Methodology

Evidence Score

Versioned algorithm evidence_score_v1.0.0 produces a 0–100 score with an inspectable breakdown. Components (maximum points before conflict adjustment):

• Source authority / reliability — 30
• Required-indicator completeness — 25
• Data freshness — 20
• Independent source diversity — 15
• Internal consistency — 10

A conflict penalty (up to 25 points subtracted) applies only when two or more current, comparable, non-superseded sources disagree materially (>5% relative) for the same indicator, period, geography and unit. Normal revision chains do not trigger conflict penalties. Demonstration-classified observations are excluded from national scoring inputs.

API responses include breakdown, sources[], coverage_pct, limitations[] and warnings[]. Low coverage lowers the score; it does not invent Critical severity.

Issue ranking

The National Situation Room orders issues deterministically:

1. Severity (Critical → Unknown)
2. Evidence score (descending)
3. Data freshness in days (ascending — fresher first)
4. Registry rank_hint (ascending)
5. Stable issue code (ascending)

Each row includes a ranking explanation. Programme drill-downs (for example correctional food-production) carry a high rank_hint and are excluded from the default national crisis table unless explicitly opened. Eight national issues are seeded in the current preview; this is not a claim of exhaustive national coverage.

Severity

Severity follows declared rules in the issue registry combined with indicator coverage. When coverage is below 25% or no non-demonstration evidence exists for required indicators, severity maps to Unknown — never automatic Critical. Partial coverage (40–75%) and complete coverage (≥75%) follow the issue's configured thresholds. Severity reflects evidence state, not model confidence alone.

Direction and trend

Trend requires comparable observations: same unit and geography, non-demonstration classification, and not superseded by a newer revision. With fewer than two comparable points, direction is unknown. With two points, period-over-period change is used. With three or more, an ordinary least-squares slope on period index is computed. Mixed units or geographies refuse a direction rather than forcing a misleading arrow.

Freshness

Freshness is expressed as days since the latest retrieved observation among non-demonstration evidence used for the issue or score. Publisher quality warnings surface on observations and in provenance drawers. Stale data reduces the freshness component of the evidence score but remains visible — it is not hidden or replaced.

Missing-data treatment

Gaps are first-class objects. The Missing Data Board aggregates registry gaps, required issue indicators, simulation input requirements and unapproved candidate sources. When a numeric estimate cannot be calculated from connected public inputs and declared formulas, Duplios displays exactly:

No reliable public estimate available.

This sentence appears for missing simulation outputs, unavailable cost-of-inaction figures, people-affected estimates without versioned inputs, and any metric where required observations or formula prerequisites are absent. Duplios does not impute, interpolate national totals from anecdotes, or use silent defaults in simulations.

Provenance

Observation-backed values expose organisation, dataset, version, observation identifier, period, geography, unit, classification, retrieval and publication times, transformation version, quality warnings and revision state. Users can open a provenance drawer from Situation Room and simulation surfaces. Formula-derived values carry formula ID, version, inputs and calculation context in API payloads and evidence snapshots.

Transformations

Ingest parsers and normalisers may convert publisher formats into canonical observations. When a transformation version is recorded, it appears in provenance. Transformations do not upgrade classification — a third-party or demonstration source remains labelled as such after parsing.

Derived metrics

Values computed from other observations use declared formulas with version identifiers. Derived metrics inherit input provenance and appear with formula context. If any required input is missing or fails validation, the derived output is withheld and the unavailable sentence is shown instead.

Simulation versioning

Models are registered with stable IDs and versions. The connected preview includes correctional_v1.0.0 — a pure-math correctional programme model. Inputs are validated; rates must fall in [0, 1]; negative prisoner counts are rejected. Each run persists to immutable simulation_run rows with model version, parameters, outputs and linked evidence snapshots. Other national models are not connected in this preview.

Assumptions

Simulations and severity rules declare assumptions explicitly — for example model coefficients that are not national statistics, or coverage thresholds that map evidence state to severity bands. Assumptions are documented in model modules and registry metadata; they can be reviewed and challenged. Duplios does not present assumptions as observed facts.

Limitations

Phase 2 preview limitations include: no calibrated national people-affected or cost-of-inaction coefficient tables; several issue indicator series not yet connected to live feeds; a descriptive Knowledge Graph that does not propagate national simulations; no AI Cabinet; SQLite in preview versus PostgreSQL in production-oriented paths; and fixture fallbacks where live connectors are blocked or incomplete. Limitations appear in API payloads, docs and on product pages — not only here.

Audit records

Registry seeds, graph edge create/deactivate actions, simulation runs and evidence snapshots write through the central audit service. Audit events are append-only in the preview listing. Model runs never bypass audit. This supports reproducibility — a reviewer can see what ran, when and with which version — but it is not a full security information and event management system.

Data conflicts

When current comparable sources disagree materially, conflicting values are preserved. Duplios does not silently prefer one publisher. The evidence score applies a conflict penalty and surfaces conflict pairs in the breakdown. Users should inspect both values and publisher context before citing a single number.

Revisions

Observations may link to a prior row via revision_of. Normal revision chains represent publisher updates or corrections — they are tracked in provenance and counted in limitations where relevant, but they do not automatically trigger conflict penalties. Superseded rows are excluded from current conflict detection.

Classification labels

Every observation carries one of: official, third_party, uploaded, inferred, demonstration. Labels are never visually merged. Official means published by an authoritative public body when connected live or mapped from an official series. Demonstration means fixture or curated demo content — never treated as current official national fact. See also the Data Policy page.

Fixture and demonstration data

When live machine-readable feeds are unavailable — for example due to connector blocks — curated fixtures may load with explicit source_mode and demonstration classification. Fixture-backed connector results are labelled demonstration when fixtures are in use. Marketing demo panels use qualitative labels and “Demonstration view” framing; they do not display invented national statistics as live facts.

What Duplios refuses to calculate

• National totals or costs without versioned formulas and connected inputs
• People-affected figures without declared, input-complete models
• Severity Critical from low coverage or absent evidence alone
• Trend directions from incompatible units or geographies
• Simulation outputs when required parameters are missing or invalid
• LLM-generated statistics presented as observed data
• Silent imputation of missing registry indicators or model inputs
• Graph-propagated national impacts — the Knowledge Graph is descriptive only in this phase

In all refused cases, users see No reliable public estimate available. or an explicit unknown/missing state — never a plausible guess dressed as official data.