The plan — three tracks, 90 days

Hire a Paraguayan bilingual research analyst. Send the outreach emails. Get one cooperative meeting on the calendar.

Why this is #1

Every other goal — gauge data access, government endorsement, cooperative pilots, publication paths — depends on Paraguayan institutional reach. That reach currently passes through one collaborator who isn't moving consistently. Track 1 routes around that single point of failure.

What gets done in 90 days

By day	Outcome
7	Workana job posted. Itaipú & US Embassy emails sent. Trial-week analyst applicants reviewed.
14	Trial deliverable complete: 25+ named contacts at target Paraguayan institutions, with confidence scores and notes. Hire confirmed.
30	DMH waiver letter delivered. Itaipú email acknowledged. MAG email acknowledged. First weekly status report.
60	Itaipú technical call held. First cooperative call held (Cooperativa Colonias Unidas — most technologically sophisticated of the soybean coops).
90	Decision gate: how many of the four target doors are open? Extend or wrap based on result.

Cost

$500–1,500 for the analyst, paid through the Workana platform's escrow over 30–60 hours of work spread across 90 days.

What success looks like

By day 90: at least three of the four institutional doors (DMH, Itaipú, MAG, one cooperative) are at "in progress or better." That earns a 60-day extension to convert open doors into actual data-sharing or pilot agreements.

The current demo is excellent for a technical reviewer. For a cooperative agronomist, it's still a regional-aggregate RMSE story. Closing that gap is cheap and makes Track 1 conversations land harder.

What gets built (~3–5 days of my time)

• Per-cooperative probabilistic views. For each of the 5 priority cooperatives, a per-coop forecast panel showing probability of >1 mm, >5 mm, >10 mm precipitation at each lead day. Math is just sampling the existing distribution at that grid cell — no new modeling.
• Spanish translation of the demo. Top sections, headline numbers, captions, caveats. So that a Spanish-speaking stakeholder doesn't have to ask for a translator.
• "What this means for me" page. One short paragraph per audience type — cooperative agronomist, MAG planner, hydroelectric operator. Plain language, no acronyms. Concrete: this is what better forecasts let you do differently.
• Printable one-pager in Spanish (PDF, A4). Headline number, QR code to the demo, contact info, three sentences of context. The analyst hands this to stakeholders before any meeting.

Cost

$0. All time, no compute.

Why parallel and not sequential

Track 1 outreach lands harder when the analyst can show stakeholders something tailored to them, not a 25-km grid full of ERA5 acronyms. Both tracks need to be in motion at the same time.

A list of things that are tempting because they are familiar technical work, but the math doesn't pencil right now. Each one has a specific reason for the deferral.

Thing	Cost	Why defer
Run validation on 120 dates instead of 60	$5–8	Won't change the headline meaningfully. Pursue when a stakeholder specifically asks about temporal stability.
Add Atlas as a 4th ensemble member at full sample	$8–12	Adds robustness; doesn't unlock anything new.
Debug AIFS to add a 5th member	$5–15	Likely moves headline 0.5–2 percentage points. Not strategic.
Phase 5: kilometer-scale diffusion model (CorrDiff)	$600–20k	Real blocker is gauge data, not compute. Until Track 1 unlocks data, this is a science project, not a product unlock.
Write a methodology paper	~3 months of writing	Don't write a paper before a stakeholder is using the work. The paper without a real user is invisible.
Apply for grants	~3 months of effort	Don't pursue grants before a partner institution and a stakeholder testimonial.

The trap I'm avoiding

Doing more model work because it's familiar and feels productive. The model is fine. The gap is operational reach. Adding another ensemble member or another fifty validation dates won't get me one inch closer to a Paraguayan cooperative caring about this.

When does Track 3 unlock?

When a Track 1 conversation produces a specific question that requires the work. "We need probability of more than 25 mm in a 10-day window for our insurance product" — that justifies the heavy-event refit. "What about sub-25-km resolution for our irrigation district?" — that justifies a downscaling spike. Don't pre-build for hypothetical asks.