Pillar 4 — Frontier Intelligence

The thesis. Give the private sector the intelligence to accelerate the frontier in Latin America: a terminal of comparable markets and indicators, BI exploration and analysis with deterministic provenance, post-AGI economy and compute-sovereignty lenses, the decision layer that turns verified data into investment and deployment of frontier technology.

It is the fourth pillar of the cycle: it consumes the outputs of the Sovereign Model (2 → 4), and the companies that profit from this intelligence pay the Data Trust back with data and fees (4 → 1).

In the grant frame

In the three-pillar proposal Futuros presents to funders, these surfaces appear as outputs of the Governance Engine (P3), the layer that turns the governed corpus into decisions. The platform canon remains the four-pillar cycle: the mapping between the two frames lives on the four-pillar cycle.

The subnational layer gained behavioural ADM1 and country series on 2026-10-15 (Anatel, Pix, Chile SII, Conab, CNJ Justiça em Números, ENOE TIL1, ICFES Saber 11, IMSS, PNCP, MapBiomas, INPE PRODES, OSM amenities, SECOP II, SESNSP crime, SIAP agriculture, TSE turnout, IBGE PNAD Contínua) with citations to the concrete record and no imputation; see the data pipeline.

The country-series layer adds in 2026-10-15 behavioural evidence (CEPALSTAT time use, UNESCO LLECE ERCE, GBIF occurrences, GFW tree cover loss, expanded GitHub Innovation Graph; lote 07: Glottolog, Comtrade HS, DPA, EIA IES, Ember, EM-DAT, GEM APS; lote 08: HRMI ESR, ITU DataHub v3, IUCN Red List Index, LAPOP, Latinobarómetro and Media Cloud Wayback; lote 09: ReliefWeb HDX, UCDP by violence type, WDPA via WDI, OpenAlex OA/SDG facets and Overture Places; and batch 10: OECD PISA, PSMSL sea level, IMF IFS via DBnomics and WIPO IP Statistics) with per-record citations and no imputation; see data pipeline.

The frontier-technology corpus

public/data/frontier-indicators/ no longer serves eight fabricated columns (provenance: "synthesised") and publishes 97 cited indicators across ten domains (AI/compute, semiconductors, nuclear, biotech, quantum, space, patents, research/talent, energy/datacenter, industry) for the 25 countries. Every value comes from the parameter-cache with its citation_id; a zero the source reports is tagged sourced-zero; a country with no source carries no-source and value: null, never an invented zero. The check-frontier-provenance gate enforces this in prebuild.

Typical bake coverage: on the order of 1,200 non-null valued rows, ~580 cited zeros and ~630 with no source. Live spots (October 2026): Brazil 1,884 MWe nuclear (IAEA CNPP), Mexico 1,552 MWe, Chile a cited zero; Epoch AI counts 8 clusters in Brazil and 1 in Mexico.

Seven measured absences (probe-confirmed, not filled): installed AI compute in FLOPs per country; IFR robot density for the 25 economies; submarine-cable capacity in Tbps (TeleGeography's free tier has no capacity field); public-sector AI deployment registries; deep-tech VC per country; ITU orbital slots; firm-level AI adoption. Also deliberately unwired: Overpass/OSM datacenters (ingest host gets :443 refused; a failed fetch is not a zero), cloud_regions (all three probed sources are redistribute: false), and the industry-deployment lane (no bakeable indicator). Detail: state/FRONTIER_CORPUS_SHIPPED.md. Chat answers from frontier-ranking / frontier-country / frontier-gaps; the PillarExtras pillar card shows four real indicators, not the fabricated ones.

The surface: /frontera

/frontera is the corpus page: ten domains, 25 countries, every figure tied to its source record. It does not rebuild /soberania-computo (readiness to host compute); it links to that lens. What it shows:

1. Corpus coverage (countries, indicators per country, cited values, fetched zeros, cells with no source, open sources).
2. Measured absences as headline content: cards that count how many countries answered zero and which ones do report a value (for example, zero frontier AI data centres in the region; three countries with operable power reactors).
3. The ten domains (by domain or by country), with per-indicator strips and a 25×10 coverage map.
4. Verdict on the previous corpus (retired / replaced / kept) and the source list with republication licence.

