Pillar 3 — Governance & Democracy (Governance Engine)
The thesis. Put model outputs at the service of democratic governance and the region's regulators: cited observatories of democratic resilience and regulation, election and transparency monitoring, verifiable situational intelligence — every figure with its source, every alert auditable. The region regulates from its own legal traditions, not with US or EU templates.
The Governance Engine. Seen as infrastructure, this pillar is the platform's decision engine: it takes verifiable indicators (the Democratic Resilience Index, the observatory's freedom layer, the Regulatory Coverage Index) and turns them into inputs for decisions: auditable alerts, machine-readable feeds, a forecast calibration board and a public corrections ledger. The chain is indicators → decisions, with every link traceable. One honesty caveat: today "verifiable outputs" means verifiable provenance, every figure traces to its source; the demonstrated track record of resolved forecasts and published corrections is still to be accumulated, and the two surfaces that will make it visible (the calibration board and the corrections ledger) are already live, starting from zero.
It is the third pillar of the cycle: it consumes the outputs of the Sovereign Model (2 → 3), and the institutions that come to depend on those outputs become contributors to the Data Trust (3 → 1).
From the region's legal traditions, not imported templates
The pillar's guiding principle is an explicit rejection: you do not take a US or EU AI governance framework and rename it. You start from Latin American legal traditions, the realities of the informal economy, and the region's constitutional frameworks. A regulation observatory that measures the region against a European template measures the wrong thing; the design starts from the region's law outward.
The live surfaces
Democratic Resilience Observatory
/gobernanza/observatorio — the Democratic Resilience Index, a cited composite of three sub-indices — rule of law, transparency and anti-corruption, democratic integrity — built on WGI, V-Dem, WJP, IDEA GSoD, BTI, TI-CPI and OWID, with goalpost normalization (p5–p95), equal weights and documented gap-filling. Beyond the ranking, the observatory offers five views: ranking, backsliding, authoritarian drift, freedom (an economic-freedom ranking on the Heritage Foundation index; the per-country detail adds layers for human, press and digital freedom, each with its own source) and an explorer. Every component of the index traces back to its source; the composite's construction, weights and assumptions are documented on its methodology page. It is not an opaque ranking: it is an index whose recipe is in plain sight and whose figures can be audited one by one.
How the index is built
The composite is baked with scripts/build-democracy-index.ts (bun run) and served statically from public/data/democracy/resilience.json; the page reads it, it does not recompute it.
The goalposts are computed over the full regional pool — one country's outlier does not define anyone else's scale. Sub-index composition (19 indicators expected per country):
| Sub-index | Indicators | Sources |
|---|---|---|
estado_derecho | 10 — wjp_rol_overall, WJP factors 1 and 8, vdem_rule_of_law, vdem_liberal_democracy, rl_est, ge_est, rq_est, idea_gsod_rule_law, bti_governance_index | WJP, V-Dem, WGI, IDEA GSoD, BTI |
anticorrupcion | 3 — ticpi_score, cc_est, owid_corruption_vdem | TI-CPI, WGI, OWID/V-Dem |
integridad_democratica | 6 — vdem_electoral_democracy, idea_gsod_representation, idea_gsod_rights, va_est, owid_political_regime, bti_democracy_status | V-Dem, IDEA GSoD, WGI, OWID, BTI |
Invariants of the script (all verifiable in the code):
| Rule | Exact value |
|---|---|
| Time window | 2010–2025 (YEAR_MIN/YEAR_MAX) |
| Per-year fill | valueAt carries the nearest observation prior to the year forward; if there is none, back-fill from the earliest |
| Trajectory | 5y Δ ≥ +2 → mejora; ≤ −2 → retroceso; otherwise estable |
| Confidence | observed/expected ≥ 0.85 → alta; ≥ 0.6 → media; less → baja |
| Quarantined caches | read as empty — they never feed the index |
| Binding constraint | the lowest-scoring sub-index (the reform priority) |
| Auditability | every contributing[] entry carries a citation_id → source record in citations.json |
25/25 coverage is guaranteed by design: Cuba (no WJP) or Suriname (thin V-Dem/BTI) appear with media/baja confidence instead of vanishing from the ranking. The backsliding board orders countries by 5-year composite drop and, for each one, identifies the individual indicator with the steepest normalized fall over ~5 years (the worst_indicator, with its citation). July 2026 cut of the baked file: 25/25 scored, regional median 51.9, best Uruguay, worst Venezuela, 12 countries backsliding and 4 improving.
Regulation observatory
/gobernanza/regulacion — the regulation tracker across the 10 axes in all 25 countries: more than 2,200 legal instruments (2,257 in the July 30, 2026 bake; laws, decrees, frameworks) — the count grows with every ingest — currently extracted from six connected catalogs: Constitute and Climate Change Laws of the World as external sources, plus the curated Futuros legal catalog (web-verified) and three internal project trackers. The source registry lists 18 catalogs in total; NATLEX, ECOLEX, FAOLEX, WIPO Lex and SHERLOC are already wired in src/lib/source-registry.ts (cite-only, with a deep link to the catalog) and pending ingest into the baked catalog; Digital Policy Alert and OECD.AI remain registered and pending. An instrument browse view at /gobernanza/regulacion/instrumentos searches and filters the flat index public/data/regulation/_instruments.json (country, pillar, type, status, free text). The Regulatory Coverage Index (ICR) has a public formula — 0.45·breadth + 0.30·force + 0.25·recency — and is computed only over verified instruments: synthetic records coming from internal project trackers are flagged (synthetic: true) and excluded from the index. A temporal layer shows how coverage evolves over time. Verified instruments are cited to their source, so a regulator can see not only what exists, but where the claim that it exists comes from.
