The 15 to 25 percent gap is mostly physics
A half-acre plot walked once on a single-frequency handset carries about a tenth of its own area as error before anyone has done anything wrong. That is not a training problem and it does not have a training fix. It has a measurement fix: six lifecycle captures of the same boundary, reconciled by weighted least squares, constrained by the edges the plot shares with its neighbours in the Club, on a handset whose second civil signal cuts the error in half again. This page runs that in your browser over 179 synthetic holdings and 954 simulated captures, and shows the places it does not work as well as the places it does.
One boundary, six observations, one answer
A capture is an observation, not a fact. This plot holds 6 walks of the same line, taken across a season on whatever handset was in the technician’s pocket. Switch them in and out: the heavy line is the weighted least-squares consensus over whatever is left on, and the interval under it is what that evidence can actually support.
With no captures switched on there is no estimate at all, only the contracted 0.82 ac on the paperwork. That is the situation the product exists to end.
Waiting for captures.
Two independent implementations, and one that does not fit
The headline claim is only worth something if it survives being computed twice. The closed-form error model in the engine and the 4,000-trial Monte Carlo that generated this world share no code, and they do not share a noise model either: the Monte Carlo treats the corners as independent, the engine does not. So they disagree on the level and agree on the reduction, and the table below shows both rather than quoting whichever one flatters the claim. Then the same closed form is checked against every simulated capture in the world, where one of the two device classes does not fit, and saying so is the point.
| Rung | Closed form | Monte Carlo | Delta |
|---|---|---|---|
| One walk, single-frequency 0.5 acre plot, sigma 5.0 m, 72 fixes at 2.5 m spacing, decorrelating over 7 m. | 12.59% | 15.76% | -20.1% |
| Six captures reconciled 6 independent looks at one boundary on 6 different days. Random error falls as 1/sqrt(N), so 2.45x. Assumes no shared path offset. | 5.14% | 6.31% | -18.5% |
| Six captures, dual-frequency Sigma 5.0 m to 2.20 m. Area error is linear in sigma. Handset refresh, no external hardware, no new training. | 2.26% | 2.78% | -18.6% |
| Plus club shared-edge constraint 75% of the perimeter is shared with a neighbour and therefore observed twice. Variance halves on that portion: sqrt(1 - f/2). | 1.79% | 2.22% | -19.5% |
| Total reduction, no external hardware | 7.04x | 7.10x |
Read the delta column before the levels. The two methods sit about 20% apart on every rung, and in the same direction: the generator observes 4 vertices independently, while the engine integrates an error that stays correlated along the whole walked track, and independence is the more optimistic assumption. That offset is close enough to constant that it cancels out of every ratio, which is why the reductions agree to 1.9% and the totals to 0.8%. The reduction is the claim this study makes; the absolute level is the engine’s, and it is the model every other figure on this page uses.
On a 0.5 acre plot at Rs 60,000 an acre, that is disputed value falling from Rs 3,777 to Rs 536 per plot. Two of the three rungs cost nothing at all: the six visits are already in the calendar and the shared edges are already in the ground. Only the handset rung costs money, and it buys more than the other two together.
Eight rules, run over the whole world, with their working shown
Two families. Measurement rules judge whether a capture is a sound observation at all. Agreement rules judge whether sound observations disagree with each other or with the paperwork. Keeping them apart is what stops a settlement conversation from starting with “your technician was careless” when the real answer is that a half-acre plot cannot be walked to better than a tenth of itself.
A shared edge is the only line two people measured
Everything on the previous panels reduces random error. None of it can find a boundary that was deliberately moved, because a claimant’s six captures agree beautifully with each other. The constraint that finds it is structural: inside a Club, plots abut, so one line is recorded twice, by two people, from two sides, on two different days. That is a stronger statement about where the line is than any number of walks by one person.
| All four planted claims, at this handset | Overlap | Floor | Verdict |
|---|---|---|---|
| OVL-01 - 39/4 vs 435/9 | 0.023 | 0.179 | inside noise |
| OVL-02 - 459/2 vs 309/6 | 0.050 | 0.110 | inside noise |
| OVL-03 - 195/4 vs 71/2 | 0.040 | 0.107 | inside noise |
| OVL-04 - 442/3 vs 21/8 | 0.090 | 0.026 | raisable |
All four of those are genuine encroachments: the generator moved somebody’s corners in every one. At sigma 4.9 m, 1 of 4 can be put to a farmer without the accusation resting on receiver noise. That is what a handset refresh actually buys on this tab: not a prettier polygon, the standing to enforce a contract.
| Corner both records claim | Apart |
|---|---|
| 8,0 | 4.95 m |
| 8,1 | 4.99 m |
A shared corner is one physical point. Where both records carry it to the last decimal, the two parties agree and there is nothing to discuss. Across the whole world 798 of 812 shared vertices are identical; the 14 that are not are exactly the corners the four planted claims moved.
| Option | 459/2 | Change | Value |
|---|---|---|---|
| Uphold the claim | 0.823 ac | - | - |
| Split the difference | 0.784 ac | -0.039 | Rs 2,347 |
| Restore the abutting line | 0.747 ac | -0.077 | Rs 4,604 |
- 1Topology check flags the pair. Neither farmer is contacted yet.
- 2Noise floor at the recorded handsets decides raisable or not. This pair: not yet raisable.
- 3Both technicians re-walk the shared line together, on the same day, as one capture each.
- 4The agreed corners are written back to both records at once, so the two polygons stay exactly adjacent.
