Project proposal · NASA Space Apps 2026

Aluni: NASA evidence for every field, tested against what really happened

Version 1.0 · 15 Nov 2026 · Challenge: Adapting Farms with NASA Data

Aluni is a live, open, bilingual decision-support tool that compares crop rotations for one field using nine NASA datasets. We backtested its calculations against 25 years of independent records, published the failures with the passes, and fixed the app where it was wrong. This proposal sets out the plan built on that evidence.

How to read the evidence labels
Label Meaning What may rest on it
[VERIFIED: source] Read in the cited repository file or web source, or computed by a script in the repository (section 11) Decisions and numbers
[VERIFIED (secondary): source] Read only on a reseller, directory or summary page Context; confirmed before any purchase
[UNKNOWN] Not yet verified. Listed in the open-items register (section 12) with the action that resolves it Nothing. No cost, date or go-decision depends on it
[ASSUMPTION A-n] A deliberate planning value, logged in section 12 with the value that would change the decision Illustration and budgeting only
[DESIGN] Our proposal (the vision), not a measured fact The plan; the pilot tests it
Swipe sideways to see the whole table.

1. Executive summary

The problem. Farmers make rotation decisions that shape their soil for years. The NASA evidence that could inform them lives in research portals, units and file formats built for scientists, and most farms cannot pay an agronomist to translate it. There are at least 570 million farms, more than 475 million of them under 2 ha [VERIFIED: FAO, Lowder et al. 2014], and erosion removes 25–40 billion tonnes of topsoil a year [VERIFIED: FAO SWSR 2015 via ReliefWeb].

What is live today at aluni.us. A farmer locates their field (from the device's location, a town search or the map), answers a short profile in English or Spanish with Aluni, the assistant, and gets their current rotation compared with at least three alternatives. Each strategy is scored on seven criteria and comes with why it was suggested, the evidence behind it (each value names the dataset that measured it), benefits, information gaps, risks to talk over with an adviser, and a three-season month-by-month timeline of bare and exposed soil. Nine NASA datasets feed it: POWER, GPM IMERG, SMAP L4, MODIS NDVI, ECOSTRESS, GRACE/GRACE-FO, Landsat (HLS), NASADEM and Worldview/GIBS [VERIFIED: config/field_profile.yml, docs/ROADMAP.md]. Sign-in is passwordless; the iOS and Android apps wrap the same site.

What we proved, and what we disproved. We ran Aluni's calculations on two real US fields for every season from 2001 to 2025: a humid one in Story County, Iowa, and a drier, rainfed one in York County, Nebraska. We scored them against records the app never reads: 136 rain gauges, county yields, crop insurance losses by cause, hundreds of published field trials and 13 years of satellite images. Pass marks were written before the results were seen and were the same on both fields (section 10):

  • Passed on both fields: the rotation advice (corn after soybean yields +9.9% [7.2 to 13.2%] in Iowa conditions and +12.7% [5.6 to 22.0%] in Nebraska conditions, at farm nitrogen rates) and the erosion ranking (it orders strategies as the RUSLE soil-loss model does; a rye cover saves 12–14% of the soil lost).
  • Passed on the rainfed field: NASA weather data at the pin; the water model, which tracked rainfed county yields (correlation 0.66 for corn, 0.77 for soybean); and choosing the crop by expected margin, which earned $752 per hectare more over 15 seasons than a farmer who had grown corn three years running.
  • Reported: the pest and disease weather rules point at the right season but none is statistically significant on either field, so they say when to scout, not that a crop will be hit.
  • Failed, then fixed in the app: NASA POWER rain read 12% below the gauges in Iowa (5% in Nebraska), so the water need now uses GPM IMERG rain, which passed on both; planting months now come from USDA crop progress, not a textbook table; the timeline now counts a young crop's open rows as exposed, which raised agreement with the satellite record from 76% to 90% (Iowa) and from 65% to 82% (Nebraska); the water model, which does not track yields on humid land, is shown as a drought risk, never as a harvest number.
  • Failed, still open: GAEZ suitability and water alone cannot rank replacement crops on either field, and in Iowa the margin rule earned $255 per hectare less than a farmer who already alternated strictly. Aluni helps most where a rotation has slipped; it does not beat a well-run one.

See the backtest evidence

Every number in this proposal can be rerun. The data are public and the scripts are in the repository. On 15 Nov 2026: the worked example passes its 24 checks, the production services reproduce its strategy totals exactly (82 / 68 / 60), and the claims audit recomputes all 52 audited figures in this proposal, both fields included, with 0 differences (section 11).

Where it goes [DESIGN]. From a rotation comparator to one season plan per field that the farmer runs with their own people, machines, supplies and budget, plus the Sonda, a low-cost field station that puts ground truth next to SMAP and IMERG. The work is gated: each stage starts only when the previous one shows evidence, and each has a written kill criterion (sections 15 and 24).

What the next 12 weeks cost. Gate 1, a two-country pilot (York, Nebraska and Buin, Chile) with 10 farmers and 10 engineering prototypes of the Sonda: $25,537.58, AACE Class 5 range $12,769–$51,075. Every line is a quantity times a sourced rate (section 17).

What we have not shown yet. No farmer has used Aluni in a season, so its value to farmers is untested; that is what Gate 1 measures. The backtest covers two US counties with county-level yields; Buin, Chile cannot be scored the same way, because no open source has Chilean rain gauges, county yields or crop insurance losses. The Sonda exists as a design and a 3D catalogue page, not as hardware (TRL 2).

Back to contents

2. How Aluni answers the challenge

The challenge asks for a decision-support tool that uses NASA Earth observations, local soil, crop traits and farmer priorities to explore crop rotations [VERIFIED: docs/ROADMAP.md]. Each requirement, where it lives in the app, and the independent test behind it:

Challenge requirements against the live app and the evidence
Requirement Where it is in Aluni (live) NASA and other data Independent check
Use NASA Earth observations, local soil, crop traits and farmer priorities Field profile and Your plan; every value names its dataset and resolution POWER, GPM IMERG, SMAP L4, MODIS, ECOSTRESS, GRACE, HLS (Landsat), NASADEM, GIBS; SSURGO, 3DEP and CDL in the US; FAO GAEZ and ECOCROP V1, both fields: daily highs within 1.7–1.9 °C and lows within 2.4–2.5 °C (RMSE), frost days agree 95–96%; IMERG rain within 15% of the gauges in 83% of seasons
Compare at least three strategies, including the current one Rotations::Comparator: the current rotation plus legume, small-grain, other-family and perennial strategies, seven criteria scored 0–100 and weighted by the farmer's goals Crop history from the USDA CDL; FAO-56 crop coefficients; ECOCROP tolerances V4, both fields: the rotation rules agree with the field trials (+9.9% and +12.7%); section 9: production services reproduce the worked example's totals exactly
Explain why, the evidence, benefits, information gaps and risks for local review Per strategy: why it was suggested (rule text, narrated by Gemini when on), evidence with sources, benefits, gaps, and a "Talk this over with your adviser" list Each line cites its dataset V3 and V7 failed honestly, so gaps (prices, yields) are shown as gaps, not filled
Show bare-field periods and vulnerable seasons Three-season month-by-month timeline: crop, establishing crop, bare soil, and storm, dry, frost and heat bands; cover-crop windows; observed NDVI on Your field from space IMERG heavy-rain days, POWER frost and ET0, HLS and MODIS NDVI V8: after the fix the timeline agrees with 13 years of satellite NDVI in 90% (Iowa) and 82% (Nebraska) of months; V6: the erosion ranking holds under RUSLE on both fields
Stay readable for farmers and support their judgment Plain-language copy in English and Spanish; scores explained; the current rotation is always shown and can win – At Ames the farmer's own rotation scored highest (82) and the field trials agree; the pilot measures readability (O1)
Swipe sideways to see the whole table.

Back to contents

3. Problem and strategic rationale

Farmers have to make season decisions that hold up for years, and the evidence and the records they need are spread across many places. Aluni's job is to bring them together.

Case for change

  • Scale. At least 570 million farms worldwide, more than 500 million family farms, and more than 475 million under 2 ha [VERIFIED: FAO ESA Working Paper 14-02, Lowder, Skoet, Singh 2014]. A later count gives 656 million farms in 2020 [VERIFIED: CIMMYT repository paper].
  • Soil loss. Erosion removes 25–40 billion t of topsoil a year and costs about 7.6 million t of cereal a year. If nothing changes, cereal losses reach about 253 million t by 2050 [VERIFIED: FAO SWSR 2015 via ReliefWeb]. FAO put the possible yield loss at up to 10% by 2050 [VERIFIED: UN News].
  • Getting worse. The FAO Status of the World's Soil Resources 2026 was launched on 27 Aug 2026 and calls erosion the most severe threat [VERIFIED: FAO Global Soil Partnership]. 65% of assessments report a deteriorating trend since 2015, and 81% rate erosion management poor or very poor [VERIFIED (secondary): DevelopmentAid on FAO SWSR 2026].
  • Access. NASA data lives in scientific portals, file formats and units built for researchers, and small and medium farms often cannot pay agronomists [a team claim, not yet measured; the pilot interviews test it (U-06)].
  • Why now. NASA point services (POWER, AppEEARS) return per-field values with no GIS skills, LTE-M/GNSS modules cost about £19–23 each [VERIFIED: RS UK], and language models can narrate computed facts under guardrails; Aluni's Gemini narration is live with a daily cap of 2,000 requests [VERIFIED: docs/ROADMAP.md].

Business as usual (fair version). BAU is not "nothing". Farmers rotate by habit and adviser input. Large farms buy commercial platforms (section 13). Others keep records on paper or in spreadsheets, and extension offices give advice where they reach. BAU's weakness is the break between evidence (satellites, soil surveys) and execution (labour, machines, inputs, cash). BAU cost to the farmer: [UNKNOWN]; it needs pilot interviews (U-06). BAU is also a high bar: at Ames the farmer's strict corn–soybean alternation beat a margin-switching rule over 18 years, while at York, where the farmer had drifted into corn after corn, the same rule earned $752 per hectare more (section 10, V7). Aluni has to earn its place next to experience, not replace it.

Theory of change [DESIGN]: per-field NASA and soil evidence, plus farmer goals → a ranked rotation with exposed months flagged (built) → a season execution plan with tasks, machines, inputs and costs (vision) → the farmer runs it and records what happened (vision) → next season recomputes from real outcomes (partly built: "Use as plan" saves next season's crop [VERIFIED: docs/FIELDS.md]). The key untested assumption: farmers will record operations if the plan pre-fills them. Kill criterion K1 tests it.

Strategic fit. It answers every item in the challenge brief, item by item, in the section above [VERIFIED: docs/ROADMAP.md, challenge requirements]. It also matches FAO's 2026 call for the adoption of sustainable soil management [VERIFIED: FAO GSP page].

Back to contents

4. Objectives and success criteria

The objectives are outcomes for farmers, not features. Baselines are missing because no farmer has used Aluni yet, so every baseline is set during Gate 1.

Objectives O1 to O5
# Objective (SMART) Baseline Target By Measured with
O1 New farmers get from landing to a usable plan quickly [UNKNOWN] under 5 min; 4 of 5 without help [VERIFIED: team overview §17] Gate 1 Usability sessions, en and es
O2 Fewer exposed bare-soil months in the chosen plan Method run on Ames: 12 vulnerable months over 3 seasons, 4 with cover (section 9); the app now also counts a young crop's open rows as exposed (V8 fix, section 10); pilot fields measured at enrolment [DESIGN] −50% against the current rotation 1 season after Gate 1 Rotations::Timeline exposed-month count; NDVI bare windows
O3 The plan is executed, not only read 0 (module not built) [DESIGN] 3 of 10 pilot farmers log operations weekly Gate 2 + 6 weeks Task and operation logs
O4 The farmer can see the cost of each strategy 0 (prices are always a gap [VERIFIED: docs/FIELDS.md]) [DESIGN] cost per ha shown for all strategies in 2 countries Gate 3 Budget module coverage
O5 Ground readings sit next to satellite readings 0 (hardware not built) [DESIGN] 10 Sonda units paired with SMAP and IMERG for 60 days Hardware DVT device_readings beside the observations
Swipe sideways to see the whole table.

