active pilot · one city, five verified layers, canonical desks planned — 2026-08
I run a research program whose deliverable is a city. Not a city rendered — a city whose every claim can be checked against a source, or marked UNKNOWN. The rendering comes later. The program itself doesn’t touch a texture.
Most simulated-city work starts from the opposite end. Reference photos, concept art, “does this feel like a real street.” That’s a legitimate craft, and I’m not running it. What I run is a desk: one city at a time, five subsystems, each with a state machine pulled from public records, each with a named piece of the simulation that has to consume it, and each gated behind adversarial tests it has to survive before it’s trusted with anything. The city that comes out the other end isn’t a mood board. It’s a stack of claims, each one either backed or labeled UNKNOWN.
One city, one desk, no mixing
The first rule is the one everything else depends on: research one primary city deeply per cycle, using only public, attributable sources, and never let facts from one city bleed into another city’s model. Los Angeles is the pilot. Other cities are planned, not started — the desk explicitly keeps them out of scope until the LA patterns stabilize, rather than half-building five cities at once and calling all of them in progress.
The reasoning is the same reasoning behind every other measured-or-UNKNOWN system I run: a fact is a fact about the place it was measured in, not a fact in general. A curb-sweeping cycle timed in one city’s sanitation schedule is not evidence about a different city’s schedule, even if the two look similar on a map. Mixing them would let the desk answer questions it hasn’t actually researched, by borrowing confidence from a city it did research. The rule against that isn’t caution for its own sake — it’s the same discipline as refusing to let a stale note answer a question nobody re-checked.
What comes out of a desk is described, deliberately, as an interpreted layer, not a warehouse. The desk doesn’t hand the simulation a dump of FilmLA permit records or LADWP outage feeds. It hands the simulation a state machine, a confidence number, and a list of things the simulation is allowed to do with that state. Raw data doesn’t get to render. Only claims that have been triangulated do.
Five subsystems, five state machines
By the last recorded delta — signed August 28–29, 2026, the most recent state this piece draws from — the LA desk had five active layers, up from roughly two at the start of the pilot, and five state machines modeled, up from roughly one. Each layer follows the same shape: sources, a state machine, a named consumer, and a falsifier list.
Production and street occupation, sourced from FilmLA’s own Q2 2026 production records — 4,711 on-location shoot days that quarter — runs as prep → low/high-impact active → teardown → residue/recovery. The distinction that matters isn’t “is there a shoot,” it’s low-impact versus higher-impact: only the higher-impact state actually rewrites parking, curb, pedestrian, and lane states downstream. A low-impact shoot on a residential block shouldn’t reroute traffic in the simulation any more than it does in life. This layer’s source count moved 40 to 46 across the delta, confidence 85 to 87 percent.
Curb-state, sourced from LADOT’s colored-curb rules, the Code the Curb dataset, LADOT’s Zero Emission Delivery Zone reporting, and StreetsLA’s ADAPT sweeping routes — four independent official sources, not one — treats curb space as a resource with a clock, not a static strip of asphalt. Loading windows open and close on a schedule. Sweeping runs before → active → after, with a recovery pattern once it’s done. Nightlife blocks get passenger-loading surges at specific hours; destination curbs saturate. Sources moved 51 to 55, confidence 89 to 90.
Seismic preparedness and recovery, sourced from LADBS, USGS, Cal OES, LA Emergency Management, and LADWP, runs baseline → possible alert → shaking → localized assessment/service state → staged recovery → repair memory. The word doing the work there is localized: the desk explicitly rejects a generic citywide-destruction model in favor of heterogeneous recovery that varies by building and by district, backed by retrofit-status data, ShakeAlert, statewide earthquake early-warning coverage, and NotifyLA’s documented scope and limits. Sources moved 61 to 66, confidence 91 to 92.
Air quality, sourced from South Coast AQMD and EPA AirNow, refuses to treat “air quality” as one dial. Ozone and heat, smoke-relevant PM2.5, dust and ash, visible haze, and AQI are modeled as five separate, sometimes-independent variables instead of one wildcard “bad air” flag. As of the last recorded state, the desk read an ozone advisory driven by heat, with the smoke and dust/ash advisory categories both inactive — three of five variables doing different things at once, which a single AQI dial can’t represent. Sources moved 132 to 135.
Utility power-outage and restoration, sourced from LADWP’s own reliability reporting, runs reported → assigned/queued → crew en route → crews working → restored, anchored to LADWP’s system-level reliability numbers for the fiscal year: SAIFI 0.60, SAIDI 82.74 minutes. The desk is careful to keep those as system-level anchors, not address-level truth — it explicitly preserves the outage map’s own stated limitation that its coverage is estimated and partial-neighborhood, and keeps a major-alert-issued state distinct from a confirmed address-level outage state, instead of collapsing the two. Sources moved 144 to 147.
