Private beta · Runs on your machine
A memory archaeologist for your own filesystem. It digs through what you've already built, assembles the pieces into something runnable, and then grades that artifact by whether it actually runs.
You ask “have I built something like this before?” The scroll answers from your own disk.
Private beta — macOS, Python 3.8+. No hosted version, no public install yet · hello@llmdorado.com
The dig
Most of a dig costs nothing. The model is called once, for the patterns that already fired, and everything it produces has to get past a filter that has seen the rest of the hoard.
MayanRick walks a directory for cheap filesystem signals and matches them against a library of eleven glyphs. Nothing is sent anywhere at this stage.
Only the glyphs that fired reach a model. The pin is grok-4.5 for its 500k context — a dig wants the whole cave in one prompt, which is why the higher version number is not the better model here.
Fired patterns get assembled into a runnable artifact and written to the hoard alongside notes on what it is, which pieces it came from, and why they belong together.
IDF-weighted cosine over stemmed, boilerplate-stripped tokens, judged against the three nearest neighbours — not just the closest one, which is how a paraphrase used to slip through.
Every artifact is executed with HOME and CWD redirected into a temp directory and API keys stripped. Verdicts: real, needs-args, no-data, hollow, broken, slow.
The hoard
The index marked all 38 of these ok: true, which only ever meant a model returned a file. These are the verdicts after each one was executed.
The verdict that matters
A script that prints nothing and a script that works looked identical in the index. The gate separates them: hollow means it exits clean and produces almost nothing. One artifact in the hoard is hollow. Before the gate existed, it read as gold.
Measured, not asserted
Against the real hoard it caught 4 of 4 known duplicates, including a paraphrase that shared no vocabulary with its original, with no false positives. The substring check it replaced caught none of them — four artifacts built on one idea in different nouns all passed.
Built for builders
The dig reads your files on your machine. There is no upload step and no hosted index — the only thing that leaves is the prompt for a glyph that already fired.
Glyph matching is filesystem signals and string work, so a scan of a directory that holds nothing interesting costs nothing. The model is the second pass, not the first.
hoard lists every artifact grouped by whether it needs input, with a handle that is round-trip tested against the command it prints. hoard try runs one against a temp HOME before you trust it.
mayanrick, dms, hoard, reality-gate and ghost-hours-digest are ordinary CLIs. An MCP server exposes the same commands to any host that speaks the protocol.
Scope
Working today
Not yet
Elsewhere
Two papers
Why AI companies forgot to use the human brain, and what to do about it. Abstracts, the key claims, and the full PDF.
Read it →The method
Seven questions, two minutes. Work out the role you already know how to describe, then hand it to a model.
Try it →Founding access
Early access to the Rafiki method and templates, and the setup walkthrough. Free waitlist, or ask about a founding place.
Join it →Fixed-price engagement
Where your team's AI workflow leaks quality, time and trust — diagnosed against a failure taxonomy, with a 30-day plan.
See the scope →The scroll is in private beta and access is handed out one at a time. Tell me what your cave looks like and how much of it you've forgotten.
Ask for access