Back to contents

5. Scope

In scope (end goal): the eleven platform modules in section 7 and the Sonda station (Lite and Pro) in section 8. Global coverage with richer US layers, two languages, web plus iOS and Android shells.

Out of scope until a later gate [DESIGN]: farm accounting and tax (Farmbrite and others already do this, section 13); machine guidance and variable-rate prescriptions; livestock management beyond the profile list; payments and marketplaces; drones.

Interfaces: NASA POWER, AppEEARS, Earthdata, Planetary Computer, FAO GAEZ, SSURGO, USGS, CDL, Meteomatics, Gemini on Vertex [VERIFIED: docs/FIELDS.md, CLAUDE.md]. Planned: price, pest and nutrient sources already documented in docs/data_sources/ [VERIFIED: docs/DATA_SOURCES.md], and the Sonda ingestion API [VERIFIED: docs/DEVICES.md].

Back to contents

6. Options analysis

The recommendation [DESIGN] is to build the integrated platform in stages (B) and to treat the Sonda (C) as a separate, gated track. Option A (lean on an existing open-source farm system for records) is the main alternative, and the team should choose between A and B at Gate 1.

Longlist considered: BAU; do-minimum (keep the comparator only); A, comparator plus integration with farmOS for operations records; B, an integrated platform built in the current Rails app; C, B plus own hardware; D, license a commercial platform and add the NASA layer to it. D is dropped: no commercial platform's API terms were read [UNKNOWN], and it gives up the open-source and free-data objectives.

Shortlisted options against the objectives
Criterion BAU Do-minimum A · Comparator + farmOS B · Integrated platform (preferred) C · B + Sonda
Meets O1 (fast plan) No Yes (built) Yes Yes Yes
Meets O2 (fewer exposed months) No Partly: shows them, cannot track execution Yes, if farmers log in farmOS Yes Yes
Meets O3 (plan executed) No No Yes, but across two apps Yes, in one app Yes
Meets O4 (cost per strategy) No No Partly: farmOS logs carry no prices by default [UNKNOWN] Yes (Gate 3) Yes
Meets O5 (ground vs satellite) No No farmOS has sensor support [VERIFIED: Drop Times project page]; pairing [UNKNOWN] No Yes
Whole-life cost Farmer's time, advisers [UNKNOWN] $0 on credits [VERIFIED: docs/deploy/DEPLOY.md]; $24.46/month VM after (section 17) Low build; Drupal hosting [UNKNOWN] Gate 1 $25,538 (section 17); later gates priced at G1 Highest: Sonda Pro planning BOM $493.73/unit [ASSUMPTION A-4], certification, connectivity
Key risk Status quo erosion A plan that is never executed Two stacks, two logins, Drupal skills Scope creep into a generic farm-management tool Hardware cost and certification
Evidence status Weak Strong (live) Medium Medium Weak (TRL 2)
Swipe sideways to see the whole table.

Why B over A [DESIGN]: Aluni's season data already lives in its own models (fields, crop zones, timeline months, observations) [VERIFIED: docs/DATA_MODEL.md, docs/FIELDS.md]. Tasks, machines and inputs need to hang off those same months and zones, which is easier in one data model. A stays the fallback if team capacity is the binding limit. farmOS is GPL and models equipment, maintenance logs and inventory [VERIFIED: farmos.org guide], so it is also a pattern to copy.

Why C is gated, not cancelled: the challenge rewards satellite-ground comparison, and the design is ready [VERIFIED: docs/DEVICES.md]. Hardware is the costliest and least mature part (TRL 2), so it only moves past EVT with pilot demand (section 24).

Switching values (what would flip the choice) [DESIGN]: if fewer than 3 of 10 pilot farmers log operations weekly, B falls back to do-minimum. If farmOS integration costs under half of B's build effort, A wins. If the Sonda Pro BOM goes above the price ceiling the team sets, C stops after EVT.

Back to contents

7. Technical design: the end-goal platform

The end goal is one season plan per field that every module reads from and writes to. Rotation, weather and satellite evidence decide what to grow. People, machines, supplies and money decide whether it can be done. Sensors and records show what actually happened. Six modules exist today, one is designed (the Sonda), and four are vision.

The integrating idea: the Season Execution Plan [DESIGN]. When a farmer taps "Use as plan" (built), Aluni expands the chosen strategy into a dated list of field operations for each crop zone: prepare, sow, fertilise, spray, irrigate, cover-crop, harvest. Each operation carries its window (from the timeline's months and the 7-day forecast), the machine it needs, the inputs and quantities, the people and hours, and its cost. Done operations are logged against the same zone, and the harvest closes the loop. Today the plan stores only next season's crop [VERIFIED: docs/FIELDS.md, Crops::PlanService].

Aluni end-goal architecture · 11 modules around one season plan

Evidence on the left and resources on the right both meet in the season plan. Solid boxes are live today; dashed boxes are the Sonda and the four vision modules.

The eleven modules
# Module Built today End goal [DESIGN] Main data and APIs Status
1 Land and Fields Pins, area, irrigation, water source, soil class; up to six crop zones on a 10×10 grid; field workspace [VERIFIED: docs/FIELDS.md, DATA_MODEL.md] Field boundaries (the boundary column is empty until a map exists [VERIFIED: DATA_MODEL.md]); tenure and lease terms; parcel documents; soil lab results SSURGO, 3DEP, NASADEM, CDL, GIBS, FCC Built, extend
2 Crops and Rotation Planner Current rotation plus 4 strategies, 7 criteria scored 0–100, 3-season timeline, scenarios (+1/+2 °C, −20% rain), Gemini "why" [VERIFIED: ROADMAP.md] Own yields recorded (needs a column [VERIFIED: FIELDS.md]); GEOGLAM crop calendars; gross margin as an 8th criterion GAEZ, ECOCROP, FAO-56, POWER, IMERG, SMAP, MODIS, ECOSTRESS, GRACE Built, extend
3 Weather and Risk 7-day Meteomatics forecast; daily frost, heat, heavy-rain and dry-spell alerts; push and email wait for switches and the shells [VERIFIED: ROADMAP.md] Pest and disease weather rules; alerts tied to scheduled operations ("spray window closes Thursday") Meteomatics, POWER hourly; pest rules documented [VERIFIED: data_sources/pest_weather_rules.md] Built, extend
4 Field from Space Monthly NDVI (Landsat, Sentinel-2), MODIS normal, bare windows, NDVI map [VERIFIED: ROADMAP.md] Observed bare months against the planned timeline (listed as next [VERIFIED: ROADMAP.md]); anomaly alerts Planetary Computer, AppEEARS Built, extend
5 Reports and Sharing Season summary, GAEZ attainable harvest, CSV/JSON export [VERIFIED: FIELDS.md] Read-only plan link for an adviser or extension office; cooperative roll-ups Internal Built, extend
6 People and Tasks Profile fields only: role, team type, team size [VERIFIED: farmer_profile.rb]. "Team members and tasks need their own models" [VERIFIED: FIELDS.md] Several users per farm with roles (owner, manager, worker, advisor already exist as profile roles); tasks generated from the plan, assigned, timed; hours, wage rate, labour cost; workload score from real hours Passkeys per member (existing auth) Vision
7 Equipment and Machinery A checklist of 8 types: hand tools, draft animals, tractor, planter, sprayer, harvester, irrigation system, drone [VERIFIED: farmer_profile.rb]. The workload score counts a combine for small grains [VERIFIED: FIELDS.md] Asset registry (make, model, year, hours, location); maintenance logs and reminders; use per operation; fuel; ownership vs hire cost; strategies checked against machines actually owned Manual entry; machine-data import [UNKNOWN which standard] Vision
8 Supplies and Inputs Nothing Stock of seed, fertiliser, crop protection, fuel, cover-crop seed; consumption per operation; reorder points before a sowing window; nitrogen balance per zone Nutrient removal sources documented [VERIFIED: data_sources/nutrient_removal.md] Vision
9 Costs, Budget and Markets Nothing. Market prices are a gap in every strategy [VERIFIED: FIELDS.md]. The profile has budget level low/high [VERIFIED: DATA_MODEL.md] Cost per ha per strategy (inputs + machines + labour); price feeds; gross margin; monthly cash need on the timeline FAOSTAT, AMS, NASS, WFP, World Bank and others, with licence verdicts already checked [VERIFIED: DATA_SOURCES.md]; most keys not configured Vision
10 Field Station (Sonda) Catalogue page with a 3D viewer only [VERIFIED: DEVICES.md] Ground soil moisture, temperature and rain next to SMAP, IMERG, POWER; device card, alerts (section 8) POST /api/v1/devices/readings (to build) Designed
11 Integrations Hub 13 sources fetched per field pin, all through one client with retries and an operator pause switch [VERIFIED: FIELDS.md] Sources as plug-ins with status per farm; credentials per source; partner exports (adviser, cooperative, farmOS) Index of about 60 documented sources [VERIFIED: docs/data_sources/] Built, extend
Swipe sideways to see the whole table.

Architecture fit. The current architecture absorbs the new modules without a rewrite [DESIGN]. Each module is one Rails area with service objects and Turbo screens, following the house rules (thin controllers, Result objects, every outside call through the switchboard) [VERIFIED: CLAUDE.md]. New tables follow the existing conventions (money as integer minor units or decimal(10,2), never Float; integer enums starting at unknown) [VERIFIED: DATA_MODEL.md]. Proposed new tables [DESIGN]: farm_members, tasks, operations, equipment_assets, maintenance_logs, supplies, stock_movements, cost_items, price_quotes, devices, device_readings (the last two are already specified [VERIFIED: DEVICES.md]).

Data sources in production. 14 sources are connected in the field registry, 9 of them from NASA (POWER, GPM IMERG, SMAP L4, MODIS NDVI, ECOSTRESS, GRACE/GRACE-FO, HLS, NASADEM, Worldview/GIBS); FAO-56 and ECOCROP are reference tables; MODIS phenology and OpenET are planned [VERIFIED: config/field_profile.yml]. SMAP, MODIS and ECOSTRESS arrive through AppEEARS, which runs one task at a time per account, so a new field gets them within an hour or two (R6).

Back to contents

8. Sonda field station: modules, cost, manufacturing

The Sonda is fully specified on paper (docs/DEVICES.md) but not yet built: only the catalogue page with a 3D viewer exists in production [VERIFIED: docs/DEVICES.md]. Roughly half of the core electronics have verified distributor unit prices, while soil probes, power components, enclosure mechanics and market-by-market certification paths remain open (U-07 to U-11, U-21, U-22). Consequently, no final unit BOM or retail price is stated at this Strategic Outline Case (SOC) stage. The station is designed as a single global product: the same hardware, firmware and backend serve any agricultural region, with only the connectivity variant (Lite or Pro) and the regulatory approvals differing by deployment market.

