The Strategic Preparedness Index (SPI)

Step 1 — variable progress (0–100)

Each tracked variable has a fixed baseline (its value when tracking began) and a programme target. Progress is the distance travelled from baseline toward target, clamped to 0–100:

• Higher-is-better variables (e.g. defence spending % GDP): 0% at baseline, 100% at target.
• Lower-is-better variables (e.g. gross debt, China export share): 0% at baseline, 100% at target — symmetric with the above, not a 50% midpoint. A metric sitting exactly at its baseline reads 0% (“no progress yet”), never a misleading halfway score.

The baseline is fixed at the point tracking began and never moves — this matters because an earlier version of the formula anchored to the live value instead, which silently ratcheted scores toward zero as ordinary reporting noise drifted the number. That bug was fixed; the fixed baseline is what makes the score honest rather than a moving target.

Step 2 — pillar score

Each pillar score is the weighted average of its variables' progress. Every variable gets full weight (1.0) exceptthose flagged “estimate”-confidence (author-assessed rather than sourced from an official dataset), which are down-weighted to 80% — so a shakier figure can still move a pillar, just less than a hard number.

Step 3 — overall score + the pessimism nudge

The overall score is a weighted average of all eight pillar scores. The weights (foundational pillars count more):

Weights sum to 8.9; the average is normalised by that total.

That weighted average is the calculated score — it is always derived from live variable data, never manually set. An administrator can apply one further, deliberately limited adjustment: a score nudge that can only move the published score downward(it is clamped to ≤0) and never up. It exists as an explicit, disclosed pessimism correction for known gaps in what's currently modelled — not a lever to make the headline number look better. The overall score is floored at 1 (it never displays 0, to avoid implying total, literal collapse).

The exact recipe

1. For every enabled cluster, compute its own ratio: mean(last 3 days) ÷ median(last ~90 days) of its daily volume — the identical ratio already shown per card on the Signals page.
2. Average the ratios within two categories: coverage clusters (press attention) and events clusters (CAMEO conflict-event counts).
3. Blend the two category averages with equal weight — 50% coverage / 50% events — into one combined ratio. If an entire category has no cluster with enough history, the gauge falls back to the other category alone rather than assuming the missing one is at a neutral “normal” ratio.
4. Map the combined ratio onto a 0–100 gauge: a ratio of 1 (both categories exactly at their own historical normal) reads 50; a ratio of 2 or more (double normal) caps the gauge at 100; a ratio of 0 floors it at 0.

Self-scoring

Coverage-direction calls resolve automatically from later GDELT data (did the volume actually go up or down); scenario probabilities resolve by human judgement at the 90-day horizon and are graded with a Brier score (a standard proper scoring rule for probabilistic forecasts — lower is better).

Ensemble forecasting — independent personas, aggregated

Each scenario probability is produced by four independent analyst personas — a base-rate statistician (anchors on historical frequency), an escalation hawk (weights current tension signals), a status-quo dove (expects regression to the mean), and an economic-structural realist (weights trade and fiscal constraints) — each answering the identical question independently, on a cheap model. The four numbers are combined with a trimmed mean(the highest and lowest replies are dropped before averaging, so one outlier persona can't dominate), and the spreadbetween the personas — how much they disagreed — is recorded alongside the result as an honest signal of how uncertain the call really is. Every forecast row's provenance records which model actually produced it, the aggregation method, the spread, and each persona's individual number and reasoning, so nothing here is a black box. If the ensemble can't be formed (an API outage, a parsing failure), the engine falls back to a single call rather than skip the week's forecast — and records honestly which path produced the row.

Two self-improvement mechanisms, both gated

Statistical calibration:a conservative correction derived from the engine's own resolved track record. Below 20 resolved scenario calls it is the identity — it changes nothing. Above that gate, it applies a heavily shrunk (weight 0.25) nudge toward the engine's observed historical accuracy, and every forecast row records whether the correction was active and by how much.

Reasoned lessons: after each forecast resolves, a separate process asks Claude to write a short post-mortem — a verdict (hit / miss / partial) plus a transferable lesson about why — and the accumulated, human-prunable lessons are injected into future forecast prompts. An operator can deactivate any lesson that turns out to be spurious. This complements calibration: calibration adjusts confidence, lessons adjust reasoning.