The ICR formula, component by component
The ICR is computed in scripts/bake-regulation.ts and published to public/data/regulation/_analytics.json — the formula the page displays is the same one the code runs:
| Component | Weight | Operational definition |
|---|---|---|
| Breadth | 0.45 | pillars with ≥ 1 registered instrument / 10 × 100 |
| Force | 0.30 | pillars with framework_status in-effect or partial / 10 × 100 |
| Recency | 0.25 | pillars with an instrument enacted in the last 6 years / 10 × 100 |
Two honesty invariants protect the index: (1) an instrument without an official URL must be flagged synthetic: true — the bake discards any record that violates this rule — and (2) the country×pillar cell rollup that feeds the ICR filters synthetic !== true before counting, so no demo record moves the score. Each country row publishes its full inputs (amplitud, vigencia, actualidad, covered_pillars, inforce_pillars, recent_pillars, total_instruments, gaps, rank) — the score can be recomputed by hand. The ranking orders by ICR with covered pillars as tiebreaker; in the current cut Brazil leads with ICR 100 (10/10 on all three components, 289 instruments). The score does not reward raw volume: 289 instruments and 30 yield the same ICR if they cover the same pillars with the same force.
AI governance
/gobernanza/ia — who regulates AI in each LATAM country, today? The per-country map for all 25: data law, national AI strategy, AI-specific framework, regulatory sandbox and designated authority. It is perhaps the surface most aligned with the pillar's thesis: regulating AI from the region's own frameworks, not from imported templates.
Constitutional processes
/gobernanza/constitucional — the tracker of the region's constitutional processes: Chile's two constituent attempts, Mexico's judicial reform, Colombia's constituent debate. Constitutions get rewritten while nobody is watching; this surface watches.
Oracle: early reading of bills
/gobernanza/oracle — the per-country bill tracker: each bill is summarized, cross-referenced against the 10 Futuros axes and flagged as risk, conflict or opportunity. It is a demo — the page itself says so, and most of the bills are synthetic scenarios with no real sponsor. The surface demonstrates the early-reading format, not a live legislative feed.
Elections
/gobernanza/elecciones — the region's election monitoring: calendar, results of concluded elections with vote percentages and signals, with sources per event. It serves institutions and observers who need a comparable cross-country picture without giving up traceability for every figure. The calendar also exists as its own surface at /calendario: 130+ regional and electoral events across the 25 countries plus the regional category.
Transparency
/gobernanza/transparencia — the AI registry that Futuros proposes states adopt, shown with demonstration data: records of model executions, prompts and agent runs, in the format a public AI registry would require. The page itself warns: the data is illustrative and not yet connected to real telemetry. The surface exists to publish the format before asking for its adoption — practicing what it preaches — not to report institutional indicators.
Situation room
/sala — the situation room: a personalized aggregator that gathers, per layer (instability, indicator movements, anomalies, contradictions, forecasts, elections, attention), the most severe signal of the moment — all under a citation-integrity contract, so that every alert is auditable. It is the pillar's situational intelligence: not "the system says there is risk", but "this figure, from this source, moved this much, and here is the record".
Consensus
/consenso — deliberation by anonymous voting in the style of Pol.is: participants vote on 12 statements and the system aggregates the anonymous votes of all visitors into opinion groups (k-means, served by /api/consensus with an Upstash Redis store). One operational condition worth stating plainly: the cross-visitor aggregation is live only when the Upstash Redis environment variables are configured in the production deployment; without that store, the page degrades gracefully to individual self-positioning, local to the browser, with no cross-visitor aggregate. The two axes are editorially fixed policy tensions — sovereignty ↔ global integration and state ↔ market — they do not emerge from clustering as in real Pol.is. It is the channel through which civil society enters the governance pillar — not as observed data, but as deliberative voice.
Mechanics of the aggregate
A participant's vote is a 12-character vector of a|d|p (agree/disagree/pass), one per statement, regex-validated on the server. The store is a single Redis HASH (consensus:v1; the key version is rotated if the statement set changes) whose field is an anonymous id generated in localStorage — a returning browser updates its record instead of duplicating it. Anti-abuse: before inserting a new id the server checks HLEN against a cap of 50,000 participants (updating an existing id is always allowed: it does not inflate n), and reads are bounded to 20,000 vectors. The response distinguishes two failure modes — configured: false (no Redis) and degraded: true (Redis present, call failed) — and in both the page falls back to local self-positioning without breaking.