1. Expected end-to-end operation [DESIGN] + [VERIFIED: docs/DEVICES.md]

The Sonda operates as an autonomous agrometeorological field station that turns subsoil and microclimate measurements into actionable decisions inside Aluni's season plan. Operation follows four stages:

  • Field setup (once, about 2 minutes). The farmer mounts the weatherproof enclosure on a post inside the target crop zone and buries two capacitive soil probes at 10 cm (seedbed and evaporation zone) and 30 cm (active root zone). Scanning the exterior QR label in the Aluni app validates a one-time claim token and binds the hardware (device_id) to the field pin or crop zone [VERIFIED: docs/DEVICES.md].
  • Edge acquisition (every 15 minutes). Powered by a 5 W solar panel and a LiFePO4 battery, the station stays in deep sleep and is woken by the RV-3028 real-time clock to sample soil moisture and temperature at 10 and 30 cm, air temperature and relative humidity (SHT45), plus cumulative rainfall and solar radiation on the Pro model [VERIFIED: docs/DEVICES.md]. Each sample is written to a 16 MB SPI NOR flash buffer with a monotonically increasing sequence number (seq) and an RTC timestamp, so no data is lost during multi-day WiFi or cellular outages [VERIFIED: docs/DEVICES.md]. Timestamps are stored in UTC and converted to local time from the field's location, so the same firmware runs in any time zone [DESIGN].
  • Secure telemetry (every 60 minutes). The Sonda batches pending readings, signs the batch inside its ATECC608B secure element (ECDSA P-256) and uploads it over TLS through local WiFi (Lite) or LTE-M/NB-IoT with a soldered multi-carrier eSIM (Pro) [VERIFIED: docs/DEVICES.md]. Flash is freed only up to the highest seq acknowledged by the server [VERIFIED: docs/DEVICES.md].
  • Cloud pairing and action (continuous). Readings are ingested through POST /api/v1/devices/readings, validated (signature, seq, clock skew ≤ 10 min) and plotted side by side with NASA SMAP, IMERG, POWER and Meteomatics series in the "Your sensor next to the satellite" view [VERIFIED: docs/DEVICES.md]. When readings cross agronomic thresholds (water stress, rainfall that makes a planned irrigation unnecessary, frost, heat or disease risk), Aluni updates tasks in the Season Execution Plan and sends alerts [DESIGN].
Figure 8.1 · Sonda end-to-end flow in four stages
Stage Cadence Key operations and data flow
Stage 1 · Field Setup Once (~2 min) Post mount in crop zone; bury 10 cm and 30 cm capacitive probes; scan QR label in the Aluni app to bind device_id to field pin or zone
Stage 2 · Edge Acquisition Every 15 min RV-3028 RTC wakes ESP32-S3; sample soil (10/30 cm) and SHT45 air T/RH; write to 16 MB SPI NOR flash buffer with monotonic seq and UTC timestamp
Stage 3 · Secure Telemetry Every 60 min Batch pending samples; sign in ATECC608B (ECDSA P-256); upload over TLS via WiFi (Lite) or LTE-M/NB-IoT with soldered eSIM (Pro); free flash up to server ACK
Stage 4 · Cloud Pairing & Action Continuous Ingest via POST /api/v1/devices/readings; validate signature, seq and clock skew ≤ 10 min; store in device_readings; pair against SMAP, IMERG, POWER and Meteomatics; update Season Execution Plan and send alerts
Swipe sideways to see the whole table.

2. Scientific rationale: closing the satellite-to-ground gap [DESIGN]

Earth observation datasets give Aluni global, zero-cost coverage, but their spatial resolution, sensing depth and revisit time leave blind spots at the field-zone scale. The Sonda is designed to ground-truth and downscale four data streams already active in Aluni [VERIFIED: docs/ROADMAP.md, docs/DEVICES.md]:

What each satellite source misses at field scale, and what the Sonda measures
Variable Satellite / gridded source Limitation at field scale Sonda measurement
Soil moisture NASA SMAP (~9–36 km) Senses only the top ~5 cm of soil Capacitive moisture and temperature at 10 cm and 30 cm [VERIFIED: docs/DEVICES.md], validating root-zone water where SMAP cannot see
Precipitation NASA IMERG (~0.1°, ~10 km) Can miss localized convective rainfall 0.2 mm tipping-bucket rain gauge, Pro only [VERIFIED: docs/DEVICES.md], separating rain from irrigation at field level
Microclimate NASA POWER (~50 km); Meteomatics (~1–9 km) Grid cells average out local frost pockets, slopes and canopy effects On-site air T/RH (SHT45), solar radiation (Pro) and optional leaf wetness [VERIFIED: docs/DEVICES.md], driving frost, heat-stress and pest/disease rules (pest_weather_rules.md)
Crop water stress NASA ECOSTRESS (~70 m) Intermittent overpasses leave temporal gaps Continuous 15-minute soil and air time series [VERIFIED: docs/DEVICES.md], filling gaps between overpasses for irrigation timing
Swipe sideways to see the whole table.

3. Hardware architecture: Lite vs. Pro [VERIFIED: docs/DEVICES.md]

Both variants share a common solar-powered, weatherproof core and differ only in backhaul connectivity, positioning and meteorological expansion:

  • Common core platform (Lite and Pro): 5 W solar panel → MPPT charger → LiFePO4 battery → regulated rails; Espressif ESP32-S3-WROOM-1 MCU (WiFi 2.4 GHz + BLE); Microchip ATECC608B secure element (ECDSA P-256); 16 MB SPI NOR flash buffer; RV-3028 RTC; IP67 UV-stabilised polycarbonate enclosure with vent, M12 sealed probe connectors and pole mount. Samples every 15 minutes and uploads hourly in batches.
  • Sonda Lite: for fields within reach of a local WiFi network. Uses WiFi 2.4 GHz plus Bluetooth for local setup, inherits its geolocation from the field pin in the Aluni app, and carries the base sensor set (soil at 10 and 30 cm, air T/RH).
  • Sonda Pro: for remote fields that depend on cellular IoT coverage. Adds a Quectel BG95-M3 module (LTE-M / NB-IoT with built-in GNSS), a soldered MFF2 multi-carrier eSIM, a 0.2 mm tipping-bucket rain gauge, a solar radiation sensor and an SDI-12 / RS485 expansion port.
Figure 8.2 · Hardware architecture: common core, Sonda Lite and Sonda Pro
Subsystem Common core platform Sonda Lite Sonda Pro
Power 5 W solar panel → MPPT charger → LiFePO4 battery → regulated rails Included Included
Compute and security ESP32-S3-WROOM-1 MCU + ATECC608B (ECDSA P-256) Included Included
Storage and time 16 MB SPI NOR flash buffer + RV-3028 RTC with backup Included Included
Base sensors 2× capacitive soil moisture/temperature probes (10 and 30 cm) + SHT45 air T/RH Included Included
Enclosure and mount IP67 UV-stabilised polycarbonate, vent, M12 connectors, pole mount Included Included
Backhaul and positioning Variant-specific WiFi 2.4 GHz + BLE; location from field pin in app BG95-M3 LTE-M / NB-IoT + built-in GNSS; MFF2 soldered eSIM
Weather and bus expansion Variant-specific None (base soil and air sensing only) 0.2 mm rain gauge, solar radiation sensor, SDI-12 / RS485 port
Target deployment Variant-specific Fields with local WiFi coverage Remote fields with cellular IoT coverage
Swipe sideways to see the whole table.

4. Hardware modules and unit costs normalized in USD (BOM status)

Distributor quotes in other currencies are normalized to USD with fixed reference rates ([ASSUMPTION A-3]: 1 GBP = 1.27 USD; 1 EUR = 1.08 USD), keeping the original verified list price in parentheses. All prices exclude VAT, sales taxes, import duties and freight; landed cost varies by destination market and is not estimated at this stage. "Qty 100" is the 100-unit price break.

Hardware modules and unit costs (BOM status)
Block / subsystem Part (as designed) Lite Pro Unit (Qty 1) Batch (Qty 100) Evidence
MCU, WiFi + BLE Espressif ESP32-S3-WROOM-1-N8R8 Yes Yes $5.82 (£4.58) $4.39 (£3.46) [VERIFIED: DigiKey UK]
Cellular + GNSS Quectel BG95-M3 (BG95M3LA-64-SGNS) No Yes $29.60 (£23.31) $24.66 (£19.42) [VERIFIED: RS UK]
Secure element Microchip ATECC608B-SSHDA-B Yes Yes $0.86 (€0.796) $0.69 (€0.642) [VERIFIED: Farnell IE]
Air temperature + humidity Sensirion SHT45-AD1F-R2 Yes Yes $4.10 (€3.80) $3.70 (€3.43) [VERIFIED: SOS electronic]
Embedded SIM MFF2 soldered eSIM No Yes $0.70 $0.70 [VERIFIED: Telnyx pricing]
Rain gauge (0.2 mm/tip) Tipping-bucket collector; final SKU open No Yes $150.00 (benchmark) [UNKNOWN] [VERIFIED: NovaLynx store]
Soil moisture + temperature (×2) Capacitive, never resistive; SKU open (U-07) Yes Yes [UNKNOWN] [UNKNOWN] Research-grade benchmark: $218 each [VERIFIED: METER TEROS 11]
Offline buffer 16 MB SPI NOR flash (U-08) Yes Yes [UNKNOWN] [UNKNOWN] Not yet quoted
Real-time clock RV-3028 RTC + backup (U-08) Yes Yes [UNKNOWN] [UNKNOWN] Not yet quoted
Power subsystem 5 W panel, LiFePO4 cell, MPPT charger (U-08) Yes Yes [UNKNOWN] [UNKNOWN] Not yet quoted
Mechanics IP67 PC enclosure, M12 connectors, radiation shield (U-08) Yes Yes [UNKNOWN] [UNKNOWN] Not yet quoted
PCB + assembly Custom 2/4-layer board (U-22) Yes Yes $8.00 + $0.0016/joint $8.00 + $0.0016/joint [VERIFIED: JLCPCB]; joint count [UNKNOWN]
Priced core electronics subtotal MCU + secure element + SHT45 (Lite); + BG95-M3 + eSIM (Pro). Excludes probes, rain gauge, flash, RTC, power, enclosure and PCB Lite / Pro ~$10.78 / ~$41.08 ~$8.78 / ~$34.14 Computed from verified items [ASSUMPTION A-3]
Swipe sideways to see the whole table.

Optional SDI-12 / RS485 expansion sensors [DESIGN]: soil electrical conductivity (TEROS 12, $258.00) and leaf wetness for fungal-disease models (PHYTOS 31, $164.00), both at METER 2025 fellowship-sheet prices [VERIFIED: METER 2025 fellowship sheet]. The solar radiation sensor SKU for the Pro remains [UNKNOWN] (U-21).

BOM sensitivity: the priced core electronics are inexpensive (~$8.78 Lite and ~$34.14 Pro at Qty 100, of which the BG95-M3 module is ~$24.66). External transducers dominate: two research-grade soil probes alone (~$436 at the benchmark price) would exceed the whole priced core of either variant, and the rain gauge benchmark ($150.00) is more than four times the Pro core. Choosing between an industrial capacitive probe and a research-grade probe is therefore the main hardware trade-off before EVT (U-07).

Planning BOM used for the budget [ASSUMPTION A-4]: until EVT fixes the parts, the budget (section 17) prices each unit with benchmark parts on the open lines (soil probes, rain gauge, power, enclosure; listed under Sources): Pro $493.73 and Lite $316.96 at 100 units. Of the Pro figure, the electronics are $41.37 and the eSIM $0.70; transducers, power and mechanics are $451.66, which is why the probe choice (U-07) moves the price more than any chip. The electronics in this figure were priced at earlier exchange rates and sit about $1.84 above the table above, so the planning BOM is a ceiling for budgeting, not a quote.

5. Firmware security, backend ingestion and engineering estimates

Hardware-rooted identity and firmware safety [VERIFIED: docs/DEVICES.md]: each unit generates its private key inside the ATECC608B during factory enrolment, and the key never leaves the chip. Every batch is signed with ECDSA P-256; the backend rejects duplicate seq values and clock skew above 10 minutes. Firmware enforces secure boot, flash encryption and signed A/B OTA updates with automatic rollback.

Backend ingestion and UI integration (to build) [VERIFIED: docs/DEVICES.md]: dedicated endpoint POST /api/v1/devices/readings feeding the devices and device_readings tables; non-destructive range validation (out-of-range values are flagged, not dropped); rate limiting with Rack::Attack; an operator pause switch in the Integrations Hub; and a field-workspace device card with battery and signal health, threshold alerts and the "Your sensor next to the satellite" charts.

Cellular payload and connectivity cost [ASSUMPTION A-1]: 4 readings/hour × 24 h × 30 days = 2,880 readings/month (~720 hourly batches). A compact JSON/CBOR payload (~180 bytes/reading plus TLS overhead) is expected to stay under 1–2 MB/month if TLS sessions are resumed between uploads; the model uses 5 MB/month to cover full handshakes, OTA checks and retries. At 5 MB/month, Telnyx ($0.70 one-time eSIM, $0.00/month platform fee, $0.078/MB) costs ~$0.39/month versus Hologram ($3.00 SIM, $1.00/month, $0.03/MB) at ~$1.15/month [VERIFIED: Telnyx and Hologram pricing]. Excluding one-time SIM costs, Telnyx stays cheaper up to ~21 MB/month per device. Per-country tariffs, roaming restrictions and LTE-M/NB-IoT availability in each target market are [UNKNOWN] (U-09, U-10).