The three states are distinct at a glance; a real zero is never collapsed into an empty cell:

State	In the JSON	In the UI
Fetched value	status: "sourced", value ≠ 0, with citation_id	solid fill; the number opens its citation
Source zero	status: "sourced-zero", value: 0, with citation_id	flat wash; shows 0 and the citation
No source	status: "no-source", value: null, no citation	dashed hatch; shows —, never an invented zero

Why provenance makes the intelligence bankable

The pillar's central argument in one line: a cited figure is an investable figure. An analyst taking a number to an investment committee, or a development-bank officer structuring an instrument, cannot defend a figure from a black box. They can defend a figure that traces back in two clicks to its primary source. Deterministic provenance is not a compliance feature, it is what turns intelligence into something you can act on with real money.

That is why the same grounding discipline that serves a regulator (pillar 3) serves an investor (pillar 4): both drink from the same Sovereign Model grounded in the same verified corpus.

And the argument is not asked to be taken on faith: /confianza is the platform's public trust report, provenance coverage, freshness of every source, anomaly load, coverage gaps, expired sources and the corrections log. The claim "a cited figure is an investable figure" is auditable on that page.

The live surfaces

The hub: /inteligencia

/inteligencia: the gateway to the entire pillar: a hub for the BI suite (the analog of Bloomberg's Launchpad) that groups every analytical surface by the job it does, explore · screen and build · monitor · analytical lenses · trust and data. It makes the full terminal legible and discoverable instead of leaving it scattered across loose routes.

Markets terminal

/mercados: the markets screener: dozens of indicators comparable across the 25 countries, with ranking, as-of year, base-100 index, CSV export and watchlists; it integrates IMF WEO projections. The per-country view /mercados/:iso drills into a specific economy. It is the terminal from which an analyst builds a thesis with traceable figures.

The exact semantics of the transformations (all client-side, in src/routes/mercados.tsx):

Transformation	Semantics	Honesty guards
"As of" cut	each cell takes the latest point of the series with year ≤ the chosen year; if the series starts later, the cell stays empty ("—"), never interpolated	opens by default at min(current year − 1, latest observed WEO year) — the terminal never opens on projections
WEO projection boundary	derived from the data itself: horizon = max year of any WEO series; first projected year = horizon − 5; last observed = horizon − 6	self-corrects when a new vintage is re-ingested (horizon 2030 → observed 2024); every WEO cell with year ≥ first projected carries a projection mark, whether the slider moves or not
Rebase to 100	base = the series value at or before the base year (2015 by default); each point = v / base × 100, rounded to one decimal	a country with no base value, or base = 0, is excluded from the trajectory rather than plotted wrong
Ranking and heat	position within the cohort at the as-of year, respecting the indicator's good direction (if "lower is better", lower value = better rank); per-column min–max heat	the ≥ / ≤ filters live in the URL (f=id:op:val), so every filtered screen is shareable and reproducible as-is

BI exploration in the browser

/explorar: the self-service exploration workbench: the analyst crosses whichever metrics they want against whichever countries they want over a static dataset of 200+ metrics × 25 countries, with table, trend and scatter views; normalizations (raw value, base-100 index, z-score, year-over-year change); regional presets; cross-highlighting between views; named saved views; URL-shareable state; and CSV/SVG/PNG export.

The four normalizations have an exact, deterministic definition (src/routes/explorar.tsx), two analysts with the same URL see the same number:

Mode	Definition	Exclusions and guards	Reference line
Raw	the value in its original unit	—	—
Base-100 index	v / base × 100 to one decimal; base = latest value at or before the chosen base year	a series with no base, or base = 0, is omitted from the chart	y = 100
Z-score	per year, across the countries on screen: (v − mean) / σ (population σ), to two decimals	σ = 0 → divisor 1 (never infinite)	y = 0
Year-over-year (YoY)	% change vs the immediately preceding, consecutive year — a gap in the series cuts the computation, it never skips	no consecutive previous year, or previous value = 0 → empty point	y = 0

And for advanced analysis, a fourth view: a SQL console on DuckDB-WASM, loaded on demand when the tab opens, executed entirely in the browser, no backend: queries never leave the analyst's machine. Serious analysis, provenance intact.

Situation Room

/sala: the personalized, cited situation room for decision-makers: it aggregates (never duplicates) the single most important item from each analytical layer: instability, biggest mover, top anomaly, cross-source contradiction, nearest forecast, next election, rising attention, into one screen ordered by severity, with a link back to each full surface. Purely client-side assembly over already-baked JSON; zero backend, zero LLM until the user chooses to drill down in the chat.

Movements

/movimientos: the movement detector: which indicators moved, by how much and in which direction, to catch significant shifts across countries and dimensions without sweeping tables by hand.

