/ › expansion › now-next

Now / Next

Three streams in parallel. Each a 100× lever, each shipping concrete deliverables, each tracked with falsifiable success metrics.

46Citations · curated

5,486Citations · auto-extended

22LATAM countries

30indicators · WB

5named targets · Anchor 0 sprint

14dsprint window

Stream progress

Stream A · Data spine — WB ingestion live → 5,486 facts 60%

Stream B · Anchor 0 sprint — 5 named targets, 14-day plan 15%

Stream C · Conversational — Design + truth-gate spec 25%

Stream A — Data spine ignited

The Phase 2 charter shipped on /expansion/02-data-spine. Today we lit the first adapter: World Bank Indicators v2 ran across LATAM × top-30 and produced 5,486 typed facts, each auto-extended into _citations_scraped.yml with provenance chain, last-verified timestamp, drift-detection baseline, and license class.

Facts ingested per country

Top 12 LATAM countries by WB fact count · indicator coverage uneven across countries (data gaps for HTI / VEN).

What ran

Component	State
WB adapter (single + batch, feature-flagged)	✓ live
X / Twitter adapter (paid, dry-run only)	✓ scaffolded
`tools/ingestion/_citations_writer.py` · atomic YAML writer with drift-detection	✓ live
`tools/ingestion/run_batch.py` · driver decoupling adapters from registry writes	✓ live
5,486 World Bank facts ingested · 22 countries × 30 indicators × ~8 years	✓ done today
5,486 citation entries auto-extended into `_citations_scraped.yml`	✓ done today
DuckDB graph (`data/graph/knowledge.duckdb`) · backfill from registry	blocked · needs `pip install duckdb` auth
IMF SDMX, OECD SDMX, ECLAC, USGS adapters	spec'd, not built

Stream B — Anchor 0 distribution sprint

14 days · 5 named targets · falsifiable response targets. Cartes excluded (net-negative reputational); Peña has no documented channel.

14-day Gantt

5 named targets

Tier	Target	Role	Brief	Channel
A	Eduardo López	Google Cloud LATAM Lead	`02-google-deepmind-outreach`	Direct + LinkedIn
A	Sergio Díaz-Granados	Presidente CAF	CAF custom (financiamiento)	Carta institucional
A	Ilan Goldfajn	Presidente IDB	IDB custom (Anchor 0 + Anillo 0)	Vía rep IDB-PY
B	Gustavo Villate	MITIC Paraguay	`05-ande-mitic`	Vía IDB rep o Cancillería
B	Anthropic Policy	Frontier policy LATAM	`01-anthropic-outreach`	Coefficient-brokered

Falsifiable success metrics

3 / 5target · responses

≥2target · meetings scheduled

≥1target · LOI/MOU

0 / 5pésimo · re-evaluate package

Full plan: state/expansion/anchor-0-distribution-sprint.md.

Stream C — Conversational Futuros

Architecture

Cost per query (cache-aware)

Sonnet 4.6 · cache-hit $0.08 / Q

Sonnet 4.6 · cache-miss $0.55 / Q

Opus 4.7 · cache-hit $0.38 / Q

Opus 4.7 · cache-miss $2.30 / Q

Mixed 90/10 average · target $0.11 / Q

40 Q/day cap × $0.11/Q = $4.40/day, fits the existing $5/day daemon guardrail. Excess returns graceful 429.

Truth-gate flow

System-prompt instruction: cite every quantitative claim with [citation_id] or refuse.
Regex linter: every numeric token (USD, MWh, %, etc.) must have a citation anchor within ±1 line.
Citation existence check across _citations.yml ∪ _citations_scraped.yml.
On any failure → HTTP 200 with { truth_gate: "refused-no-anchored-evidence" }. Never fabricates.

Full design: state/expansion/conversational-futuros-design.md.

What ships next, in order

Authorize DuckDB → init knowledge.duckdb → backfill facts_v1 from the 5,486 facts already in _citations_scraped.yml.
Build IMF SDMX + OECD SDMX + ECLAC adapters → reach ~30K facts.
Resolve the 3 Anchor 0 VERIFY items → start sprint day 1.
Decide sidecar deploy target → ship /preguntar.
Ship MCP server at mcp.futuros.xyz.