Energy budget envelope [ASSUMPTION A-2]: the MCU spends more than 98% of the time in deep sleep, waking ~1–2 s every 15 minutes for sensing and ~5–10 s per hour for the radio upload. With a 3.2 V / 6,000 mAh LiFePO4 cell (~19.2 Wh nominal) and a 5 W panel, an estimated consumption of ~0.3–0.6 Wh/day gives a conservative autonomy of more than 15–20 days without sun (after derating for cold, ageing and depth of discharge) and recovers a day's use in under one hour of peak sun. Because deployments will span all climate zones, the charger must block LiFePO4 charging below 0 °C and the power design must be checked against high-latitude winter irradiance and hot-climate enclosure temperatures [DESIGN]. All figures are to be measured on EVT units (U-08).

6. Manufacturing path [DESIGN]

Manufacturing gates
Gate Units Build Exit criteria
EVT · Engineering Validation Test 10 Commercial development boards (ESP32-S3 and BG95-M3) in a standard IP67 enclosure with candidate soil probes Store-and-forward firmware loop validated; power draw measured; 60-day pairing against SMAP and IMERG on pilot fields in at least two contrasting agro-climatic regions (O5)
DVT · Design Validation Test 20–30 Custom PCB through turnkey PCBA; final UV-stabilised enclosure and M12 connectors Full-season field deployment; soil-texture calibration curves indexed by Field.soil.texture covering the main soil classes of target markets (open design item in docs/DEVICES.md)
Certification and PVT · Production Validation Test 100 (pilot batch) Per-unit factory provisioning: flash firmware with secure boot → ATECC608B internal key generation → register public key in Aluni's devices table [VERIFIED: docs/DEVICES.md] → print and apply QR claim label → thermal burn-in and IP67 water-ingress test → pack Repeatable yield; units ready for market-specific certification and shipment
Swipe sideways to see the whole table.

Make-vs-buy gates: off-the-shelf versus custom enclosures and probes are re-evaluated at each gate [VERIFIED: docs/DEVICES.md].

7. Regulatory certification path (global, module-based)

Using pre-certified radio modules (ESP32-S3-WROOM-1 and Quectel BG95-M3) lets the Sonda reuse existing module approvals in many markets, so that the host product needs a lighter declaration and EMC assessment instead of full end-device radio certification [DESIGN]. Each destination market (or regional bloc) has its own type-approval regime, and which module approvals are recognized in each target market is [UNKNOWN] (U-09, U-10). The only market costed so far is the United States, used here as a reference benchmark:

Certification paths, United States as the reference benchmark
Certification path Applicability Benchmark cost (USD) Benchmark lead time Evidence
Host declaration reusing pre-certified module approvals Sonda Lite; Sonda Pro if the BG95-M3 approval covers the host design $2,500 – $5,500 3 – 6 weeks [VERIFIED (secondary): MarkReady] (US FCC SDoC); other markets [UNKNOWN] (U-10)
Cellular IoT EMC / radio testing Sonda Pro if the host board or antenna needs a full radio test $15,000 – $40,000 8 – 16 weeks [VERIFIED (secondary): MarkReady] (US); other markets [UNKNOWN] (U-10)
Full cellular end-device + industry certification + operator approval Sonda Pro only if mobile operators require end-device certification $50,000 – $200,000 6 – 9 months [VERIFIED (secondary): MarkReady] (US, PTCRB route); other markets [UNKNOWN] (U-10)
Swipe sideways to see the whole table.

Gated rollout strategy [DESIGN]: Sonda Lite ships first because it relies only on a WiFi/BLE module and stays on the cheapest, fastest certification route in each market. Sonda Pro follows market by market once module manufacturer and test-lab consultations confirm BG95-M3 approval coverage (U-09), national type-approval requirements (U-10) and LTE-M/NB-IoT network availability in each target market.

Back to contents

9. How Aluni decides: use cases, formulas and a worked example on real data

Every recommendation Aluni makes is a deterministic calculation on public data, and it can be reproduced line by line. The formulas below are copied from the production services. They were run on a real field: the team's captured dataset for Ames, Iowa (42.0, −93.75), with NASA POWER daily data for 1 Jan–31 Oct 2025 (304 days), plus SSURGO soil, USGS 3DEP slope and the USDA Cropland Data Layer [VERIFIED: farm_data_42.0_-93.75.csv / .json in the repo].

Use cases: problem → data → calculation → farmer decision

Use cases
Farm problem Data source (NASA first) Calculation in Aluni What the farmer decides
Will rain cover this crop's water need? NASA POWER (T2M, RH2M, WS2M, ALLSKY_SFC_SW_DWN, PRECTOTCORR); FAO-56 Kc Weather::Et0 (Penman-Monteith) → Crops::WaterNeed (Kc curve, CROPWAT effective rain) → water score Which crop fits rainfed; whether to irrigate
Will bare soil wash away? GPM IMERG 20-year daily rain; USGS 3DEP or NASADEM slope Weather::RainExtremes (25/50 mm days) → erosion score; Rotations::Timeline storm months Rotation with fewer row crops on slopes; cover crop timing
When is the soil bare and exposed? Landsat + Sentinel-2 NDVI; MODIS 20-year normal; POWER frost days Fields::Vegetation, Rotations::Timeline (bare month under storm, dry or frost) Cover crop and sowing window
Is irrigation water secure? GRACE/GRACE-FO water storage; SMAP L4 root-zone moisture; ECOSTRESS ESI Weather::WaterStorage (−15 water points when depleted), Weather::SoilMoisture, Weather::WaterStress Reduce irrigated area; choose drought-tolerant crops
Is the crop suited to this pin and soil? FAO GAEZ v4; FAO ECOCROP; SSURGO Crops::Potential, Crops::Fit (−20 points per soil property outside tolerance) Drop unsuitable crops
Does the rotation break pests and build nitrogen? USDA CDL crop history; crop families Crops::RotationCheck (same family, monoculture, no legume in 4 seasons) Insert a legume or another family
Frost, heat, storm this week Meteomatics 7-day; POWER Weather::Forecast, Weather::DailyAlertsJob Delay sowing, protect, harvest early
What will the plan cost and earn (vision) FAOSTAT, USDA AMS/NASS, WFP, World Bank (licence verdicts recorded) Module 9: cost per ha + price → gross margin Pick the strategy that pays
Swipe sideways to see the whole table.

Sources for the calculations column: [VERIFIED: app/services/weather/et0.rb, crops/water_need.rb, rotations/comparator.rb, rotations/timeline.rb, weather/rain_extremes.rb; docs/FIELDS.md; docs/DATA_SOURCES.md].

The formulas (as implemented)

1. Reference evapotranspiration, FAO-56 Penman-Monteith (Weather::Et0):

ET0 = [0.408 Δ Rn + γ (900 / (T + 273)) u2 (es − ea)] / [Δ + γ (1 + 0.34 u2)]

Here Rn = 0.77 Rs − Rnl; Rs comes from ALLSKY_SFC_SW_DWN, γ = 0.000665 P, and P is set by elevation.

2. Crop water need over the FAO-56 stages (Crops::WaterNeed). Kc is flat, rises, holds, then falls:

ETc = Σ (d = 1 to L) Kc(d) × ET0(d)

3. Effective rain per month, FAO/CROPWAT (Smith 1992):

Pe = 0.8 P − 25 when P > 75 mm; Pe = max(0.6 P − 10, 0) when P ≤ 75 mm

4. Water score (rainfed): 100 × the mean of Pe/ETc over the strategy's crops. Season totals let a wet month hide a dry one: sunflower is 91% covered by the season total but 80% month by month. Irrigated: 60 + 40 × that share − 15 when GRACE shows the regional store depleted.

5. Erosion score: 100 − (row crops × penalty). The penalty is 25 on slopes of 5% or more, else 8, plus 7 where 50 mm days come at least once a year (IMERG).

6. Soil health = min(100, 25 + 20 × legumes + 10 × families + 10 × non-row crops). Nitrogen = 100 × legume seasons / seasons. Pest break = 100 − 25 × same-family pairs − 30 × monoculture runs. Workload = 100 − 15 × (distinct crops − 1) − 20 for small grains without a harvester.

7. Total, weighted by the farmer's goals (weight 3 for a criterion that matches a goal, else 1):

S = Σc (wc × sc) / Σc wc, with wc ∈ {1, 3}

8. Exposed month (Rotations::Timeline): a month with no crop in the ground whose 20-year IMERG mean is at least one heavy-rain day a year (storm), rain under half of ET0 (dry), or ≥2 frost days. A cover crop protects every month of a bare window of 2+ months except the first.

Worked example: Ames, Iowa, 2025 season

Field inputs [VERIFIED: repo dataset]: elevation 321.7 m; slope 4.05% (USGS 3DEP); Clarion loam, well drained, 3.5% organic matter, available water capacity 0.21 cm/cm (SSURGO). CDL history: soybeans 2020, 2022, 2024; corn 2021, 2023, 2025. Profile and goals as in the team's Nebraska demo farm (rotation + soil health; owns a harvester) [VERIFIED: config/demo_farms.yml].

Method note. Production uses 25-year POWER normals and 20 years of IMERG. This example uses the 2025 daily series actually captured, so it shows one real season. Wind is not in the file, so each day's NASA POWER WS2M at the pin is used (mean 3.15 m/s in 2025; the FAO-56 default of 2 m/s understated it). Heavy-rain days come from POWER rain here, not IMERG. Every number comes from docs/proposal/ames_worked_example.rb (run mise exec -- ruby docs/proposal/ames_worked_example.rb). It writes ames_worked_example_out.json, lists every gap it does not fill, and exits with an error if any of its 24 checks fails. Since the backtest, the app reads rain from GPM IMERG and plants in the region's observed month (section 10); this example keeps the 2025 POWER series and the FAO-56 calendar so that its figures stay comparable with the cross-check below.

Monthly weather, 2025 (computed)
Month ET0 (mm) Rain (mm) Frost days Days ≥ 25 mm Risk flags (timeline rule)
Jan 18.1 7.0 31 0 dry, frost
Feb 25.4 10.0 26 0 dry, frost
Mar 67.2 66.7 22 1 storm, frost
Apr 86.6 89.8 7 1 storm, frost
May 127.0 75.8 0 1 storm
Jun 132.3 187.4 0 2 storm
Jul 138.7 236.0 0 3 storm
Aug 114.1 82.6 0 0 –
Sep 94.5 39.9 0 0 dry
Oct 76.8 59.1 0 0 –
Swipe sideways to see the whole table.

Jan–Oct totals: ET0 881 mm, rain 854 mm. Last spring frost: 8 Apr; no fall frost by 31 Oct. Growing degree days (base 10 °C, 86/50 method, May–Sep): 1,702; the landing page's 1,915 is the simple method (daily mean − 10 °C) over Jan–Oct. Wettest day: 46.8 mm (May), so there were no 50 mm days and no storm penalty.

Crop water balance, 2025 (Crops::WaterNeed algorithm)
Crop (FAO-56 season) ETc need (mm) Season rain (mm) Effective rain (mm) Deficit (mm) Covered by rain
Corn, 1 Apr–28 Aug 506 663 408 119 80%
Soybean, 1 May–12 Sep 499 588 365 138 73%
Spring wheat, 1 Mar–13 Jul 428 508 297 165 69%
Sunflower, 1 Apr–8 Aug 418 610 382 83 91%
Swipe sideways to see the whole table.
Strategy scores (Rotations::Comparator; crop fit left out because GAEZ is not in this dataset, exactly as the app does when a criterion is missing)
Strategy (next 3 seasons) Pest break ×3 Soil health ×3 Nitrogen ×3 Water ×1 Erosion ×3 Workload ×1 Total
Current: soybean → corn → soybean 100 85 67 76 76 85 82
Small grain: wheat → soybean → corn 75 75 33 74 84 70 68
Other family: sunflower → corn → sunflower 100 45 0 88 76 85 60
Swipe sideways to see the whole table.
Bare and exposed months, 2025 weather (Jan–Oct)
Strategy, seasons 2026–2028 (Jan–Oct known; Nov–Dec have no data) Vulnerable months, FAO-56 seasons With cover Vulnerable, corn on the farmer's Apr–Oct record With cover
Current: soybean → corn → soybean 12 4 11 3
Small grain: wheat → soybean → corn 11 4 10 3
Other family: sunflower → corn → sunflower 12 6 11 5
Swipe sideways to see the whole table.

Source: timeline.lines in docs/proposal/ames_worked_example_out.json. A cover protects every month of a bare window except the first. Unknown months break a window, so these counts are low: November and December were open soil in all 26 clear satellite views of 2013-2025 (section 10, V8).

