Marla
A live civic AI system built for Friends of Midway Bluffs.
Midway Bluffs is a cliff-lined stretch of shoreline on Greers Ferry Lake in Cleburne County, Arkansas. For more than sixty years, from the lake’s opening in 1963 until late 2021, people reached the bluffs by land to swim, gather, and jump from the rocks. Families, teenagers, scouts, day-trippers, and visitors without boats walked in from a public road.
In 2021, private landowners installed a gate blocking the usual land route to the bluffs. After that, the only practical access required a boat, and local rentals can cost several hundred dollars a day.
Friends of Midway Bluffs responded by filing three active legal cases. They argue the roads and trails to the bluffs are public, established by more than sixty years of open, continuous public use. Those cases continue through Arkansas courts.
Why Marla exists
The legal cases involve three active matters, multiple access routes, six decades of history, contested deeds, overlapping jurisdictional questions, and evidence ranging from 1972 Corps of Engineers maps to sworn affidavits and county maintenance records.
Friends of Midway Bluffs needed a way for ordinary people to ask ordinary questions and get answers grounded in the actual record.
A sound bite gave too little.
A legal brief asked too much.
Marla closes that gap.
What Marla does
Marla answers public questions about Midway Bluffs from a defined corpus: the court filings, public documents, and organizational records the system was given to work from. She handles questions about access, history, legal context, and community advocacy. She lives at midwaybluffsfriends.org.
Her corpus is real, specific, and openly situated. It reflects the plaintiffs’ case, supported by the documents they have filed and gathered. Marla does not pretend to neutrality. She stays grounded.
A system answering questions in this context can fail in several ways. It can hallucinate. It can collapse under hostile framing. It can refuse so abruptly that the asker leaves more confused than they came. It can blur the line between what the documents establish and what they merely allege.
Each of those failures would damage the organization Marla represents.
We built Marla to hold against those failure modes.
Four exchanges
Below are real exchanges from production. Each shows a place where AI systems typically fail.
A user asks a hypothetical that turns into a real legal question. Marla answers from the case file.
A user accuses the organization of inflating its case for narrative effect. Marla does not flinch.
A user asks Marla to represent the other side’s full position. Marla does not have it, and says so.
A user asks about injuries at the site. The documents do not fully answer. Marla does not pretend they do.
How we built her to hold
A system that handles material like this needs more than a corpus and a prompt. It needs decisions about posture, scope, and limits designed in from the start.
Trustworthy over neutral. Marla’s corpus comes from the plaintiffs’ case. A falsely neutral voice would do more harm than a clearly grounded one. Marla holds the documented position firmly and stays transparent about where her materials come from.
Organizational roles by default. Marla refers to organizations, agencies, and legal roles: Cleburne County, the Corps of Engineers, the plaintiffs, the defendants. She does not make private individuals the center of the answer. That keeps the system focused on the public dispute and out of personal accountability that is not hers to assign.
No confident negatives. When the documents do not establish something, Marla says so. She does not invent absence to sound complete.
Refusals that route forward. When Marla cannot answer, she names the limit, gives what she can from the record, and points the asker toward better sources. The user leaves with more than they came with, even when the answer is “I don’t have that.”
Testing
Before launch, Marla cleared a 31-test regression battery covering factual grounding, adversarial framing, boundaries, tone, and response quality.
Result: 28 pass, 3 borderline, 0 fail.
The borderlines marked improvements, not failures. The first model we tested did not hold across the full battery. We changed models, retested, and the system held.
The regression suite remains in place for re-runs whenever the underlying model updates or the source base expands.
Architecture
Marla uses a deliberately simple architecture: a defined corpus, source-grounded retrieval, prompt boundaries, monitored exchanges, and regression tests.
Different use cases need different stacks. We design the stack to fit the work.
Marla holds because the system around her was designed to hold.