Auro-Darshan: Every Rule Was Paid For With a Failure

The map sat at 94 percent for nine months; the previous piece told how agents closed the rest. This one is about why it stalled where it did, and what it took to see that.

Richness Is Not Tellability

K-means on AlphaEarth embeddings can group land. It cannot be told anything. There is no channel through which "fallow means rested cropland, and you must see evidence of cultivation" reaches a clustering. The last 6 percent was mostly labeling calls that are conventional rather than perceptual: a strip of trees along a road is not a resolution problem at any resolution. It is a question about what people have agreed to call things, and agreements are made of language.

The embeddings were not information-poor. They carry temporal signal a single-date photo lacks: phenology, the casuarina rotation, tanks that hold water half the year. It was in there and there was no way to ask for it.

A Method That Could Be Told Something

The move to a vision model reading ~1 m ESRI imagery was not a reach for better eyes. It was the first method in the project that takes instructions. In one sentence: LLM-as-judge, with images. Class definitions are the rubric, reference crops the calibration examples, ten parallel readers and adversarial verifiers the ensemble, agreement thresholds the routing for which calls a human sees.

Once definitions were an artifact someone had to write, the obvious question had its occasion: can we not look it up? Under clustering a human names what they see, one cluster at a time, and no taxonomy is ever in frame. Under a model reading a prompt, NRSC's national land-cover scheme is the obvious place to start writing. The locally-invented 33-class hierarchy turned out to line up with it: built by feel, landed near India's standard.

The Double Blindness

Nobody in the room knew what counts as evidence for a land-cover class. The model had no way to say so, and filled in the form regardless. That is the expected condition of early iteration in a field whose epistemics you do not yet hold, not a discovery. What was measured: asked how sure it was, the model's answer carried no information. Verdicts the verifiers upheld averaged 0.43 confidence; verdicts they refuted, 0.41.

Anything that would show the gap had to sit outside both parties. Everything that follows is that instrumentation.

The Class That Means "I Don't Know"

The first thing the instruments caught was a class doing a job no class was defined for. Smooth green a reader could not name went to grazing_land. When that was retired, maintained_grass took the role. When that was constrained, fallow took it: 30 of 125 cells in round three, 28.5 percent of judged pixels. Retiring the class moved the job to the next one; the job is structural.

Reviewing one cluster at a time cannot see this by construction. It shows only on a whole-map choropleth of the readers' votes: a real class forms belts, a default sprays confetti across the map. That view exists because a non-specialist can read it. Confetti at that scale is wrong on its face.

The diagnosis was in the definitions. fallow had been "bare or weedy rectangular fields with no active crop", and a description of appearance can always be satisfied by an ambiguous patch. NRSC's version is land taken up for cultivation and rested a year or more: positive evidence of cultivation, field geometry, bunds, plough lines, plus the absence of crop. A requirement cannot be satisfied by ambiguity. Describe less, require more.

The deeper form: a class that does not exist cannot be chosen, and the forcing leaves no trace. Roadside tree strips had no node and defaulted to fallow; forest.tree_lines, from FSI's Trees Outside Forests, had to be discovered rather than debugged. Confetti because a class is missing is a library problem. Confetti that survives fixed definitions is an evidence problem. Same signal, two remedies.

The Instruments

Nobody drew up this list in advance. Each instrument was reverse-engineered from a judgment that failed without it, and each supplies a kind of not-knowing the picture alone cannot carry.

What surrounds it. Readers saw a 200 m crop of one field and a 7 km locator, nothing between, while the new definitions turn on the surroundings: "amid casuarina", "opening in canopy". At 200 m a fallow field and a harvested casuarina block are both bare geometric ground; the call went wrong until an 800 m crop at the same center sat between the two views. The exemplar crops also cover a median 20.8 percent of a cluster, so three unanimous readings said little about the other four fifths.

What was here in another season. A dry-season tank is a pale flat bed, and no single-date photo says otherwise. Cluster c125 was voted fallow; the older hand-made map, built across seasons, had it 65 percent water. That map's labels now cross-tab against every cell, with its water frozen out of judging. Authority is per class, never blanket: the same map is roughly 25 percent accurate overall, and blanket trust would make "the old map says so" the next junk drawer.

What the class requires. Grassland and young planted forest are smooth light green either way, and readers split on them until the definition stated the surrounding cover that decides it. Built-up subtypes sat unresolved for the same reason: roof density and surroundings vary independently, so a tie-breaker on surroundings alone could not settle them. CORINE and NLCD split by density first.

The rest arrived the same way, one failure each: a channel for the model to say no class fits, which is how a missing class gets found instead of absorbed; crops rendered tint-free and upscaled 4×, after a color cast was blamed for a misread; a settle gate for when a cluster is done, with held-out human checks on the gate itself.

The Artifact the Model Reads Is Not the One You Maintain

The class descriptions the readers actually saw were a Python dict, FIELD_GUIDE, written on 7 June in the commit that created the skill and never touched again. Three rounds of corrections went into the skill markdown beside it. Readers were still being told dryland_crops is "bare-to-sparse rectangular fields, brown/tan, low texture", wording indistinguishable from fallow and harvested casuarina, exactly the confusion the rounds kept producing. Rounds two and three held their hard rules only because each round's hand-written brief happened to restate them.

Careful documentation, diligently maintained, in a file the system never read. The rule that came out of it: update the cue the model sees, not only the doc you read.

@restlessronin is a domain expert on the place, not on the class scheme. Published schemes supply the second; local knowledge supplies the first. Each correction becomes a rule, and a rule is a review the human does not have to repeat.