What this shows. On this field Aluni does not push change for its own sake: the current corn–soybean rotation scores highest (82). That agrees with 670 field trials of maize after soybean in a climate like Ames (+9.9%, section 10, V4) and with what the farmer has grown since at least 2008 (CDL). The data flags 12 vulnerable bare months over the next three seasons on the FAO-56 calendar, under frost, drought or the 25 mm storms of March–July. A cover crop in each bare window cuts the count to 4, but the soil it saves is about 13% (RUSLE, section 10, V6): the riskiest months are May and June under young crops, which the month count treats as covered. This is one season of model output, checked against the app's code (below) and against 25 years of records (section 10).

The same formulas in Buin, Chile (demo farm, canal-irrigated). GRACE shows the region's water store about 20 cm below its 2004–2009 level [VERIFIED: docs/DEMO.md]. Because the field draws on canals, irrigation_risk takes 15 points off the water score of every strategy, and the plan flags "water storage" as a risk to check locally [VERIFIED: comparator.rb, SUPPLY_RISK = 15].

Cross-check: the app's own services give the same answer

docs/proposal/ames_cross_check.rb feeds the same 2025 data, as monthly means, into the production Weather::Et0.monthly, Crops::WaterNeed and Crops::RotationCheck. Run on 15 Nov 2026 with mise exec -- bin/rails runner docs/proposal/ames_cross_check.rb in the project's own environment. It now also totals the strategies with the app's own water shares:

Cross-check against the production services
Check App services Worked example Result
ET0, Jun–Aug 129.3 / 138.0 / 114.7 mm 132.3 / 138.7 / 114.1 mm App 0.5% higher to 2.3% lower
ET0, Jan–Feb 16.1 / 21.6 mm 18.1 / 25.4 mm App 11–15% lower: monthly means lose the daily Tmax/Tmin spread, as the script header predicts
Corn ETc / effective rain / deficit 497 / 408 / 111 mm 506 / 408 / 119 mm Same rain; ETc 1.8% lower
Soybean 490 / 371 / 124 mm 499 / 365 / 138 mm Within 3%
Spring wheat 408 / 306 / 150 mm 428 / 297 / 165 mm Within 5%
Sunflower 408 / 382 / 76 mm 418 / 382 / 83 mm Within 3%
Pest break (current / small grain / other family) 100 / 75 / 100 100 / 75 / 100 Identical
Strategy totals with the app's water shares 82 / 68 / 60 (water 78 / 78 / 90) 82 / 68 / 60 (water 76 / 74 / 88) Ranking and totals identical
Swipe sideways to see the whole table.

The worked example reruns cleanly: all 24 of its checks pass.

Robustness: does the answer change if one input changes?

Robustness of the ranking
Variant Current Small grain Other family Winner
Baseline (3DEP slope 4.05%) 82 68 60 Current
SSURGO mapped slope 8% (the steep-slope penalty applies) 71 60 49 Current
No harvester 82 66 60 Current
No goals (all weights 1) 82 69 66 Current
Swipe sideways to see the whole table.

Water-limited yield (FAO-56 soil buffer, FAO-33 yield response)

The soil buffer is RAW = p × 1000 × AWC × Zr, with AWC 0.21 from SSURGO and p and Zr from FAO-56 Table 22. The rain deficit beyond the buffer is stress, and FAO-33 converts it to yield:

1 − Ya / Ym = Ky × (1 − ETa / ETc)

Water-limited yield, 2025
Crop Deficit (mm) Soil buffer RAW (mm) Stress (mm) ETa/ETc Ky Yield factor
Corn 119 116 4 0.993 1.25 0.991
Soybean 138 63 75 0.85 0.85 0.872
Spring wheat 165 116 50 0.884 1.15 0.867
Sunflower 83 76 7 0.982 0.95 0.983
Swipe sideways to see the whole table.

The model gives soybean a 13% cut in 2025 because its shallow root zone (0.6 m, the low end of FAO-56 Table 22) holds half the buffer corn has. It did not happen: Story County soybeans yielded 4.2% above trend (NASS). With 1.3 m roots, or with gauge-based rain, the modelled cut falls to 0–2%. Over 2001–2025 the same model ranks drought insurance losses well but does not track yields in humid Iowa; in rainfed Nebraska it does (correlation 0.66 for corn, 0.77 for soybean; section 10, V2). So the buffer belongs in a drought flag, not a yield forecast, until each field shows which of the two it is. The app itself still ignores stored soil water (docs/FIELDS.md); this buffer exists only in the proposal script.

Pests the daily data can already flag

  • Tar spot (corn): 20 of the 150 corn-season days had a 30-day mean temperature of 18–23 °C, the most predictive range in the rule documented in pest_weather_rules.md. The first was 16 Jun and the longest run 16 days. Only 4 days reached a daily mean RH of 90% or more; the rule's humidity term is hourly, so that count is a proxy.
  • Fall armyworm: 1,402–1,890 degree-days from May to October, enough for 3.6–3.9 generations after arrival. Two published thresholds were used (du Plessis 2020; Malekera 2022). The arrival date needs FAMEWS or trap data.
  • Needs hourly data, which the Sonda or POWER hourly will supply: gray leaf spot, northern corn leaf blight and Asian soybean rust (current rotation). The small-grain rotation adds stripe, leaf and stem rust and Fusarium head blight.
  • Backtest (section 10, V5): over 2001–2025 every rule leans the right way against insured disease and insect losses in Central Iowa (Spearman 0.14 to 0.32) but none is significant, and around York they scatter around zero (−0.19 to 0.18). The rules say when to scout; they need leaf-wetness hours and scouting or trap records before they can say more.

Nitrogen and margin per strategy (illustrative inputs)

These two results rest on 20 inputs tagged placeholder in the script: yields, prices, costs and N rates. Real values replace them through docs/proposal/ames_inputs.json once the NASS, AMS and ERS/ISU keys exist (U-20). They show the calculation the Costs and Markets module will run, not a forecast.

Nitrogen balance and margin per strategy
Strategy N applied + fixed − removed, 3 seasons (kg N/ha) Margin, $/ha/yr Corn −20% Soybean +20%
Current 150 + 224 − 631 = −257 (partial balance) · −259 (soil mined) 614 478 789
Small grain 240 + 112 − 484 = −132 473 336 560
Other family 340 + 0 − 303 = +37 409 273 409
Swipe sideways to see the whole table.

The margin ranking agrees with the agronomic ranking in all four price scenarios in the script. What the balance column adds is a cost the score does not show: in this partial balance (fertiliser plus fixation minus grain removal) the current rotation removes the most nitrogen. The backtest tested the margin idea on 2008–2025 with real NASS yields and prices (section 10, V7): choosing the crop each spring by expected margin earned $255/ha less over 18 years than the Ames farmer's strict alternation, but $752/ha more over 15 years than the York farmer, who had grown corn three years running. That is the trade-off a farmer should talk over with an adviser.

What the Sonda measures that the satellites cannot (from the same run)

Wind sensitivity of the corn result
Wind at 2 m (m/s) Corn ETc (mm) Deficit (mm) Corn yield factor
1.0 481 98 1.000
2.0 (FAO-56 default) 497 112 1.000
3.0 510 124 0.979
4.0 521 134 0.956
Swipe sideways to see the whole table.

Wind moves the corn yield factor by up to 4.4% between 2 and 4 m/s, but NASA POWER publishes daily wind at 2 m for every pin, and the worked example now uses it, so an anemometer is not what the Sonda adds. Leaf wetness is: every disease rule counts wetness hours, and the humidity proxy showed no significant link to real losses (section 10, V5). Rain is the other: NASA POWER read 12% below the 59 CoCoRaHS gauges near Ames in every season of 2008–2025, and 5% below the 77 near York (section 10, V1). Each sensor in the design has a validation metric against its NASA counterpart:

  • Air temperature and humidity: bias and RMSE of daily Tmin/Tmax against POWER, and hours at RH 90% or more against POWER hourly.
  • Soil moisture: correlation of daily anomalies against SMAP L4.
  • Rain: bias of event totals and the hit rate for 25 mm days against IMERG.
  • Radiation: bias of daily MJ/m² against POWER ALLSKY.

Back to contents

10. Backtest 2001–2025 on two fields: would Aluni have been right?

Section 9 shows that the formulas run and match the app. This section asks the harder question: if Aluni's calculations had been run every season from 2001 to 2025, would they have matched what really happened? We asked it on two fields that differ where it matters: Ames, Story County, Iowa (42.0, −93.75), humid, where rain rarely limits the harvest; and York, York County, Nebraska (40.95, −97.5), drier, where USDA publishes rainfed (non-irrigated) county yields. Each study states its pass criterion before the results were seen, uses the same criterion on both fields, uses only data available before each decision, and scores Aluni against records it does not read: rain gauges, county yields, crop insurance losses by cause, field trials and satellite images. Nothing is fitted to the outcome; every coefficient comes from FAO, USDA or the cited papers. [VERIFIED: docs/proposal/validation/, sites.yml]

See the backtest evidence

Eight studies on two fields, each with its criterion set before the results
Study Question Truth (independent record) Criterion, set before the results Ames, Iowa (humid) York, Nebraska (rainfed)
V1 data Is the NASA data at the pin factual? CoCoRaHS gauges within 30 km (59 at Ames, 77 at York); Daymet 1 km Tmax and Tmin RMSE ≤ 2.5 °C; May–Sep POWER rain within ±15% of gauges in ≥ 80% of years FAIL (rain). Tmax 1.7 °C, Tmin 2.4 °C, frost days agree 95%. POWER rain below the gauges in all 18 seasons (−12%), within 15% in 67%. IMERG within 15% in 83% PASS. Tmax 1.9 °C, Tmin 2.5 °C, frost days agree 96%. POWER rain −5%, within 15% in 83%. IMERG within 15% in 83%
V2 yield Would the water model have flagged the bad years? NASS county yields (detrended; rainfed at York); RMA losses by cause r ≥ 0.5 with the yield anomaly and 2012 among the 3 lowest modelled years FAIL. r = 0.03 (corn), −0.05 (soybean). Ranks drought and heat insurance losses well (ρ 0.61 and 0.80); 2012 lowest PASS. r = 0.66 (corn, 17 years), 0.77 (soybean, 13 years; the rainfed series stop in 2017 and 2013); 2012 lowest. Calls all 25 seasons stressed, so it overstates the size of losses
V3 crop choice Does the crop evidence point at the crops that thrive? CDL 2008–2025; NASS yields Corn and soybean ranked above wheat and sunflower; GAEZ yield within ±30% of NASS FAIL. GAEZ rates all 8 crops at the top (cannot rank); GAEZ maize +20%, soybean +99% vs NASS FAIL. GAEZ maize +21%, soybean +14% vs NASS, but the water criterion ranks wheat and sunflower first
V4 rotation Is the rotation advice backed by evidence? Zhao et al. 2022: maize after soybean in a climate like the site's, at farm N rates 95% interval of the soybean pre-crop effect excludes zero PASS. +9.9% maize yield [7.2 to 13.2%], 670 comparisons; +46% without N fertiliser PASS. +12.7% [5.6 to 22.0%], 194 comparisons from 17 sites; +57% without N fertiliser
V5 pests Do the pest weather rules match real losses? RMA disease, mycotoxin and insect losses in the surrounding counties Reported, not scored REPORTED. All five rules lean the right way (ρ 0.14–0.32), none significant; first favourable day about 3 months before the main loss month REPORTED. ρ −0.19 to 0.18, none significant; lead about 1 month
V6 erosion How much soil does each rotation lose, and how much does a cover save? RUSLE (AH-703) on Daymet rain, SSURGO K and slope length The app's erosion score orders the strategies as RUSLE does PASS. Order holds. Rye cover saves 13–14%. Losses 17–24 t/ha/yr (assumed residue and cover) PASS. Order holds. Rye cover saves 12–13%. Losses 20–27 t/ha/yr (same assumptions)
V7 money and operations Would choosing by expected margin have paid? Can the farm plant on time? NASS yields and prices, days suitable for fieldwork, planting progress; placeholder costs Margin rule's realised margin ≥ the farmer's FAIL. 2008–2025: margin rule $11,995/ha vs farmer (strict alternation) $12,250 vs continuous corn $10,366 vs hindsight $13,567. Corn 50% planted 25 Apr–19 May PASS. 15 seasons, 2008–2024: margin rule $13,335/ha vs farmer (corn three years running) $12,583 vs continuous corn $10,388 vs hindsight $13,886. Corn 50% planted 2–19 May
V8 timeline Do the timeline's crop and bare months match the field? Landsat + Sentinel-2 monthly NDVI 2013–2025 with the CDL crop ≥ 80% monthly agreement and both season edges within one month in ≥ 80% of years FAIL. 76% agree; canopy appears 2.2 months after the FAO-56 planting month; Nov–Dec open soil in 26 of 26 views. 90% with the app's current rule FAIL. 65% agree; same 2.2-month lag; Nov–Dec open soil in 19 of 21 views. 82% with the app's current rule
Swipe sideways to see the whole table.