Alerts and embeds

/alertas: the alerts interest list: you register an email (with optional country and topic) and a confirmation arrives; the weekly email digest is in preparation, and the page itself says so. What already operates today: an anomaly digest in JSON at GET /api/alerts-digest, and a Monday cron (0 13 * * 1) that distributes it to the configured webhooks.

The white-label embeds (/embed/board, /embed/stat, /embed/chart) let you embed boards, stats and charts with your own brand and accent on third-party sites (Futuros intelligence inside the client's workflow, not in a separate tab. Provenance travels with the embed. And there is no need to build the URL by hand: /incrustar is the self-service builder) pick parameter, country, indicator, type and theme, see the live iframe preview and copy the one-line snippet, always with the cited value and its source chip.

Strategic lenses

• /economia-agi: the post-AGI economy lens: the scarce complements in a world of abundant cognitive labor, with six cited dimensions, capital and net-position sliders, and scenarios. It helps an investor reason about where value accrues when intelligence gets cheap.
• /frontera: the technology frontier measured (ten domains, source zeros and absences without invention). Detail above in The surface: /frontera.
• /soberania-computo: the compute sovereignty index for 25 countries (clean energy, connectivity, financing and scale, with the binding constraint) and, on the same country, the cost of a sovereign AI stack for all 33 CELAC states. Three ambition tiers — access, capacity and full stack — plus a regional hub. The model subtracts GPUs, cables and IXPs that already exist. Every cost figure links to its source in the verified 15 Oct 2026 study. The same page sets the annual bill against public expenditure, education, health and R&D, the fiscal balance, the power system, a published H100 rental rate, and the country's EAR group.
• /dependencia: the trade and technology dependency analysis (on UN Comtrade data): what and whom each country depends on, to read geopolitical and supply-chain risk from a non-aligned position.
• /horizonte-2035: the backcasting module: it extends each country's 10-year historical slope of the composite index (least squares) out to 2035 and compares it against a user-set target. It is explicitly a trend extrapolation, not a forecast, the assumption is declared on the page and the underlying series is shown in full. The natural companion of the chat's forecast operation, applied to the pillar's underlying question: are we on trajectory for 2035?

Under-the-hood mechanisms

Two mechanisms with no route of their own that define what this terminal can promise:

• Deterministic forecast engine (server/chat/compute.ts): the platform's projections are not improvised by an LLM, they are computed by auditable code via OLS trend or CAGR. Every result comes out stamped projection: true and cited to the endpoints of its fit window, and the engine refuses to project with fewer than 3 historical points: it would rather not answer than fabricate a trend. The operation-by-operation detail is in the decision chain, just below.
• MCP institution feed (server/mcp/handler.ts): the governed corpus, exposed to third parties' AI agents (the internal copilot of a development bank or a ministry) through a closed exposure of 15 read-only tools. It is an allowlist, not a denylist: open-web tools are excluded and any new non-corpus tool fails closed by default. The institution asks with its own tools and every answer arrives with its citations. Connection guide: MCP server.

The chat's BI tools: the decision chain

The assistant (Sovereign Model) exposes BI tools that chain the data all the way to the decision, every result tied to its source:

• compute: deterministic computations over the series: change, cagr, rank, gap_to_frontier, correlation, forecast, explain_change. forecast projects by OLS trend (95% band from in-sample residuals, requires ≥3 historical points) or by CAGR (requires positive endpoints); the result comes out marked projection: true and the endpoints of the fit window carry their citation_id: the projection inherits the provenance. explain_change ranks co-movements: the Pearson r of each candidate indicator from the same country against the subject's series, and declares it explicitly as association, not causation.
• recommend_action: the pilot → financing → decision-maker chain (server/chat/tools-action.ts). Given a country (ISO3) and a pillar, it walks the corpus's own graph with deterministic joins, zero LLM in the chain:
  1. Pilot: filters pilots/_index.json by country (primary or alternate) and pillar (primary or tag); reproducible score = 0.6 primary pillar (0.3 tag) + 0.1 primary country + maturity as tiebreaker (scaling > mvp > concept). No match → returns pilot: null with the explicit instruction to state it as a gap, not to invent one.
  2. Financing: the pilot's own funding sources (with % share) plus its instruments, joined by slug against financing-instruments/_index.json: issuer, ticket range, eligibility and application URL.
  3. Decision-makers: personas/_index.json filtered by country and pillar affinity, discarding expired mandates; score = closeness to the decision (layer 1 > 2 > 3) + real-name holder over placeholder + relationship status (warm-intro +1 … hostile −1) + readiness + bonus if the office itself IS the pillar (health minister for salud) + executive bonus. Top 3, each with its own tailored_ask; an empty ask is returned as null so the chat drafts one from the office, never a blank.
• get_discoveries: the proactive radar, with two sources: anomalies from data-health.json (year-over-year jumps and drops, magnitude = |z-score|) and documented contradictions from pulse/contradictions.json (an open one weighs more than a resolved one). Ranked by magnitude, filterable by pillar and country, every finding with its citation_id. It is the tool behind "what should I be watching?".
• Plus the retrieval tools (search_corpus, get_dataset, resolve_metric) that keep every answer grounded.