- 5Settlement runs on the reconciled areas, with the interval attached.
The gap between the second and third tile is the entire argument. A raw geometric test flags 5 overlapping pairs in this club; after the noise floor, 0 are worth putting to a farmer. Shipping the first number would have a field team knocking on 5 doors about slivers a receiver invented.
80% of the perimeter is shared with a neighbour and therefore observed twice. Variance halves on that portion: sqrt(1 - f/2).
A 0.7 m systematic path offset, for example always walking the outside of the bund, puts a 6.2% floor under all of this. No rung removes a bias; only agreeing the line with the neighbour does.
Walking consistently on the outside of a bund is not carelessness, it is where the path is. Six captures by one technician reproduce that offset six times and report a beautifully tight interval around the wrong line. Two technicians walking the same line from opposite sides disagree by twice the offset, and the disagreement is the only signal that the offset exists. That is why a Club is the unit of reconciliation and a plot is not.
The wrong instrument for a boundary, the right one for a season
Free 10 m imagery is the obvious way to check a field team, and at this holding size it is the wrong instrument. The grid below is drawn at true scale over a real holding from this world. Count the pixels that sit wholly inside it, then count the ones straddling two farmers who are paid separately.
At 10 m this plot carries a time series (16 clean pixels) but not a boundary (11.4% area error, 36% of the area in mixed pixels).
| Size band | Plots | Retained at 10 m | at 3 m |
|---|---|---|---|
| 0.5-1 ac | 58 | 60.4% | 86.8% |
| 1-2 ac | 52 | 68.8% | 90.1% |
| 2-5 ac | 54 | 74.5% | 92.1% |
| 5-10 ac | 15 | 80.7% | 94.1% |
| Sub-0.5 hectare, separable under the 75% rule | 69 | 1.4% | 100.0% |
Computed over every pixel that touches cultivated land in this world. A mixed pixel covers two or more separately contracted holdings, so it cannot settle either one. Across this world 17.0% of the 17,420 pixels over cultivated land at 10 m cover two or more separately contracted holdings, against 5.5% at 3 m. A mixed pixel can arbitrate neither holding, whatever the classifier says about it.
Faint rules are a 5 day revisit across the 134 day window this lifecycle covers, from the onboarding walk two weeks before sowing to post-harvest on day 106: roughly 13 to 26 usable passes, against six technician visits. Averaging the 13 clean pixels inside this holding gives a time series that answers did the crop emerge, did it die back, did it recover, and answers it between visits. None of those questions needs to know where the boundary is, which is exactly why the imagery is good at them.
- study-input
Roughly 7.5% of sub-0.5-hectare fields stay separable as distinct polygons at 10 m resolution, against roughly 88% at 3 m.
Carried as a study input. Cite the underlying remote-sensing source before publishing this figure. Checked against separability.subHalfHectare.
- study-input
Dual-frequency L1+L5 GNSS on mid-tier Android reaches roughly 1.75 to 3 m horizontal accuracy, against 5 to 8 m for single-frequency L1.
Carried as a study input. The simulator uses 2.1 to 2.4 m and 4.9 to 5.6 m, inside these bands. Checked against devices[].sigmaM.
- study-input
Dual-frequency L1+L5 GNSS reached mid-tier Android handsets from 2022 onward.
Carried as a study input. Cite handset or chipset availability data before publishing.
- study-input
Contracted against actual area varies by 15 to 25% under eye-estimated capture.
Carried as a study input. The synthetic contract set reproduces a comparable spread; see derived.contractVariance. Checked against derived.contractVariance.
- study-input
Benchmark gross crop value of INR 60,000 per acre used for value-at-risk.
Carried as a study input. The synthetic per-crop economics put cotton at INR 62,050 per acre, so the benchmark sits inside this world's own range. Checked against crops[].grossValueInrPerAcre.
The separability numbers in the table above are not from this list. They were measured on this world’s own geometry, which is why they disagree with the carried figure: 1.4% separable at 10 m here against the 7.5% the study input claims. Where a measurement and a carried claim differ, the surface shows both and says which is which.
- Every reconciled boundary, interval, residual and gain on this page comes from src/lib/traverse/reconcile.ts, run in your browser on the fixture geometry.
- Every overlap area is an exact polygon intersection computed by slab decomposition at render time, not a number read off the fixture.
- Every tolerance band, noise floor and value at risk is derived from plot size, handset sigma and the decision the stage feeds.
- The anomaly findings are produced by running eight of the engine's ten typed rules over all 179 plots, which is why the sweep has a progress bar.
- Every capture is SIMULATED from seed 20260912. No device GNSS was read and no field visit took place.
- Villages, clubs, farmers, survey numbers, contracts and settlements are fabricated. Warangal district and Telangana are real only as a geographic anchor.
- Synthetic ground truth is shown where it helps, and it exists only because a generator drew it. In the field there is no truth column, which is the whole reason an interval is attached to every number.
- Five claims are carried as study inputs rather than computed, and are labelled as such wherever they appear.
World clock 2026-03-14T09:30:00+05:30 - seed 20260912 - schema v1 - equirectangular about each village origin
SYNTHETIC DATA. Every company, village, club, farmer, plot boundary, capture, contract, rate and settlement in this file is fabricated for a product study. No real person, holding or agreement is represented. District and state names are real Indian administrative names used only to anchor the belt; village centroids are approximate synthetic points inside it and do not correspond to any actual settlement.