calibration discipline, publicly scored

Every quarter, we score our standing forecasts against actual outcomes and publish the Brier scores, calibration plots and reliability diagrams. Some quarters this is unflattering. We publish anyway. Without this discipline, "AI-powered horizon-scanning" is a marketing claim. With it, it is a verifiable service that survives audit.

Worked example · Brier 0.087 Our current 24-month Brier score on grid-code revisions across the UK, Germany, Ireland and Texas is 0.087, based on 318 closed forecasts since 2023. That number moves quarterly. We publish the calibration plots in the appendix to the methodology document and in the Library; the underlying outcome ledger is available to clients on request.

What we will not do

Four things you will never hear from this desk.

Single-number forecasts

"It will happen in Q3 2027." Not a Horizon output. We do not strip uncertainty out of probabilistic claims, even when the IC asks for a single number. We will give a most-likely point, but it ships with an interval, every time.

Off-the-record carbon claims

If a Scope 3 number is published without uncertainty intervals, methodology version and underlying activity data, it is not a Horizon Scope 3 number. We will not run a "best-effort" lane.

Backtests on closed forecasts

If we run a backtest, we publish the calibration plots and outcome ledger that go with it. We will not show you a hand-picked subset of correct calls without disclosing the denominator.

Confidential-by-default outputs

The standing methodologies, calibration data and outcome ledgers are open or shared with clients under straightforward terms. We do not hide the workings — the workings are the product.

How a pilot runs

Two weeks to scope. Six to first defensible output.

Week	Phase	What the desk does	What you receive
W0 – W2	Scope	Brief intake, jurisdictional scoping, success-metric definition, calibration target agreed.	One-page scope memo with calibration target.
W3 – W4	Corpus & baseline	Corpus ingest, retroactive backtest on the calibration target, baseline forecast.	Backtest report with calibration plots.
W5 – W6	Forward forecast	Forward probabilistic forecasts on the agreed target, with counterfactuals and provenance trail.	First defensible Horizon output, ready for IC.
W7+	Steady state	Continuous refresh on the target; quarterly Brier scoring; expansion into adjacent jurisdictions if scoped.	Monthly refresh, quarterly calibration review.

Begin a two-week scope