The Hyperstition Corpus and the show: a bridge (2026-10-20, ET)

Source: KB's Downloads, corpus_bundle/ (the engine and a 291-node genesis snapshot: 212 entities, 79 claims, 76 sources, 472 relations, an 840-row hash-chained ledger; the validator passes clean on a scratch copy) and Hyperstition Corpus.html (the current explorer: 612 graph nodes, 431 entities, 181 claims, 124 sources, 1,187 relations). The corpus's own labels and grades are used here as given; none of its claims were re-verified for this note.

What the corpus is

A working knowledge, social, network and market graph "for two artists studying hyperstition". Append-only JSONL (sources, entities, claims, relations) plus a hash-chained ledger; agents contribute by writing a patch to inbox/, and ingest.py validates the whole patch against a scratch copy and writes nothing if any rule fails. Four principles: loop, affect, authorship (optional in the art, mandatory in the ledger), hyperstition (veracity and force are independent). Two layers: a fiction layer (declared, never a living target, never a forged record) and a ledger layer (nothing deleted; a wrong record is superseded). Its scale tiers are S1 (a scene), S2 (a niche) and up.

The ladder, applied to the show

Baked-reference candidates (direct and indirect; KB's call)

The architecture

Two parts, kept apart. The engine (build/, schema/vocab.json, the ledger and the patch/ingest rule) has nothing to do with any node; it is a boilerplate. An instance is data in that schema. Instance A is the Hyperstition Corpus (documented real cases). Instance B would be the Jazz Cigarettes / Department of Jazz corpus: personas (Jazz Cigarettes, Gice), works (the trailer, the episodes), the show's operators, the Telegram ledger and the meme-vibes register (which already uses graded rows and literal/metaphor tags) as claims. The link is a thin bridge of relations that already exist in the vocabulary (evokes, exemplifies, applies, tested_at, targets): a show scene evokes a case; a designed claim tested_at a case. Updating either instance updates the other's view without merging them. The show's story graph, the register and the candidates lane are the same pattern (append-only, hash-chained, candidates gated).

Limits and open

Instance B, built (same day)

The engine run on this show's own material, KB-only at internal/corpus-jc/ (rebuildable from the repo by seed/seed_jc.py; not in git). 196 entities, 32 claims (29 designed fiction, 3 analysis), 306 relations, 21 sources, a 555-row hash-chained ledger; the validator passes (0 errors, 25 warnings). Each entity carries a frame state: 63 at -1 (the canon: 24 personas, 15 texts, 8 site products...), 119 at 0 (81 craft rules, 14 operators, 12 concepts...), 14 at +1 (the trailer, the 12 episodes, the Ep1 cut).

What the three levels look like as data:

Open for the next pass: wire the uncited rules and operators through a Pendulum round 3; connect the remaining isolated nodes (Focus, Mr. Atrium, Mr. Ortiz, five site products, the essays); read the six /texts on the site and the JC lore into claims; replace the explorer-derived id list for instance A with its data files.

After Pendulum round 3

Rebuilt from the repo (23 sources, 202 entities, 33 claims, 355 relations; validator clean): rules named in an episode rose from 16 to 41 of 81, operators exemplified by an episode from 3 to 8 of 14, and second-order nodes touching an episode from 20 to 57. The five complete-internet lenses (hypertext, interest, typed ontology, content graph, discovery) are concepts in instance B, each carried by the episodes that show it. The bridge now checks against the full 612-node snapshot of instance A (55 links, none missing). See research/archive/2026-10-20-pendulum-round-3-report.md.