What the second field changed. Five studies pass at York against two at Ames, with the same pass marks. The difference is informative, not lucky: the water model tracks yields where water is what limits the harvest (rainfed Nebraska) and only flags disasters where it is not (humid Iowa), so the app should tell each farmer which of the two their field is. The margin rule helps where a rotation has slipped (York) and not where the farmer already alternates strictly (Ames). The rotation evidence, the erosion ranking and the weak pest rules repeat on both fields, and the timeline fix holds on both. One planter of 12 rows covers 400 ha in all 25 springs at both sites, so planting capacity is not the bottleneck there.

Answers to the questions this proposal must answer

  • Could Aluni have predicted the facts on these fields? Partly, and we now know where. Temperature and frost from NASA are reliable at both pins (V1). Rain from GPM IMERG passes on both; NASA POWER rain read 12% low in Iowa. The water model tracks rainfed yields in Nebraska but only spots drought disasters in Iowa (V2).
  • Can it suggest a better rotation or a replacement crop? A better rotation, yes, and the evidence is strong on both fields: corn after soybean yields +9.9% and +12.7% at farm nitrogen rates (V4), and RUSLE confirms the erosion ranking (V6). A replacement crop, not yet: neither GAEZ suitability nor water alone ranks the crops farmers actually grow (V3). That needs the margin criterion of Gate 3 with each farmer's own yields and prices.
  • Can it help prevent pests and disease? It can say when weather favours a disease, as a reminder to scout; it cannot say a crop will be hit. No weather rule tracked real losses significantly on either field (V5). Leaf-wetness hours and scouting or trap records are needed, which is the strongest case for the Sonda (section 8).
  • Can it inform costs, supplies, people, equipment and land productivity? The calculations work on real data: field capacity against NASS days suitable for fieldwork, planting windows from NASS progress, nitrogen balances and margins per strategy. Rotating clearly pays against continuous corn (the Ames farmer's alternation earned $1,884/ha more over 18 years); choosing by margin paid at York (+$752/ha over 15 years) but not at Ames (−$255/ha). These margins use placeholder costs; the app uses each farmer's own recorded costs instead (U-26).

What changed in Aluni because of this

Fixes raised by the backtest
ID Cause found Change Status
U-27 NASA POWER rain read 12% below local gauges in Iowa; IMERG passes on both fields (V1) The water need reads GPM IMERG's 20-year monthly rain where the field has it, else the POWER normal [VERIFIED: app/services/crops/water_need.rb] Done
U-28 The FAO-56 table plants corn on 1 April; Iowa plants late April to mid May (V7, V8) Planting month from the farmer's record, else the region's USDA NASS crop-progress calendar, else FAO-56 [VERIFIED: app/services/crops/planting_calendar.rb, config/crops/planting_calendars.yml] Done
U-29 The timeline counted the field as covered from planting; NDVI shows open soil until the canopy closes (V8) A crop is establishing until FAO-56 effective full cover and its months count as exposed; agreement with 13 years of satellite NDVI rose from 76% to 90% of months at Ames and from 65% to 82% at York [VERIFIED: app/services/rotations/timeline.rb, validation_out.json] Done
U-30 The water model ranks drought years but tracks yields only on rainfed land (V2) Shown as a drought-risk guide, never as a harvest figure [VERIFIED: docs/FIELDS.md] Done
U-31 RUSLE losses rest on assumed residue and cover values (V6) Report erosion as soil loss in tonnes with documented NRCS residue values and a published Iowa reference Open (G1)
U-32 No weather rule tracked insured disease losses significantly (V5) Validate the rules with leaf-wetness hours and scouting records in the pilot Open (G1, HW EVT)
Swipe sideways to see the whole table.

Limits of this backtest. Two fields in two US counties; yields and insurance losses are county aggregates; York's rainfed yield series are short (17 years for corn, 13 for soybean); insured losses are a thin proxy for disease; costs are placeholders; the RUSLE losses rest on assumed residue and cover values. Buin, Chile, the other pilot site, cannot be scored the same way: no documented open source has Chilean rain gauges, county yields or crop insurance losses. How to rerun everything is in section 11.

Back to contents

11. Reference tests: rerun every number

No figure in this proposal has to be taken on trust. Every figure comes from public data and a script in the repository, and the scripts fail loudly when the text and the code disagree. Results of the run on 15 Nov 2026:

Reference tests and their results
Test Command What it checks Result, 15 Nov 2026
Worked example mise exec -- ruby docs/proposal/ames_worked_example.rb Recomputes the Ames 2025 season (section 9) from the captured NASA POWER, SSURGO, 3DEP and CDL data; lists every gap it does not fill 24 of 24 checks pass
Cross-check mise exec -- bin/rails runner docs/proposal/ames_cross_check.rb Feeds the same data to the production services (Weather::Et0, Crops::WaterNeed, Crops::RotationCheck, the comparator's totals) Strategy totals identical (82 / 68 / 60); ET0 within 0.5–2.3% in the growing season; water need within 5%
Backtest data mise exec -- bin/rails runner docs/proposal/validation/fetch.rb Downloads the public records: NASA POWER, IMERG, Daymet, CoCoRaHS, USDA RMA and NASS, CDL, SSURGO, GAEZ, Landsat and Sentinel-2 NDVI Both fields complete (run once per field with SITE=york for Nebraska)
Backtest mise exec -- ruby docs/proposal/validation/run_all.rb (then SITE=york) Runs studies V1 to V8 against criteria written before the results (section 10) Ames: 2 pass, 1 reported, 5 fail. York: 5 pass, 1 reported, 2 fail. Each failure has a named cause and fix
Claims audit mise exec -- ruby docs/proposal/claims_audit.rb Recomputes 52 figures and evidence claims in this proposal (budget, BOM, worked example, both backtests) from the outputs above (docs/proposal/claims.yml) 52 of 52 reproduce; 0 differ from the code
App test suite mise exec -- bundle exec rspec Request specs render every page, including this one, in English and Spanish; locale parity; the services behind the scores Runs in CI (GitHub Actions) on every pull request and push to main
Swipe sideways to see the whole table.

What the audit caught. Twelve numbers changed between earlier drafts and this version because the audit found them stale (for example the measured NASA POWER wind replaced the 2 m/s default, which moved corn's water need from 497 to 506 mm), and two earlier claims were withdrawn: that soybeans lost yield at Ames in 2025 (NASS shows 4.2% above trend) and that the Sonda needs an anemometer (POWER publishes wind at every pin).

Data and licences. All inputs are public, for both fields: NASA POWER, GPM IMERG, Daymet, CoCoRaHS, USDA NASS QuickStats, RMA cause of loss, CDL, SSURGO, FAO GAEZ v4, Landsat and Sentinel-2 via Microsoft Planetary Computer, and Zhao et al. 2022 (CC BY 4.0). The NASS API needs a free key; everything else downloads without an account.

Back to contents

12. Evidence and validation

The software rests on working code, documented sources and a published backtest. The value to farmers has not been tested with any farmer yet, so Gate 1 is about evidence, not features.

What is evidenced today

  • A live system at aluni.us, deployed automatically from main, with a two-farm demo script [VERIFIED: ROADMAP.md, DEMO.md].
  • 13 data sources fetched per field pin, each value labelled with its dataset [VERIFIED: FIELDS.md].
  • About 60 data-source files with captured requests and licence verdicts [VERIFIED: docs/data_sources/, DATA_SOURCES.md].
  • Automated checks: RSpec, RuboCop, Brakeman, Rust checks [VERIFIED: README.md]. The decision engine is reproduced on a real field (Ames, 2025) by a script with 24 passing checks, and its results match the production services under bin/rails runner (section 9).
  • A 25-year backtest on two fields (section 10) scores eight claims against independent records with the same pass marks: 2 pass at Ames and 5 at York, 1 is reported on each, and every failure has a stated cause; four fixes are already in the app. docs/proposal/claims_audit.rb recomputes 52 figures in this document (section 11).
  • Known method limits, already stated by the team: NASA POWER is about 50 km and GAEZ about 9 km; water need ignores stored soil water [VERIFIED: FIELDS.md, team overview §11].

TRL register (NASA scale; self-assessed, not independently reviewed)

TRL register
Component TRL now Evidence Target TRL De-risking step Gate
Rotation comparator, timeline, evidence 5–6 Live with real NASA data at two demo pins; reproduced and cross-checked on a third real field (section 9); backtested 2001–2025 on that field and a second one in Nebraska (section 10: 2 and 5 of 8 pass, 4 fixes shipped); no real users yet 7 Season-long pilot with 10+ farmers G1
Gemini "why" with guardrails 6 In production, capped, falls back to rule text 7 Pilot review of narration accuracy G1
iOS and Android shells 4–5 Built, not in the stores [VERIFIED: team overview §10] 7 Store release, TalkBack/VoiceOver tests G1
Season Execution Plan; People, Equipment, Supplies 1–2 Concept in this document 6 Clickable prototype, then a pilot build G2
Costs, Budget, Markets 1–2 Sources documented; nothing built 6 One-country price feed + cost per ha G3
Sonda Lite / Pro 2 Design doc and a 3D catalogue page 6 EVT then DVT field season HW gates
Swipe sideways to see the whole table.

Open-items register

Nothing in this proposal rests on these items until they are resolved; each has the action that closes it and the gate it must close by. Items U-12 (hosting after credits, now $24.46 a month) and U-14 (labour rates, now BLS medians) are resolved. U-27 to U-32, the backtest fixes, are tracked in section 10.

Open items
ID Open item Depends on it Resolution action Due
U-03 No LICENSE file in the repository root, so the open-source claim is not yet legally true sections 14 and 21 Choose and add an OSI licence Before G1
U-04 Named people, roles and hours for each role in section 16 sections 16 and 22 Assign before the pilot starts Before G1
U-05 Farmers' use of and willingness to pay for the integrated tool sections 13, 14 and 18 Gate 1 pilot G1
U-06 What BAU costs farmers (time, adviser fees, losses) sections 3 and 18 Pilot interviews G1
U-07 Soil probe for EVT: industrial capacitive or research grade Section 8, A-4 Compare candidate probes on EVT units HW EVT
U-08 Flash, RTC, power and enclosure parts and the measured energy budget (A-2) Section 8 Quote the parts; measure draw on EVT HW EVT
U-09 Does the BG95-M3 module approval cover the Pro host? Is PTCRB or carrier approval needed? sections 8 and 21 Ask a test lab and Quectel HW DVT
U-10 Radio rules in Chile (SUBTEL) and the EU Section 21 Ask a test lab HW DVT
U-11 Real data per device per month (A-1 uses 5 MB) Section 17 Measure on EVT HW EVT
U-13 Gemini input tokens per narration (output is capped and priced) Section 17 Log token counts for one week G1
U-15 Pilot travel and farmer incentives Section 17 Site plans for York and Buin G1
U-16 Machine-data import standard for the Equipment module Section 7 Survey pilot farms' machines G2
U-17 farmOS integration effort (Option A) Section 6 2-day spike G1
U-18 Commercial platform API terms (Option D) Section 6 Read 2 platforms' terms G2
U-19 Farm-data privacy duties in the US and Chile Section 21 Legal review G2
U-20 NASS, AMS and api.data.gov keys for the app; no FAOSTAT token route [VERIFIED: DATA_SOURCES.md] Module 9 Apply for keys; bulk files meanwhile G3
U-21 Solar radiation sensor part for the Pro Section 8 Pick a part HW EVT
U-22 Small Sonda parts and PCB solder-joint count sections 8 and 17 First layout and full quote HW DVT
U-23 Gate dates and the accountable owner of each gate sections 15 and 22 Set at the start of WP1 Before G1
U-24 Exposed-month baseline on each pilot field (the Ames example in section 9 is the method) sections 4 and 23 Run the timeline at enrolment G1
U-25 Leaf-wetness part for the Pro (every disease rule counts wetness hours). Wind comes from NASA POWER WS2M, so an anemometer is optional sections 8 and 9 Pick parts for the SDI-12 port and price them HW EVT
U-26 20 placeholder inputs behind the nitrogen and margin results (yields, prices, costs, N rates) sections 9 and 18 Fill docs/proposal/ames_inputs.json from NASS, AMS and ISU/ERS budgets G1
Swipe sideways to see the whole table.

