Measured against AI and best humans.

Composite indices have a well-documented trust problem: a single number cannot be both the value and a claim about that value's reliability. Mature systems — IPCC AR6, the IMF Data Quality Assessment Framework, GRADE in medicine, and the US intelligence community's ICD-203 — deliberately separate the two. Meridian follows the same pattern, captured in ADR-002.

Provenance — SDMX OBS_STATUS. Each STEEPE cell ships with one of five codes drawn from the SDMX cross-domain code list, the convention every major statistics office (Eurostat, ECB, IMF, BIS, OECD) already uses:

• A — Normal. Direct aggregation of signal observations from the live ingest pipeline (FRED, World Bank, V-Dem, ACLED, OWID, WHO, OECD, and the 22 other connectors).
• E — Estimated. Composed from upstream observations via a documented methodology (weighted mean, UCM, or factor model).
• I — Imputed. Filled via MICE / last-observation-carried-forward where coverage is thin, with the imputation policy flagged on the cell.
• F — Forecast. Projected beyond available observations.
• X — SME anchor (Meridian-custom extension). Hand-authored with cited published source — used where no per-country signal pipeline exists (the four Tech dimensions: AI capability, Bitcoin adoption, CBDC rollout, Quantum readiness).

Confidence — IPCC AR6 five-tier rubric. Each cell carries an independent confidence label computed from coverage, freshness, and source agreement: very high · high · medium · low · very low. This is the same scale the IPCC uses for climate findings; a low label here means the underlying inputs are thin or disagree, not that the score is wrong — it's a calibrated trust signal.

The two axes are intentionally orthogonal. A signal-derived cell with stale or thin coverage can rank low confidence; an SME anchor against a strong published source (METR, Chainalysis, Atlantic Council CBDC Tracker) can rank medium. Combining them into a single tier would lose information.

Hybrid anchor seam. Historical years 2020 – 2024 carry hand-authored anchors with documented provenance (Edelman, Fragile States Index, ND-GAIN, OECD PISA). 2024-forward values are signal-derived where coverage permits. The seam consistency rule: derived 2024 must agree with anchored 2024 within ±5 points, or the derivation methodology is recalibrated before promotion — not the anchor quietly rewritten.

Tech dims — modeled, not measured. The four Tech dimensions (AI capability, Bitcoin adoption, CBDC rollout, Quantum readiness) have no per-country signal pipeline equivalent to FRED for economic or V-Dem for political; the underlying data is global-aggregate at best. They carry obs_status=X against named published anchors: METR for AI capability, Chainalysis Global Crypto Adoption Index for BTC, the Atlantic Council CBDC Tracker for CBDC, and Qureca + The Quantum Insider investment-program data for Quantum (no standardized quantum readiness index exists; government investment is the best available proxy). Their accountability runs through a different path: tech-dim skill is graded by Brier on resolved scenarios that turn on these dims, not by MAE on STEEPE projections — there is no per-country signal to backtest a projection against.

Track record is the actual accuracy claim. Provenance and confidence describe inputs. Accuracy lives in section 01: per-dimension MAE against the held-out backtest sweep, plus rolling Brier against resolved scenario outcomes. Methodology versions are tracked semver-style on every derived cell so a 2024 score remains comparable across releases.

Per-dim skill vs. naive year-to-year persistence at the 2020 → 2022, 2024, 2026 backtest sweep. Skill convention follows ECMWF / Atlanta Fed GDPNow: (naive_mae − forecast_mae) / naive_mae — positive means the projection beats “next year = this year”, the trivial baseline every honest forecasting platform reports against. Per-dim JSON for any STEEPE dim at /api/track-record/[dim]; rows below cover the dims currently promoted to the live UI from signal-derivation.

