Can a listing carry a receipt you can re-run?

This is not a sample. It is everything in it.

A marketplace page that shows five listings when it has one is the defect this page exists to not have. Below is the entire current contents of the FleetPrompt registry: one entry, sealed by us, not installable. When that changes, the count changes — and nothing else on this page is allowed to change it.

Trust as a replay, not a badge.

The evidence bundle is a film. A listing is a manifest plus a film, and the install path runs through the film rather than around it. Inputs are captured at the boundary — the world an agent claims to do and the scenario a particular run fed it are different documents, and mixing them is a typed rejection. Model output enters as a claim, canonicalized on observation, so the film stays byte-exact even though the model behind it is not deterministic. The reduction is schedule-independent, which is what makes replay it yourself a coherent offer instead of a hope.

What it does not prove. A replay shows what ran, not that it should have. A faithful film of a harmful agent replays perfectly. Films make review checkable; they do not make review unnecessary, and nothing on this page should be read as saying they do.

Three commands. None of them replay yet.

This is the designed flow, numbered because the order is the point. An install engine already exists and runs all seven of its steps — permission review, manifest verification, dependency resolution, policy check, deploy, connect, audit (§7.1). Not one of them is a replay. The tag on each step below is the truth about it today, and the reason to publish a drawing of unwritten code is that it lets you check the build against it later. If the transcript ever disagrees with the drawing, that will be written here too.

Three things exist here, at three different rungs, so they are not in one list.

A single status for this domain would be false in one direction or the other. One artifact answers on a deployed host, one is a codebase with a suite, one is a document. The rung chip on each says which.

The gate this whole page is about does not exist.

Sealing a run works, at alpha. Nothing consumes a seal. So no install has ever been stopped by a replay, which means the proposition on this page has never once been exercised — stated first, at full size, because it is the thing a reader would otherwise assume works.

The suite count below was produced by running the suite, not typed onto this page.

This build shells out to mix test, parses what it prints, and refuses to emit the page if the number disagrees with the frozen record beside it. The excluded tests are published as excluded rather than quietly counted as passing, and what happens when they are included is published too.

Both rows were produced by executing the command in them and parsing what ExUnit printed. The two excluded tests are FleetPrompt.KilnPollWorkerLiveTest — the live crystallization pipeline — and they fail in setup because they need a service that is not running here. That is why they are tagged out, and publishing the tagged-out run beside the default one is the only way a reader can tell a green suite from a suite that is green because two tests were removed from it. The build refuses to emit this page if the two runs stop accounting for each other.

And the rows that are a name rather than a thing.

Same table shape, honest column. built · superseded is used where something is deployed and running but replaced by the argument above — taking neither credit for it nor pretending it was never there. The last row is the one that matters most.

Where this stands, and what would move it.

Status

live_deployed — The superseded registry service is deployed and answering. app.fleetprompt.com returns 200, GET /api/health returns {"status":"ok","version":"0.1.0","service":"fleetprompt"} — the deployed build naming the same version this page is stamped with — and POST /mcp returns 200. The search page renders and its agent grid is empty.

Last verified

2026-08-17

Source

Five measurements on 2026-08-17. (1) mix test — 112 tests, 0 failures, 2 excluded. (2) curl https://app.fleetprompt.com/ — 200, a LiveView search page with an empty result grid. (3) curl /api/health — the JSON above. (4) curl POST /mcp — 200. (5) curl /api/agents/search — 500 on every query tried, including with no query at all.

Limit

This establishes that a service is deployed and that its own health endpoint agrees with this repository about the version. It does not establish that the service is useful: its JSON search endpoint returns 500 on every query, its registry lists nothing, and the model it implements — trust computed from four weighted signals — is the one this page argues against. It does not establish anything about replay: no film has ever gated an install, because the gate does not exist. And the one identity measured across two operating systems was derived by a single implementation twice, which is not two implementations agreeing.

Next rung

external — A publisher who is not us sealing a listing, and a machine that is not ours replaying it to the same identity. Building the gate would make that possible; it would not make it true. Until then the count on this page stays at one and the count is ours.

Listings anyone can install

0. The registry holds one entry, it is ours, and it is not installable — there is no replay gate for it to pass. The deployed service's own search grid renders empty, and its JSON search endpoint returns 500 on every query.

Films that have ever gated an install

0. The install engine runs all seven of its steps and not one of them is a replay. Sealing works at alpha; nothing consumes a seal.

Three rungs, three lists. Nothing asks you to publish to a registry with nowhere to put a film.

If a number on this page is wrong, that is the message we most want.

This page is generated from a record and gated, which catches a value that drifted from its witness. It cannot catch a witness that was wrong to begin with — a count on this page was wrong for months and a gate would not have found it, because nothing had ever run the suite and compared. If you ship agents, the other useful answer is what a receipt would have to contain before you would stake an install on it. That shapes the manifest format.

Or file it in public: github.com/c-u-l8er/fleetprompt.com/issues. Both reach the same people.