The full chain, from indicator to decision-maker:

Every arrow is a join over governed corpus assets; citations travel with the result (the pilot's figures carry their citation_id, and capex/opex is declared as a design estimate unless the pilot is field-validated).

From the figure to the decision

The value of pillar 4 is not showing pretty data, it is closing the distance between a verified indicator and an investment or deployment decision. The pilot → financing → decision chain exists so that intelligence does not die in a dashboard, but becomes a bankable project.

Capital-allocator panel

/asignadores is pillar 4's consultation surface for large-scale capital allocators (and governance decision-makers). Three pieces, all primary-sourced:

1. Regulatory instrument deltas. What changed in a named instrument (law, circular, CMN, Banxico, BCB, CNBV or SFC resolution) versus its prior version, at provision level, with official permalinks to both versions. The delta is extracted from the authority's own amendment log (for example the update field of the BCB's normative system, or the resolutive sentence of a Banxico circular as reprinted in the DOF). When both versions cannot be verifiably obtained (a scanned PDF with no text layer, a JavaScript-generated listing), the record ships as "blocked" with its explicit reason: an honest "no verifiable delta" instead of an invented change. The card shows the most recent change, but the amendment chain it counts ("N recorded amendments") is the authority's complete log, never a trimmed list: a silent cap would turn an instrument amended 38 times into one amended 8. And because not all of those amendments could be parsed to provision level (many exist only in the official PDF), the card states how many carry provision-level detail, so the total is not read as if all of it were parsed.
2. Central-bank and GEMs citations. Deltas carry citations to the issuing central banks and regulators, and the GEMs consortium's default-and-recovery statistics publications are cited as context without reproducing figures that could not be extracted from the official PDF. The panel also lists the central-bank figures already cited in the platform corpus, with their original deep links.
3. Directory by policy axis. Cited institutions and public offices by country and axis (financial regulation, payment systems, monetary policy, open finance). An incumbent is named only when an official primary source vouches for it (for example the signature block of the resolution or circular in the official text); otherwise the entry stays at institution level. Not a CRM: public roles only, no personal-data harvesting.

Data lives in public/data/allocator/ (official snapshots + a deterministic bake validated by scripts/allocator/citation-validator.ts; the build fails if the published deltas stop reproducing byte for byte from the committed snapshots). The panel is not a terminal: no quotes, no order books, no real-time market data.

High-frequency series, not yet published

On 30 September 2026 the platform registered feeds that measure behaviour rather than annual institutional yearbooks: Banxico remittances and SPEI, the FIX rate, the policy rate, 28-day TIIE and reserves; WFP market prices by commodity, with no invented national index; CAMMESA monthly electricity demand; Ember's monthly generation mix (renewables, fossil, coal, demand and CO₂ intensity) beside the total-generation series that already existed; and daily PM2.5 and PM10 country means from the OpenAQ public archive. They live on the HF path (ingest-hf.ts). They are not served in the terminal: the publish gate still admits only NASA POWER, until a surface consumes them. A missing month, a station with too few hours, or a Banxico N/E cell is dropped. Nothing is interpolated.

The return to the cycle

The companies that profit from this intelligence pay the flywheel back with data and with fees (4 → 1): they fund the operation and bring their own datasets to the Data Trust, recruiting the next cohort of contributors. Frontier intelligence is not the end of the chain, it is what makes the cycle pay for itself.

Related surfaces

• /inteligencia: the hub that organizes the whole suite
• /mercados · /explorar · /movimientos · /sala
• /alertas · /incrustar · /embed/board · /confianza
• /economia-agi · /frontera · /soberania-computo · /dependencia · /horizonte-2035
• /incertidumbre: V-Dem's democracy indices with their credible interval as a shaded band: real measurement uncertainty, not a point pretending precision. /simulaciones: the frontier-intervention simulator: every persistent regional problem with its modeled pilot, re-runnable by adjusting scope, countries, capital mix and technology path.
• Technical reference for integrators: Public API v1, MCP server, embeds and white-label.

Return to the four-pillar cycle to see how this pillar recruits the flywheel's next turn.