The clustering runs entirely on the client: each vector is projected onto the 2 axes (each axis sums 6 statements with ±1 sign; "pass" votes do not count; the score is the mean of vote×sign) and on those 2D points a k-means with k=3 runs, with deterministic seeds (points sorted by x then y, no RNG — the same set of votes always produces the same groups) and a maximum of 20 iterations; it requires ≥ 3 vectors to activate. On top of the groups, statements are classified (only those with ≥ 2 non-pass votes): consensus = spread of the agreement rate across groups ≤ 0.34 and overall extremity ≥ 15 points; divisive = the 3 with the largest spread across groups.
Proposals
/gobernanza/propuestas — AI-drafted policy proposals (Opus 4.7 models, with public prompts): each carries the Futuros position, the counter-position in its strongest form (steel-man) and local voting — support, reserve, oppose — that never leaves the browser. It is a demo, not a plebiscite — the page itself says so. The surface exists to show the format of AI-assisted deliberation, not to count real votes.
Priorities
/prioridades — the index according to your priorities: the visitor sets their own weights over the 10 pillars and watches the regional ranking recompute live — the mechanics of the OECD Better Life Index — with the weights shareable by URL (?w=). There is no single true ranking: the surface makes explicit that every composite index depends on its weights. All the computation runs on the client over the in-memory score cache, with no data files and no API.
Foresight: calibration board
/foresight is the forecast board with public calibration: every forecast carries a probability, a horizon and a falsification criterion, and the board groups resolved forecasts into calibration bins to show whether a platform "70%" actually happens close to 70% of the time. The page honestly separates two populations. The first is a seed file of predictions flagged illustrative: retroactively authored methodology examples, badged on every card, which do not count as a real track record. The second is the 13 forecasts generated by the social-pulse pipeline: a live mechanism with stale data, the only batch dates from May 6, 2026 and no expired forecast has been adjudicated yet. The demonstrated forecasting record is still to be built; the format and the calibration mechanics are already published.
Corrections ledger
The public corrections ledger, falsifications.json, is the platform's commitment to document every figure it published and later corrected: each entry requires the citation_id of a claim actually published, the prior value, the corrected value, the date, the reason and the link to the corrected source. The mechanism is live and surfaced at /confianza and /linaje. Its current state is declared without makeup: zero entries since the July 23, 2026 purge, when two illustrative demo entries were removed because they described corrections that never happened, against citation_ids that never existed. The ledger starts empty and only accrues real, verifiable corrections: an empty, honest ledger is worth more than one full of fictions.
API and institutional feeds
The Governance Engine does not end at dashboards: it publishes its data in formats an institution can consume directly, without scraping and without asking permission.
- Public API v1. The entire baked corpus is served statically under
/api/v1: observations in JSON and CSV, an SDMX-CSV export (ISO 17369,observations.sdmx.csv) for statistical offices, and an XLSX workbook (futuros-data.xlsx) with the citations included. Full reference in the public API guide. - MCP server. The
api/mcp.tsendpoint (POST https://futuros.xyz/api/mcp) exposes the cited corpus as Model Context Protocol tools: an institutional AI agent can query indicators, citations and geographies from Claude, Cursor or any MCP client. Guide in the MCP chapter. - Embeds. Cited charts can be embedded on a ministry's or a newsroom's site via iframe, keeping their citations. Guide in embeds and white-label.
Why "every alert auditable" is non-negotiable
The difference between a trustworthy democratic observatory and one that is not lies in whether an institutional actor can defend what the platform tells them. A regulator acting on an alert needs to be able to show the chain: which figure, from which source, measured when. That is why the pillar inherits the platform-wide two-click traceability contract — and why the situation room carries an explicit citation-integrity contract. An alert without a source is noise; an auditable alert is defensible action.
End-to-end pilots: roadmap
ROADMAP, not operational reality. The Futuros pilot corpus contains 67 designed proposals (count of /data/pilots/_index.json as of 2026-08-13; the live figure is the one /pilotos renders) built to a bankability standard: problem, design, costing, financing and KPIs, many with named ministers as potential counterparts. None of them runs operationally with an institution today: zero pilots in execution. The distinction matters, because the full Governance Engine (indicator → decision → intervention → measurement) will only be demonstrated when the first pilot runs end to end with a real institutional partner. Until then, the pilots are negotiation-ready designs, and this section declares itself a roadmap.
The return to the cycle
The institutions that come to depend on these observatories and on the situation room become contributors: they bring their own data to the Data Trust, closing the 3 → 1 loop. The observatory is not an endpoint — it is a data magnet that feeds the next turn of the flywheel.
Related surfaces
- Governance hub: /governance — the entry point to the entire pillar.
- /gobernanza/observatorio, /gobernanza/regulacion, /gobernanza/ia, /gobernanza/constitucional, /gobernanza/oracle, /gobernanza/elecciones, /gobernanza/transparencia, /gobernanza/propuestas.
- /sala, /calendario, /consenso, /prioridades.
- Verifiability and feeds: /foresight, /confianza, /linaje, public API, MCP, embeds.
Follow the cycle on to Pillar 4 — Frontier Intelligence, the sovereign model's other output.