Assumptions log

Planning assumptions
ID Value Used in What would change the decision
A-1 5 MB of cellular data per Pro device per month sections 8 and 17 connectivity Above about 21 MB a month Hologram becomes cheaper than Telnyx; measured on EVT (U-11)
A-2 0.3–0.6 Wh a day; more than 15–20 days without sun Section 8 energy budget A measured draw that drains the battery in under 10 sunless days would need a larger cell or panel (U-08)
A-3 1 GBP = 1.27 USD; 1 EUR = 1.08 USD Section 8 prices A 10% move changes the priced Pro core electronics by about $3–4 per unit
A-4 Planning BOM: Pro $493.73, Lite $316.96 at 100 units sections 6, 17 and 18 The EVT probe and gauge choice (U-07); the hardware track stops if the DVT BOM exceeds the price ceiling (K3)
Swipe sideways to see the whole table.

Back to contents

13. Market, users and competitors

The gap Aluni can claim is NASA evidence per field, with its source shown, joined to season execution, free and in two languages. No competitor reviewed does all of that. Whether farmers value it is not yet proven (U-05).

Segments (already coded in Aluni's segment classifier [VERIFIED: docs/ONBOARDING.md])

Segments
Segment Rule in the app Likely need from the vision modules [DESIGN] Demand evidence
Smallholder subsistence Under 2 ha, solo or family, low budget, own use or local market Plan, alerts, Spanish, phone first; little equipment or inventory [UNKNOWN]
Small commercial Under 20 ha, sells beyond own use Tasks, supplies, cost per strategy [UNKNOWN]
Mid mechanized 20–200 ha, or a tractor plus planter/harvester with a small team or crew Equipment registry, People and Tasks, budgets, Sonda Pro [UNKNOWN]
Large commercial Over 200 ha, or a crew selling wholesale, contract or export Probably already has a platform; Aluni as an evidence layer [UNKNOWN]
Swipe sideways to see the whole table.

Size reference: more than 475 million farms under 2 ha, and only 1% of farms over 50 ha, which hold about two-thirds of the land [VERIFIED: FAO, Lowder et al., via People Food and Nature]. Serviceable and obtainable shares: [UNKNOWN]. The format asks for bottom-up demand, so this waits for the pilot.

Competitors and substitutes

Competitors and substitutes
Product What it covers Price Difference from Aluni Source
Climate FieldView (Bayer) Field data, yield analysis, imagery, prescriptions; row crops Basic free; Plus from $649/yr; FieldView Drive 2.0 $499 Paid analytics; not rotation-first; English-first [VERIFIED (secondary): softwares.com, yespress.io]; tiers $249/$649/$1,399 on a climate.com pricing page [VERIFIED]
Farmbrite All-in-one records: crops, livestock, accounting, tasks Crop plans from $29/mo; Essentials $59/mo, up to 10 team members Strong on operations; no NASA rotation evidence found [VERIFIED (secondary): Capterra, FitGap]
farmOS Open-source records: assets incl. equipment, maintenance logs, inventory, sensors Free (GPL); hosted from $50/yr Records, not decision support; possible partner (Option A) [VERIFIED: farmos.org]; hosting price [VERIFIED (secondary): SourceForge]
Extension offices, agronomists Local advice [UNKNOWN] Aluni sends farmers in with concrete questions, not instead Team overview §5
Swipe sideways to see the whole table.

Unfair advantage. Following the Lean Canvas advice, this stays blank until the pilot shows one. Candidates to test: per-field NASA evidence with provenance, open source, and Spanish parity.

Stakeholders [DESIGN]: farmers (users, decide); farm workers and managers (users of Tasks); extension services and cooperatives (channel, possible payers); NASA and data providers (sources; attribution); hardware suppliers and test labs; NASA Space Apps (the challenge that frames this work).

Back to contents

14. Operating model

Today Aluni runs at about $0 on cloud credits, with operator switches that pause AI, jobs, mail and push [VERIFIED: CLAUDE.md, DEPLOY.md]. After the credits, the same VM costs $24.46 a month, and Gemini is capped at $153.60 a month (section 17). Who pays later is open.

  • Service: one GCE VM running web, worker, Postgres and Redis through Kamal; deploys about 3 minutes from a push to main [VERIFIED: team overview §13].
  • Scaling: add Puma and Sidekiq containers; observations are cached per field, so traffic does not multiply outside API calls [VERIFIED: team overview §12].
  • Support and operations staff: [UNKNOWN] (U-04).

Revenue options [DESIGN], none validated (U-05)

Revenue options
Option Who pays Fits segment Risk
Free core, open source Grants and credits All Grant dependence
Sonda hardware plus a connectivity plan Farmer Mid mechanized, small commercial Hardware margin and support
Institutional licence: adviser dashboards, cohort reports Extension services, cooperatives, NGOs Smallholders through intermediaries Long sales cycles
Premium operations modules (budgets, multi-user) Farmer Small commercial, mid mechanized Competes with Farmbrite pricing
Swipe sideways to see the whole table.

Unit economics (customer acquisition cost, payback, churn): [UNKNOWN]. They come from the pilot, not from this document.

Back to contents

15. Implementation plan and gates

Four gated phases run in sequence, with a parallel hardware track. Each phase starts only when the previous gate passes. All dates are [DESIGN] targets with no reference-class check yet, so we treat them as optimistic and add buffer once named people and hours are confirmed (U-04, U-23).

Roadmap · 4 software phases, 3 hardware phases, 5 gates

The software phases run in sequence behind G1–G3. The hardware track starts after the hackathon and has its own two gates.

Work packages
WP Phase Deliverables Depends on
WP0 Hackathon close (Nov 2026) Live demo, this proposal and the two-field backtest (done); licence (U-03) –
WP1 Validate (target Dec 2026–Feb 2027) Shells in the stores; 10+ pilot farmers in York and Buin; usability sessions before and after the UX plan (team overview §17); baselines U-05, U-06, U-24; farmOS spike (U-17); backtest fixes U-31 and U-32; real prices and yields for U-26 WP0
WP2 Operations core (target Mar–Jun 2027) Season Execution Plan; People and Tasks (multi-user farms); Equipment registry and maintenance; Supplies and stock; observed vs planned bare months G1 pass
WP3 Economics (target Jul–Sep 2027) Cost items per operation; price feeds for 2 countries; gross margin criterion; monthly cash view G2 pass
HW1 EVT (target Q1 2027) 10 units from dev boards; payload defined (U-11); Seeed and DFRobot probes compared; leaf-wetness part added (U-25); small parts quoted (U-22) WP0
HW2 DVT (target Q2–Q3 2027) Custom PCB, 20–30 units, one field season, calibration per soil texture; ingestion API in Rails HW gate 1
HW3 Certify + PVT (target Q4 2027 or later) Lite certification first; 100-unit pilot batch HW gate 2
Swipe sideways to see the whole table.

Gate criteria [DESIGN]

Gate criteria
Gate Pass when Owner
G1 Validate O1 met (under 5 min, 4 of 5 unaided); at least 6 of 10 pilot farmers say they would use a season plan with tasks; U-05/U-06/U-24 baselines recorded Product lead
G2 Operations At least 3 of 10 pilot farmers log operations weekly for 6 weeks (O3); one institutional partner shows interest in writing Product lead
G3 Economics Cost per ha shown for every strategy in 2 countries (O4); price-feed licences confirmed Product lead
HW gate 1 EVT readings paired with SMAP and IMERG for 60 days on 2 pilot fields; BOM inside the price ceiling Hardware lead
HW gate 2 DVT survives one season (IP67, battery); the certification path is confirmed and costed (U-09, U-10) Hardware lead
Swipe sideways to see the whole table.

Change control: the scope list in section 5 is the baseline. Any new module needs a gate decision, because scope creep is the top risk (section 20).

Back to contents

16. Resources and team

The AstroMichis team shipped a working full-stack product in the hackathon window: 110 commits, Rails, Rust, two native shells, nine NASA datasets and a reproducible 25-year backtest [VERIFIED: git log, ROADMAP.md]. The roles below are what the plan needs; the people and hours behind each are confirmed before G1 (U-04).

Roles needed
Role needed Phase Filled by Commitment Gap / mitigation [DESIGN]
Product lead / sponsor All AstroMichis team (U-04) Owns every software gate One accountable owner per gate (section 22)
Rails + Hotwire engineer(s) WP1–WP3 AstroMichis team (built the live app) 0.5 FTE in Gate 1 (section 17) –
Agronomist / extension partner WP1–WP3 To recruit 0.25 FTE in Gate 1 (section 17) Recruit via an extension office at each pilot site
Embedded firmware engineer (ESP-IDF, cellular) HW1–HW3 None known – Critical gap for the hardware track
Hardware / mechanical (PCB, enclosure, IP67) HW2–HW3 None known – Contract design house or university lab
Field pilot coordinator (US, Chile) WP1 Agronomist line in Gate 1 (section 17) – One per site
Compliance and legal (radio, privacy, licences) HW3, G2 None – Test lab advice; pro-bono legal clinic
Swipe sideways to see the whole table.

Back to contents

17. Cost model

Gate 1 costs $25,538 (base, AACE Class 5: −50% / +100%, so $12,769–$51,075). The table below is the whole calculation: a quantity times a sourced rate on every line. Quantities are [DESIGN]; every rate is sourced.

Estimate class: AACE Class 5 (0–2% definition; −20/−50% to +30/+100%). Optimism bias: the HM Treasury upper bound applies once each base is known: 200% for equipment/development (the hardware track and new modules) and 41% for outsourcing (opex).

Gate 1 budget (12 weeks)
Gate 1 line (12 weeks) Quantity [DESIGN] Rate Rate source USD
Software development (WP1: shells, pilot fixes, UX plan) 240 h (0.5 FTE × 12 wk × 40 h) $63.20/h BLS 2024 median pay, software developers [VERIFIED: BLS OOH] 15,168.00
Agronomist / pilot lead (York, Buin) 120 h (0.25 FTE) $37.87/h BLS 2024 median pay, agricultural and food scientists [VERIFIED: BLS OOH] 4,544.40
Sonda Pro EVT units 10 $522.65 each: Pro planning BOM $493.73 (A-4) with the BG95 module ($26.07) swapped for a BG95-M3 Zero dev board ($54.99) Section 8; [VERIFIED: Waveshare] 5,226.50
Hosting, e2-medium VM, us-central1 3 months $24.46/month (disk and egress extra) [VERIFIED (secondary): Economize, VPSBenchmarks] 73.38
Gemini 2.5 Flash narration, cost ceiling 90 days × 2,000 requests/day cap × 1,024 max output tokens $2.50 per 1M output tokens Cap and token limit [VERIFIED: ROADMAP.md, ONBOARDING.md]; price [VERIFIED (secondary): Requesty, Maxim] 460.80
Connectivity, 10 Pro units 10 SIMs + 3 months $3/SIM, $1/month, $0.03/MB (A-1: 5 MB/month) [VERIFIED: hologram.io] 64.50
Total, Gate 1 25,537.58
Swipe sideways to see the whole table.

The Gemini line is a ceiling: it assumes every daily request hits the cap at full output length. Input tokens are extra at $0.30 per 1M. Not included: pilot travel and farmer incentives (U-15), and the small Sonda parts listed in section 8.

Whole-life running cost after Gate 1 (per month): VM $24.46 + Gemini at most $153.60 (30 days at the output ceiling) + $1.15 per Pro device in the field.

