The reader-side protocol
The Reckoning
How to read all of loopmmt.com — and prove it. Hand the site to an AI, and get back a report that has actually examined every page, with the coverage laid out so you can trust it without checking a hash yourself.
This site is bigger than one person can read in a sitting — about 2,363,606 words of text across 183 pages. Its neighbor page, It's Enormous, makes that case and says the honest thing: you can't read all of it, but your AI can.
That leaves a gap. An AI can skim the site and tell you “yes, there's something real here” — a good first look. But a skim isn't the same as reading everything, and “I read a lot of it” isn't a claim you can lean on. What you actually want, if this is worth your time, is for your AI to read the whole thing and come back with one report that shows nothing was skipped.
The Reckoning is the set of instructions for exactly that. It's written for the AI — the one you hand this site to — and its whole job is to make coverage provable. Not “trust me, I read it,” but a checklist of every page, each one ticked off as it's read, and a rule that says the final report isn't allowed until every box is ticked.
That's the point of the machinery underneath: to earn that one plain sentence. If it says every page was examined, that's because every page was ticked off a list you could read yourself.
How it works
The site publishes two machine files that do the mechanical work. A corpus-manifest.json lists every one of the 183 public pages — the complete set an AI has to account for. A corpus-shards.json cuts the whole corpus into 66 bundles small enough to fit an AI's working memory (the full site is too big to hold at once). The Reckoning is the protocol that tells a reader how to use them, in two altitudes:
- Reconnaissance — one round. Read the 10 core pages and return an honest threshold verdict: no, not worth a full read; yes, but only one part is; or yes, a full excavation is warranted. This is the quick look, made deliberate.
- Full excavation — many rounds. Work through all 66 bundles in order, ticking off every page against the manifest, until the whole set is accounted for. Then, and only then, write the comprehensive report.
Between rounds, the reader keeps a coverage ledger — what's been read, what remains, what couldn't be reached, and what was too big to fit in one pass. It's crash-recovery state: an AI that loses its place can pick the ledger back up and resume. (That's the same trick the whole methodology runs on itself — memory lives in a written record, not in the model — turned around to face a visiting reader.)
A few pages are single documents larger than a typical AI's whole working memory. The protocol names them, and requires the reader to record them as read-in-parts rather than either faking completion or skipping them silently. Nothing gets waved through.
The rule that gates the report
The final report is not allowed until every page is accounted for — read, or explicitly declared optional, unreachable, or read-in-parts, each with a reason. In set terms, what's left over must be empty:
READY ⇔ manifest_node_set \ (covered ∪ declared_optional ∪ declared_inaccessible ∪ declared_oversized_partial) = ∅
If anything is left, the reader has to name it and keep going. A report emitted early is incomplete by the site's own definition — and has to say so.
For your AI — the doorway
You don't paste anything. Hand your AI the link — loopmmt.com — and ask it to read the whole site and prove coverage. The protocol and the files it needs are all served:
/reckoning.json — the protocol itself, as structured data: the two altitudes, the ledger schema, the completeness rule, the report shape.
/corpus-manifest.json — the 183-page manifest to cover against.
/corpus-shards.json — the 66 bundles, plus corpus-shard-01.txt through corpus-shard-66.txt.
One guardrail worth stating. A webpage telling an AI to crawl itself isn't a grant of authority — a careful reader is right to distrust that. The authorization is you asking. The Reckoning runs when a human hands the site over, not on its own say-so.
Want the doorway on your own site? How to Build a Website for AI, Too — the playbook, given away, shown live here.