• economic — forecast MAE 8.10 vs. naive 11.61 → skill +30.2% (n=33; live on 11/11 zones)
• education — forecast MAE 5.80 vs. naive 10.24 → skill +43.4% (n=33; live on 11/11 zones)
• political — forecast MAE 6.80 vs. naive 16.03 → skill +57.6% (n=33; live on 11/11 zones)

Reliability diagram — live. The plot below bins every resolved economic-domain prediction by the probability we forecasted and plots that against the fraction that actually resolved YES. Perfectly calibrated forecasts sit on the diagonal: a 70% bin would resolve 70% of the time. Larger dots = more predictions in that bin. The summary scalar below the chart is Expected Calibration Error (ECE; Naeini 2015) — the bin-weighted mean gap from the diagonal. Source categories: money + economy templates. The other STEEPE dims do not yet carry enough resolved predictions to plot honestly — political is at 4 resolved, the tech dims at 0 by design (graded by Brier on resolved scenarios, not by reliability bins; see "Tech dims — modeled, not measured" above). Coverage expands as predictions resolve.

Methodology-version drift bands. Each row below shows one promoted (zone, dim) cell and how its value moved across every methodology version we've shipped. The light bar is the full range across our versions, the dark dot is the latest value — the source of truth. This is OUR drift, not an external peer comparison. No third-party forecaster publishes a continuously-scored 0–100 STEEPE surface against which to externally calibrate; the closest available retrospective indices (FSI, EIU Democracy Index, Bertelsmann Transformation Index) measure related concepts on different scales over historical years, not the forward forecast this page makes. Wide bands reflect genuine methodology refinement — we publish them rather than quietly overwriting prior versions. The synthetic seed version (0.0.1-synthetic) is excluded; single-version cells (no drift to plot) are omitted.

The technical name is a held-out backtest. The model only sees data on or before a chosen base year, projects forward to a target year it has never seen, and is then scored against what actually happened.

Current sweep: base year 2020, target years 2022, 2024, 2026. That produces 297 forecast-vs-actual pairs at the configuration behind the 5.61-point miss above. An expanded economy-only sweep over 2000–2024 base years (3,575 pairs, using V-Dem and World Bank WDI to backfill) is reported alongside, and produces the structural-vs-shock split shown earlier.

The harness lives in the production codebase and is reproducible on demand.

The score itself is the Brier score (Brier 1950): BS = mean((forecast − outcome)²). A rolling 7-day Brier runs continuously across all resolved predictions. A regression alert fires if it slips by more than 0.066 against the trailing baseline.

To find why a Brier score moves, we apply the Murphy decomposition (Murphy 1973), which splits the score into three independent ways a forecast can be wrong:

• Calibration. When you say 70%, does it happen 70% of the time?
• Resolution. Can you tell different cases apart, or is every forecast roughly the same?
• Sharpness. Are you saying something informative, or hedging at 50/50?

The technical name is a cross-family ensemble: five forecasters drawn from different model families. They're combined using a skew-adjusted aggregation method (Powell, Satopää, MacKay, Tetlock, 2024), an evolution of the techniques developed during the IARPA forecasting tournaments (Satopää et al. 2014).

Forecasters that miss more often get less weight, based on their track record. Not by reputation, not by how recent their work is.

We also run a second test that catches errors the Brier score misses: a model that is consistently off in one direction scores poorly on Brier but still ranks everything perfectly. Two rank-correlation metrics close that gap. Current sweep: rank correlation (ρ) = 0.906, tie-adjusted rank correlation (τ-b) = 0.759 across all 297 pairs. Both close to 1.0 means the rank ordering is nearly perfect.

The historical data is built from named, citable public sources, ingested through an audit-trailed pipeline. An ongoing recalibration process has applied 139 verified corrections to historical baselines to date, improving the ground truth against which every forecast is scored.

Each correction is dated, sourced, and linked to a specific historical cell. The audit register records what was changed, why, and against which public source the change resolves.

The forecast schema also aligns with ForecastBench (Karger et al. 2024), the open academic benchmark for forecasting systems. Concretely:

• Any prediction can be re-scored from scratch against ground truth.
• Different model versions are tracked separately, so improvements are not accidentally credited to old work.
• An external academic team can replicate any aggregate statistic this dashboard reports.

