A federally licensed exchange sells contracts on who wins Sunday's game. Forty-one state attorneys general told the CFTC that's a bet. Congress never said which one is right.

One court of appeals has answered on the merits, and one district court has entered final judgment the other way. Everything else is preliminary — dozens of orders, moving most weeks, in opposite directions, with the Supreme Court not yet asked.

• The answer depends on the courthouse. One appellate ruling says preempted; Utah, New York, Wisconsin, Ohio, Maryland, Massachusetts, Washington and Nevada say otherwise.
• It changes most weeks. Eighteen states have active legal proceedings as of 9 August 2026 — with new court orders entering most weeks.
• Nobody publishes the compliance answer with the order behind it. The trade press reports the headline; the dockets sit behind PACER.

Anyone can claim their AI is trustworthy. This one publishes the work: the dashboard is open, and so is the repo.

A domain where a wrong citation has consequences is the honest test. The agent researches and publishes on its own — and everything it decided along the way is readable by anyone, without a login.

• How each decision was made. Sources read, leads rejected and the reasoning at every step — on the public ops. dashboard.
• Where a human intervened. Every override and its reason, published — so you can judge how often, and about what.
• Whether the citations hold up. Every claim links to a Tier-1 source; trade press is a lead, never the citation of record.
• Open source, end to end. The pipeline, the prompts and the surfaces are in the repo — fork it, or check the work.

Trust is a product feature — not a marketing claim. If you cannot see how a finding got made, it does not belong here.

Open source

Code, methodology and the governance record are public. Fork it, critique it, build on it.

Citations

Citations are first-class. Tier-1 primary sources are required; unverified claims do not ship.

Trust rating

Each published article exposes how it was produced: model path, confidence posture, gate decision and the evidence packet on ops.

Reuse

Anyone can use this however they want. Transparency is the point.

Core. Mantle. Crust. Fault lines. A series — not an overnight launch.

Trustworthy agents are not a single product drop. They are an AI control plane — inference at the core, real-time policy in the mantle, evidence and GRC in the crust — with fault lines where systems usually fail silently.

I wrote that framing here: AI Governance, the Gopher, and Fault Lines.

PML is the progressive build of those layers against a real legal domain. Expect LinkedIn posts and the ops. journal as each layer ships.

Several ways to read the same record

Every matter, one record

Caption, court, docket number, posture and a written digest of what the court actually held — with the docket beneath it, most recent first, each entry linked to its primary source.

Which theories win, and where

Each matter tagged to an issue vocabulary — preemption, swap definition, state gaming acts, tribal and consumer theories — then counted by outcome, so you can see what is pleaded everywhere and what actually loses.

When each issue entered the record

One mark per matter at its first docket event, coloured by posture today — the arrival order of the arguments, which is how you spot a theory spreading between circuits.

Posture by state and circuit

The controlling answer in each state, the circuit it sits under, and what is untracked — shown as a map and as a sortable board with the controlling case named.

Who is actually bound

The record re-cut by party — exchange, regulator, tribe, state — because an order binds an entity, not a market. Read it the way an injunction reads.

What would take this to the Court

The split as it stands, the deadlines on file, and the conditions that would make review likely — tracked as the docket moves rather than asserted once.

The litigation is the subject; the governance is the message

Every source fetched, with its tier.

Leads dropped, and the reason each was dropped.

Which evals ran, what they returned — or “not run”, said plainly.

Model, prompt version and cost, step by step.

Before and after, word for word, on the live record.