Etherion / Messis / LeFARMER discovery
LeFARMER is a 963M-parameter model evaluated on one exact task: seven prior eight-day blocks from the same parent, region, and crop condition predict the eighth held-out block. The result is a measured benchmark, a public methodology, and the first proof of agriculture as a Data Enclosed Field.
Bounded claim: LeFARMER is the best measured model on the anchored same-parent next-state benchmark. It is not presented as universal agricultural state of the art.
Locked 2022-2023 test
Normalized next-state residual MSE / lower is better
Measured result
963,270,656
parameters
20,000
optimizer steps
1 × L40S
training GPU
No extensive RL
training program
01 / What is actually new
The point is that a published latent world-model principle has been carried into a real agricultural Data Enclosed Field, trained on public satellite trajectories, and opened onto a large specialization surface.
This is an applied LeWorldModel, not another agricultural score wrapped in a dashboard.
LeWorldModel describes a compact Joint Embedding Predictive Architecture for learning future state in representation space. LeFARMER applies that idea to a physical domain where the state evolves under observable environmental forcing: crop systems, soil moisture, vegetation, and weather. The scientific move is the transfer of the world-model object into agriculture, not merely another regression number.
Agriculture is the first Data Enclosed Field because its next state is constrained by measurable biology, environment, and time. The same construction can be tested where state is observable, forcing is physical, and future evolution matters.
963M parameters
A sub-billion shared world-model layer, not scale for its own sake.
20,000 steps · one L40S
A compact measured run leaves substantial room for longer training, broader temporal coverage, and specialist variants.
No extensive RL program
Forward dynamics are learned from the observed physical system without requiring a long reward-training pipeline.
One shared layer → many specialists
Crop biology, climate regimes, weather patterns, and geometry-sensitive spatial sensing are diversification axes, not decorative roadmap boxes.
The public result is a concrete agriculture application of the LeWorldModel idea, not a paper citation used as branding.
The current result uses one L40S and 20,000 steps; the proposed path scales through data, time, specialists, and sensing.
The long-term object is a family of world models for Data Enclosed Fields, with agriculture as the first falsifiable domain.
01 / The exact benchmark
This is the public contract. It makes the result interpretable, falsifiable, and distinct from non-equivalent annual-yield comparisons.
Task contract
Every model receives seven prior eight-day blocks from a single parent trajectory.
It predicts the normalized residual of the eighth block from the same parent, region, and crop condition.
Training uses earlier temporal partitions; hyperparameter selection uses validation; 2022-2023 remains locked until final evaluation.
Overlapping windows are deduplicated by parent and target date before aggregate reporting.
2,196
locked-test transitions
2022-2023
temporal test partition
Locked-test result
Normalized next-state residual MSE
LeFARMER
Best clean result
0.252477
LSTM B=128
+2.93% error reduction
0.260087
XGBoost
+8.58% error reduction
0.276176
Persistence
+18.49% error reduction
0.309742
What the metric means
This is unbounded normalized next-state residual MSE: squared error in standardized state-change units. Zero is perfect; there is no universal "good" value. The relevant scale is the error made by matched models on this exact same task.
0.5025
LeFARMER RMSE
0.5565
Persistence RMSE
Relative improvement on the same task
Honest current limit
At horizon seven, LeFARMER has lower latent MSE than latent persistence on 67 of 96 held-out parent trajectories and loses on 29. The current public artifacts do not yet provide a The three-seed result covers LeFARMER on the 17to25no2223 policy; the corrected LSTM B=128 values are single-run matched baselines, not a multi-seed LSTM confidence interval. This result proves an aggregate early advantage on the defined benchmark; it does not yet prove uniform per-crop dominance, operational decision utility, or a finished agricultural product.
01A / Evidence registry
The registry makes the scaling story measurable without pretending that MSE values from different train-fitted normalizations share one absolute scale. Every row is interpreted against its own matched baselines.
| Training windows | Protocol | LeFARMER | XGBoost | LSTM B=128 | Persistence |
|---|---|---|---|---|---|
| 2,388 | 2015–2019 | 0.3604 | 0.3930 | 0.3652 | 0.4426 |
| 4,824 | 2015–2021 / clean | 0.305731 | 0.329077 | 0.322967 | 0.374173 |
| 6,888 | 2015–2021 + 2024–2025 / test excluded | 0.252477 | 0.276176 | 0.260087 | 0.309742 |
| 9,084 | Includes locked test / leakage control | 0.259323 | 0.282347 | 0.268054 | 0.317414 |
Every row has locked-test n=2,196. LSTM was retrained with matching normalization at native B=128 for each row; the invalid 0.4026 forced-B=1024 training result is excluded. Decisions were regenerated from these corrected scores.
Three-seed evidence
0.25375 ± 0.00590
Mean locked-test MSE and 95% CI across seeds 7, 42, and 123 on the 6,888-window no-test-contamination policy.
0.25694 · 0.25658 · 0.24773
Interpretation
The clean 4,824-window result improves over matched XGBoost by 7.09%. The 6,888-window result improves by 8.58%. The leakage control is labeled as such; its absolute MSE is not used as a clean generalization claim.
Five-country operating domain
Twelve crop conditions
Maize · sorghum · pearl millet · cassava · groundnut · cowpea · rice · sweet potato · yam · sesame · soybean · coffee.
One joint, region-agnostic crop-conditioned model in the public operating domain.
02 / Fully disclosed matched baselines
LeFARMER's internal recipe remains private. The benchmark, protocol, and comparison baselines do not. A researcher can inspect exactly what it beat.