The futures cone you see on the time-branches view follows the framework popularised by Joseph Voros: it sorts scenarios into regions of plausibility — probable (high confidence), plausible (moderate), preferable (high-confidence + preferred-direction), and wildcard (low-confidence). It is ordinal, not numeric. The cone says where a scenario sits relative to others; it does not assign a probability mass to a point in scenario-space.

What it isn't:

• Not a Bayesian posterior. There is no prior, no likelihood update over data, no posterior density. Cone region membership is derived from authored scenario confidence plus rule-based cascade boosts, not from an inference machinery.
• Not a Monte Carlo distribution. The cone does not draw samples from a generative model. Each scenario carries one confidence score; the cone clusters scenarios into bands by that score.
• Not a quant-finance value-at-risk band. VaR bands are derived from historical volatility. The cone draws on authored scenarios plus signal-driven calibration; it speaks to discrete futures, not continuous distributions.

What it is: a strategic-planning tool for triaging which futures deserve serious thought. The probable region is where current trends point; the plausible region holds futures one shock away; the preferable region tracks the futures policy should aim for; the wildcard region flags tails worth pre-positioning against. Aggregate impact bars below the cone show the confidence-weighted mean of STEEPE deltas across scenarios in the active region, with the per-dim weighted standard deviation surfaced as a "scenarios disagree" badge when the mean masks material spread.

About the projection modifier. When you view the cone for a specific zone, projected scenarios are re-weighted by how well their authored STEEPE delta vector aligns with that zone's current STEEPE state — a dot-product over normalized values, clamped to a maximum ±15-point confidence shift. This shifts the scenario's effective confidence, not the scenario's authored impact. A scenario that says "techAI +22, economic +15" will deliver those same deltas if it fires; the modifier only changes how heavily that scenario's contribution is weighted into the aggregate cone delta and which cone region it lands in. The point is to focus strategic attention on the scenarios most aligned with where the zone actually stands, not to rewrite the scenarios' authored impact.

For probabilistic forecasts of specific resolvable claims, see the Brier-scored predictions track in section 01. The cone and the prediction track answer different questions: "where do strategic futures cluster?" versus "what's the probability this specific event resolves yes?"

Real-world futures are coupled: an AGI threshold crossing pulls automation deployment forward; a US debt crisis pulls Bitcoin sovereign-reserve scenarios upward; a CBDC pilot scale-up amplifies institutional-trust collapse. Meridian models these couplings as an explicit, auditable graph.

Tier 1 — scenario→scenario confidence boosts. src/lib/event-graph.ts defines roughly 150 directed edges of the form "when scenario A fires, scenario B's effective confidence shifts by N points." Each edge carries an optional rationale string. Example: cbdc-pilot-scaleup → trust-institutions-collapse (+15, "CBDC programmability erodes financial trust"); us-debt-crisis → bitcoin-sovereign-reserves (+25). Boosts compose additively, then the cone classifier reclassifies scenarios into regions using the updated confidence values.

Tier 2 — dim→dim cross-feedback. A second edge set handles cross-STEEPE-dimension propagation: when a dim's value shifts in one direction, related dims feel a fraction of that shift. Examples: education-collapse drags economic; social-erosion lifts political risk; AI surge amplifies economic displacement pressure. Each cross-edge has a multiplier and a per-edge magnitude cap so cascades cannot run away.

Tier 3 — PCMCI+ learned causal graph (offline). A Python sidecar fits causal structure over per-country signal observations using the Tigramite PCMCI+ algorithm (peter and momentary conditional independence). Results land in the ew_causal_graph Supabase table, joined into metric resolution for forward projection. This is the layer that handles the "does correlation X actually precede Y, or are both driven by Z?" question, on the long-horizon STEEPE projection side rather than the short-horizon scenario cascade.

The active scenario panel in the time-branches view surfaces the Tier-1 edges live: when you select a scenario, the cascade-chain card below the cone delta lists every scenario it amplifies and by how much. Every edge is inspectable in src/lib/event-graph.ts with its rationale comment.