Back to contents

18. Financial and social viability

Costs are now sourced (section 17); revenue is not, because willingness to pay comes from the pilot (U-05). This section fixes the tests the numbers must pass at Gate 2. Per-field economics are already shown on illustrative inputs in section 9: $614, $473 and $409 per ha a year for the three Ames strategies.

  • Social value (non-monetized, to be measured): fewer exposed bare months (O2); water and nitrogen use per strategy; earlier frost, heat and rain warnings; NASA evidence reaching Spanish-speaking smallholders [VERIFIED as team-stated aims: team overview §5].
  • Hardware unit test [DESIGN]: price per Sonda must be at least full BOM + assembly + certification amortised over the batch + 12 months of connectivity. That sum is the switching value. With the planning BOM (A-4) it is at least $507.53 per Pro ($493.73 BOM + 12 × $1.15 connectivity), before certification (U-09) and small parts (U-22).
  • Software test [DESIGN]: yearly hosting + AI + support per active farm must be covered by grants or the institutional licence (section 14). The fixed platform cost is $24.46 a month for the VM, plus at most $153.60 a month for Gemini at its cap, shared across every farm on the platform.
  • Scenarios (downside / base / upside): to be built at Gate 2 from pilot data. No figures here, to avoid an upside-only model.

Back to contents

19. Funding structure

The only resources confirmed today are in-kind: Google Cloud credits and Gemini on Vertex, Cloudflare, Meteomatics' hackathon offer and free public data [VERIFIED: team overview §14, DEPLOY.md]. Everything else is a candidate.

Funding sources
Source Amount Status Linked gate
Google Cloud credits [UNKNOWN] remaining Committed (in use) WP0–WP1
Meteomatics hackathon offer [UNKNOWN] term Committed (in use) WP0–WP1
Space Apps recognition or follow-on programmes [UNKNOWN] Indicative G1
Agricultural research or innovation grants (US, Chile) [UNKNOWN] Not researched G1–G2
Extension services or cooperatives (licence or pilot partner) [UNKNOWN] Not contacted G2
Hardware pre-orders from pilot farms [UNKNOWN] Not offered HW gate 2
Swipe sideways to see the whole table.

Back to contents

20. Risks, assumptions, dependencies

The biggest risk is not technical. The vision could grow into a generic farm-management tool and lose what makes Aluni different, while a small team spreads itself across software and hardware. Probability and impact are [DESIGN] team estimates (H/M/L), not measured.

Risk register
ID Risk (cause → event → effect) P I Response Early warning Source
R1 Many modules → scope creep → no module reaches TRL 7 H H Gates; operations modules only after G1; accounting out of scope WP2 slips more than one month [DESIGN]
R2 Farmers don't log operations → the plan stays advice M H Pre-filled tasks from the plan; one tap to confirm; kill criterion K1 Under 3 of 10 weekly loggers [DESIGN]
R3 Coarse data (POWER about 50 km, GAEZ about 9 km) misleads a small field M M Each value shows dataset and resolution; comparison is relative; Sonda adds ground truth Pilot farmer disputes [VERIFIED: team overview §11]
R4 Hardware BOM and certification cost exceed what farmers pay M H Lite first; host SDoC path; price ceiling at HW gate 1 Small-parts quote (U-22) [DESIGN]
R5 No firmware or hardware skills on the team H H Recruit or partner before HW2 No named owner by EVT Section 16
R6 AppEEARS runs one task at a time per account → new fields wait hours for SMAP, MODIS, ECOSTRESS H L Queue and show waiting states; more accounts [UNKNOWN if allowed] Pilot onboarding delays [VERIFIED: DEMO.md]
R7 Outside services fail or rate-limit M M Retries with backoff; background jobs; pause switches Error rates [VERIFIED: ROADMAP.md, team overview §11]
R8 AI narration states something wrong L M Gemini only narrates computed facts; rule text when off Pilot review [VERIFIED: FIELDS.md]
R9 Credits end → hosting and AI costs appear M M Priced: VM $24.46/month, Gemini at most $153.60/month (section 17); budget alert and pause switches GCP billing alert [VERIFIED: DEPLOY.md]
R10 A data licence bars a commercial tier L M Licence verdicts already recorded; excluded sources listed New source added [VERIFIED: DATA_SOURCES.md]
Swipe sideways to see the whole table.

Dependencies: NASA POWER, AppEEARS and Earthdata; Planetary Computer; FAO GAEZ; USDA and USGS services; Meteomatics; Vertex AI; GCP; Cloudflare; app stores for the shells; a cellular carrier and test lab for the Pro [VERIFIED: FIELDS.md, CLAUDE.md, DEVICES.md].

Back to contents

21. Legal, regulatory and compliance

Two items block a public launch: Aluni has no visible licence, and the hardware has no certification path. Neither is hard to fix, but neither is fixed yet.

Compliance requirements
Requirement Authority Status Lead time Risk if missed
Open-source licence for the repo Team No LICENSE file in the repo root (U-03) Hours "Open source" claim not legally true
Data licences and attribution Each provider Verdicts recorded per source; non-commercial sources excluded [VERIFIED: DATA_SOURCES.md] Ongoing Takedown or redistribution breach
Radio certification, US FCC Part 15 Not started 3–16 weeks; up to 6–9 months with PTCRB [VERIFIED (secondary): MarkReady] Cannot sell the Sonda
Radio certification, Chile and EU SUBTEL; CE/RED [UNKNOWN] (U-10) [UNKNOWN] Cannot pilot the Pro in Buin
Farm and personal data privacy US, Chile law Passkeys, Active Record encryption, export to CSV/JSON in place [VERIFIED: CLAUDE.md, team overview §11]; legal duties [UNKNOWN] (U-19) [UNKNOWN] Trust and compliance
App store publishing Apple, Google Shells built, not submitted [VERIFIED: team overview §10] Review times [UNKNOWN] No push alerts on phones
Swipe sideways to see the whole table.

Back to contents

22. Governance and decision rights

One person owns each gate, named at the start of WP1 (U-23). Proposed RACI [DESIGN]:

RACI
Decision Product lead Engineering Hardware lead Agronomy partner Team (all)
Pass or kill a gate A C C C I
Add a module to scope A R C C I
Hardware price ceiling and BOM C I A/R I I
Pilot design and farmer consent A C I R I
Data source and licence approval A R I C I
Swipe sideways to see the whole table.

Back to contents

23. KPIs and benefits

The team's own UX targets are the first KPIs; the farm-outcome KPIs follow once baselines exist.

KPIs
KPI Target Source of the target When
Time from landing to Your plan, new user Under 5 min [VERIFIED: team overview §17] G1
New users reaching Your plan unaided 4 of 5 [VERIFIED: team overview §17] G1
Screen-reader flows completed (sign-up, add field, read plan) All, iOS and Android [VERIFIED: team overview §17] G1
Screens that differ between app and website None except native navigation [VERIFIED: team overview §17] G1
Exposed bare months, chosen vs current rotation −50% [DESIGN]; baseline U-24 1 season
Pilot farmers logging operations weekly 3 of 10 [DESIGN] G2 + 6 weeks
Sonda days paired with SMAP/IMERG 60 days, 2 fields [DESIGN] HW gate 1
Swipe sideways to see the whole table.

Dis-benefits to watch: extra data entry for farmers; over-reliance on the tool (mitigated by the "Talk this over with your adviser" list [VERIFIED: team overview §2]).

Back to contents

24. Kill criteria

These are written before the work starts, so evidence, not momentum, decides, the same discipline the backtest used: pass marks first, results second. Thresholds are [DESIGN] and are fixed at the start of WP1.

Kill criteria
ID Gate Stop if Data source
K1 G2 Fewer than 3 of 10 pilot farmers log operations weekly after 6 weeks → drop modules 6–9, keep the comparator Operation logs
K2 G1 O1 not met after one redesign round Usability sessions
K3 HW gate 1 Full Lite BOM above the price ceiling set at WP1, or readings cannot be paired with SMAP/IMERG for the same pin EVT data, quotes
K4 HW gate 2 Pro certification only by the PTCRB/carrier path with no funder for it → ship Lite only Test-lab quote
K5 G2 No institutional partner or grant shows interest in writing → stay a free open-source tool, stop the paid tiers Partner pipeline
Swipe sideways to see the whole table.

Back to contents

Sources

Repository sources (open source, in the Aluni repository): README.md, docs/ROADMAP.md, docs/DEVICES.md, docs/FIELDS.md, docs/DATA_MODEL.md, docs/DATA_SOURCES.md, docs/DEMO.md, docs/ONBOARDING.md, config/field_profile.yml, app/models/farmer_profile.rb, docs/proposal/ (worked example, cross-check, claims and the validation scripts), docs/data_sources/pest_weather_rules.md and nutrient_removal.md; the team's Aluni Project Overview (15 Nov 2026). The proposal follows the Integrated Project Proposal format: the five-case business case, NASA TRL, AACE cost classes and a known/unknown register.

Web pages

  • FAO: What do we really know about the number and distribution of farms (Lowder, Skoet, Singh 2014)
  • People Food and Nature: Defining and estimating smallholder and family farms
  • CIMMYT repository: farm number estimates 2020/2030
  • ReliefWeb: FAO Status of the World's Soil Resources (2015)
  • UN News: Soil erosion must be stopped
  • FAO Global Soil Partnership: SWSR 2026 launch
  • DevelopmentAid: FAO report warns soil degradation is worsening
  • DigiKey UK: ESP32-S3-WROOM-1-N8R8
  • RS UK: Quectel BG95M3LA-64-SGNS
  • Farnell IE: ATECC608B-SSHDA-B
  • SOS electronic: SHT45-AD1F-R2
  • NovaLynx: Rain and snow store
  • METER Group: 2025 G.A. Harris Fellowship price sheet
  • Waveshare: BG95-M3 Zero
  • Hologram: pricing
  • Telnyx: IoT data plans
  • JLCPCB: PCB prototype assembly pricing
  • MarkReady: FCC certification cost and SDoC vs certification (secondary)
  • Climate FieldView pricing page; softwares.com: FieldView pricing; yespress.io: Climate (secondary)
  • Capterra: Farmbrite pricing; FitGap: Farmbrite (secondary)
  • farmOS: Managing assets; farmOS: Inventory tracking; The Drop Times: farmOS project; SourceForge: farmOS (secondary)

Added for the worked example, BOM and budget

  • BLS: Software developers, OOH (2024 median pay)
  • BLS: Agricultural and food scientists, OOH (2024 median pay)
  • Economize: GCE e2-medium pricing; VPSBenchmarks: e2-medium (secondary)
  • Requesty: Gemini 2.5 Flash on Vertex; Maxim: Gemini 2.5 Flash cost (secondary)
  • RobotShop: Seeed S-Soil MT-02B
  • NovaLynx: Rain and snow store
  • Adafruit: Voltaic 6 V 6 W panel
  • BatterySpace: LiFePO4 18650 1,500 mAh
  • DigiKey: TI BQ24650
  • RS UK: Hammond 1555F2GY
  • Little Bird: DFRobot SEN0601
  • exchange-rates.org: EUR/USD; currency.me.uk: GBP/USD
  • Repo: farm_data_42.0_-93.75.csv / .json, config/demo_farms.yml, config/crops/agronomy.yml, app/services/rotations/comparator.rb, app/services/rotations/timeline.rb, app/services/crops/rotation_check.rb, docs/proposal/ames_worked_example.rb
  • Backtest data for Ames, Iowa and York, Nebraska (fetched 15 Nov 2026 by docs/proposal/validation/fetch.rb): NASA POWER daily, hourly and climatology; GPM IMERG via POWER; Daymet single pixel; CoCoRaHS daily reports; USDA RMA cause of loss 2001–2025; USDA NASS QuickStats; USDA CDL; SSURGO (Soil Data Access); FAO GAEZ v4; Landsat and Sentinel-2 NDVI via Microsoft Planetary Computer
  • Zhao, Zang et al. 2022, Nature Communications 13:4926, data on figshare 20290923 (CC BY 4.0)
  • Renard et al. 1997, USDA Agriculture Handbook 703 (RUSLE); Renard and Freimund 1994 (R factor); McCool et al. 1987, 1989 (LS); Richardson et al. 1983 (erosivity exponent); FAO-33 (Ky); FAO-56 Tables 12 and 22 and Eq. 76

Prices are list prices read on 15 Nov 2026, excluding tax, duties and freight; each is confirmed again before a purchase order.

Back to contents