The continuity reference predicts zero residual: the next state equals the latest observed state. That is the correct baseline for an anchored residual task because it asks whether temporal structure beats simple continuation.
Scientific basis
LeFARMER is informed by LeWorldModel: a Joint Embedding Predictive Architecture introduced by Maes, Le Lidec, Scieur, LeCun, and Balestriero. The disclosed principle is simple: predict future system state in representation space rather than reconstruct every observation.
03 / Actual figure set
Each figure is rendered from a real frozen run artifact or is explicitly marked as a roadmap. The interpretation stays with the metric it actually shows.
FIGURE 01
The current clean-generation registry. Every row has locked-test n=2,196 and its own train-fitted normalization; native-B=128 LSTM values are used. The 17to25 row is a leakage control, not a clean generalization claim.
Source: corrected generation benchmark and native-B=128 LSTM artifacts
FIGURE 02
Across one through seven prior eight-day blocks, LeFARMER remains below latent persistence. This is a real latent-MSE curve, not a stylized growth chart.
Source: anchored all-horizons evaluation artifact
FIGURE 03
A real held-out transition trace: condition flags, prediction divergence, context-depth behavior, and test coverage. It preserves traceability without exposing private state construction, raw coordinates, or source queries.
Source: public evaluation-transition artifact
FIGURE 04
Richer temporal coverage, broader operating-domain evidence, and spatial sensing are research directions. They are not claimed as current product capabilities.
Source: roadmap diagram, not a measured performance chart
04 / What the scalar misses
The numerical margin is evidence. The more consequential object is a crop-conditioned predictive state that retains time, expected evolution, divergence, and the public path of evidence behind an interpretation.
0.252477 is the current clean proof, not the complete product thesis.
A tabular model returns a point estimate. A predictive state can preserve a time-indexed expectation and then expose divergence when observed evolution departs from it. That opens an inspectable decision surface instead of a black-box score.
Condition evolution
How crop condition changes through time, not just where it ends.
Water and vegetation response
How public soil and canopy evidence moves alongside the expected trajectory.
Meteorological response
How rainfall, temperature, and radiation contextualize change without becoming a disconnected dashboard.
Emerging divergence
Where observed evolution departs from expectation and becomes a candidate signal for review.
The public observability layer
The depth is delivered as inspectable evidence categories, not as an opaque final score.
01
Observed condition
Vegetation and soil condition with measurement history preserved.
02
Environmental forcing
Temperature, rainfall, evapotranspiration, humidity, wind, and solar-radiation context.
03
Causal agronomic summary
Vegetation movement, moisture and rainfall accumulation, water balance, heat accumulation, and evapotranspiration imbalance.
04
Crop-science context
Growth coefficients, rooting depth, depletion behavior, yield response, and salinity tolerance from calibrated crop profiles.
05
Quality and provenance
Source validity, observation gaps, parent trajectory, time window, region, crop, and split.
06
Predictive trace
Observed evolution, expected next evolution, deviation, and time-indexed surprise.
07
Decision and audit
The evidence path behind irrigation, stress, fertilization, or harvest observations.
Publicly disclosed: causal parent trajectories, source-family provenance, benchmark protocol, matched baselines, and aggregate trace evidence. Deliberately withheld: private state construction, model internals, exact coordinates, query details, and training artifacts.
05 / Public provenance
The benchmark uses real environmental observations. The source families are public; the private representation and construction recipe are not.
NASA SMAP
Public soil-moisture observations contribute the ground-water side of a crop system's evolving condition.
Sentinel-2
Public satellite observations contribute canopy and vegetation evidence across the trajectory.
NASA POWER
Public meteorological observations provide the physical forcing context for the next state.
06 / Research program
The current result was produced after 20,000 optimizer steps on one L40S, without an extensive reinforcement-learning program. That makes it an early systems result with room to scale and diversify — not a finished agricultural product.
Future specialist world-model variants can deepen phenology, stress interpretation, and management observability where crop biology demands it.
Future variants can become sharper for climate regimes and weather patterns that produce different agricultural dynamics.
For tea, fruits, and other high-value crops where geometry matters, Etherion intends to integrate LiDAR spatial insight with temporal prediction.
Measured versus future
Measured now: the anchored next-state benchmark. Future research: crop, climate, weather-regime, richer temporal coverage, and spatial-sensing specialization. The page does not present those future directions as shipped capabilities.
07 / The operating model
Messis is a thought-sorted business model: an agricultural decision engine delivered through the Etherion platform, not a dashboard in search of meaning.
Agents can translate the same evidence into practical instructions for field workers, inspectable context for agronomists, and regional briefs for ministries or cooperatives.
For deep planning, agent-led brainstorming salons can examine season plans, risks, and management choices against the model's evidence before decisions are made.
The organization retains full sovereignty over the raw data Messis produces for it. Insight can travel; ownership of underlying observations remains with the organization.
Etherion platform
The orchestration and delivery substrate Messis is designed to use.
About Etherion
Etherion is a Computer Engineering and Artificial Intelligence Startup, committed to channel the power of world models through principled Scientific Insights in "Data Enclosed Fields" such as Agriculture, Material Science, Organic chemistry by leveraging SOTA parallel computing hardware. Through the usage of JEPA World Models Etherion Aspire to create and deliver at scale SOTA Observability Layers for the greatest actors at the forefront of AI for Science.
Agriculture is the first proof. The same principle extends where state is observable, forcing is physical, and the next state is worth understanding: materials, chemistry, and beyond.