Across the whole pilot, the aggregate source count moved from 40 to 147. That’s the actual unit of progress here — not “the city looks more finished,” but “the number of independently checkable claims underneath it went up.”
The falsifier is the gate, not the finish line
None of those five layers advances on triangulation alone. Each one carries its own falsifier list — a set of adversarial and negative-control tests specific to that subsystem, and a desk only advances when a layer’s evidence has been converted into implications the simulation can actually consume, not when the sourcing merely looks thorough.
The curb layer’s falsifiers include a quiet-residential negative control — a block that should show no sweeping-recovery pattern and no loading-window churn, used to check the model doesn’t invent activity where the sourced behavior says there shouldn’t be any. The seismic layer’s falsifiers include an unaffected-district negative control alongside a same-camera baseline-and-recovery proof, so a shake event in one district can’t be shown to silently contaminate a district the sourced data says wasn’t affected. The outage layer’s falsifiers include an “alert issued but unaffected” adversarial case specifically because the major-alert coverage and the confirmed-outage state are different facts, and a naive model would collapse them.
That’s the same shape as a control that’s structurally incapable of failing being worthless as a control — a desk can’t just cite four official curb sources and call the layer done. It has to show the model that four-source layer produces, when run against a case built specifically to break it.
The ceiling nobody pretends isn’t there
Two of the five layers — air quality and utility outage — hit a hard confidence ceiling at 95 percent, and stayed there even as source counts kept climbing past that point. That’s not framed as a bug waiting on more sourcing. It’s recorded as a stated limit: confidence capped at 95 with exhaustive public sourcing, on the record, not as an aspiration to eventually close.
I take that seriously as the most interesting fact in the whole pilot. It would be easy for a research desk to keep adding sources indefinitely and let the confidence number creep toward 100, implying that enough public data eventually resolves everything. This one doesn’t do that. It states, in writing, that some irreducible uncertainty survives exhaustive public sourcing — partial neighborhood coverage on an outage map, jurisdiction and sensor-type mismatches on an air-quality reading, sampling bias that no amount of additional official data fully resolves. Whether more sources would ever move that number past 95, or whether 95 is closer to the real epistemic limit of public urban data, is UNKNOWN — the desk hasn’t run the experiment that would tell it which.
Where the swarm’s own infrastructure failed
The desk’s plan called for a GitHub repository as the durable transport layer between research sessions — the place a new session would read to pick up where the last one left off. I won’t put the repo path here; that’s exactly the kind of address this site doesn’t publish. What I’ll give you instead is the receipt, with the address swapped for a placeholder:
shaurya@sps ~ % curl -sS <transport-repo>/JOIN.md
404
Repeated, on a read the desk itself logs as harmless — no auth error, no rate limit, just consistently not found. Whether that’s a permissions problem, a moved repository, or something else was left open rather than guessed at; the record states plainly that the architectural implication is unclear. What actually happened is that the desk fell back to Notion as its working transport and kept operating, while the intended durable layer sat there, unreachable, exactly where a new session would have looked for it first.
That’s the same lesson this whole program is built to catch in the city data, turned back on its own plumbing: a transport layer you designed on paper isn’t a transport layer until a session actually reads from it successfully. Intent to be durable and being durable are different facts, and the only way to tell them apart is to try the read and record what came back — which is precisely what happened here, and precisely why the desk noticed the gap instead of quietly assuming the GitHub layer was fine because it had been built correctly once.
What I don’t know yet
I don’t know whether the 95 percent ceiling is a property of this particular pair of subsystems or a general limit that every layer eventually hits once its source count gets high enough — only two of five layers have reached it so far, and that’s not enough to call it a law. I don’t know what happens to city-truth isolation the first time a second canonical city desk actually starts, since the discipline has only ever been tested against a single active city with nothing running in parallel to tempt it. I don’t know how the named consumers — traffic, NPC schedules, story branching, the online district replication — actually translate a passed falsifier battery into rendered behavior; the desk names them as consumers and stops there, and the translation step is undocumented. I don’t know whether the GitHub transport gets fixed or gets formally abandoned in favor of the fallback that’s already proven itself by staying reachable. And I don’t know how this desk handles a rare event it hasn’t modeled yet — wildfire, flood, civil unrest, a cascading utility failure — since the five layers built so far are the ones with the cleanest public sourcing, not necessarily the ones a city actually needs most.
Until one of those gets measured, it stays what everything else on this desk that hasn’t been checked stays: UNKNOWN.
— One Desk Per City