| type | index | ||||||
|---|---|---|---|---|---|---|---|
| title | AGENTS.md — how to navigate the Position_Is_Arithmetic paper series | ||||||
| description | Entry guide for an agent or reader landing on the public Shannon-Prime paper repo. The LEDGER is the master claims index; this file gives the read order, the honest-status convention, and the receipts-first discipline so a claim can always be traced to a command + a commit. | ||||||
| tags |
|
||||||
| timestamp | 2026-06-21 00:00:00 UTC | ||||||
| resource | LEDGER.md | ||||||
| sp_status | ACTIVE | ||||||
| sp_gate | none | ||||||
| sp_commit | c74afe8 | ||||||
| sp_repro | read LEDGER.md (the master claims index); follow each row to its commit/gate |
This is the public, receipts-first paper series of the Shannon-Prime project. It holds the narrative, the papers, and the master claims ledger. The implementation lives in the linked repositories; this repo is the story and the receipts.
If you are an agent or a careful reader, enter here, in this order.
README.md— the public story: what "position is arithmetic" means, the PROVEN headline results with their caveats, the discipline, and the PROVEN-vs-WIRED-vs-DESIGN honesty section. Start here.LEDGER.md— the master claims index. This is the single source of truth. Rule: nothing appears in any paper, the README, a post, or a talk unless it is a row here, with its scope attached. Every headline number traces to a ledger row, and every row carries its number, config, gate, caveat, status, and the commit chain that produced it. If a claim is not a ledger row with a command, it is not a claim.METHODOLOGY.md— the three rules (bit-exact when off; no number without a command; scope travels with the number) and the three gates (parity / deflection <2% / poison). This is why the numbers are believable; every paper cites it instead of re-deriving it.SERIES.md— the manifest and release cadence: the paper order, each paper's front-door receipt, and the rule for slotting a new release in.HISTORY.md— a hashed, tiered commit log (MEM-OKF style). Tier-0 is a LUT of milestones; the git short-hash is the content address — follow any row withgit show <hash>for the full commit. Regenerate withpython tools/okf_history.py gen --repo . --out HISTORY.md(the tool lives in the lattice repo).- The papers themselves, under
papers/— each is a self-contained module (README.md+paper.md+receipts.md+repro/), independently citable, carrying its own one-command reproduction.research papers/holds the longer-form preprint drafts.
Every claim in this repo is tagged. Read the tag before you trust the number.
| Tag | Meaning |
|---|---|
| [PROVEN] | Measured and gated; the number has a ledger row and a reproducing command. Citable. |
| [WIRED] | Implemented and gated in-engine/in-core, running behind a flag — not a public citable headline, not "live by default." |
| [DESIGN] | Specified, with its falsification gates pre-stated; not built. |
| honest negative | Measured and refuted, kept on the record (e.g. the 32k NIAH MISS, the nine refuted recall signals, the inert content-side number-theory levers). |
Do not inflate a tag. In particular, on this public repo:
- The 32k NIAH MISS (ledger
01-R9) stays visible — it is the honest anchor for the recall budget. - The NIGHTSHIFT curator is gated-GREEN on a synthetic gate (criteria 1–4); its live in-distribution criterion is PENDING — do not call it "live."
- The byte-exact forward buys auditability (exact arithmetic / cross-machine determinism), not compression, speed, or size; it is default-off (gated-GREEN), with the OFF state byte-identical to the bf16 gold.
- The diffusion judge is not claimed here — it is unproven and held in the drawer; any 95.6% figure that appears in upstream notes is the external llama.cpp oracle's number, not ours.
- When in doubt, under-claim.
- Bit-exact when off. Every mechanism is a strict no-op in its default state — the baseline is provably the original network, so any on-state result is a controlled delta.
- No number without a command. Every figure is reproducible from a ledger row.
- Scope travels with the number. Model, context, corpus, and what it does not generalize to ride along with every figure. This is proof-of-mechanism on one dev host (RTX 2060, 12 GB) — not a scaling study, not multi-model, not independently reproduced. Say exactly that.
- No silent gate revision. A gate that can't be met surfaces upstream; gates are never quietly retuned until a number passes.
- Honest negatives stay attached. A result with its caveats is one a reader can trust without re-deriving the authors' incentives.
Shannon-Prime records its knowledge under SP-OKF — its profile of Google's Open Knowledge
Format v0.1: every knowledge .md carries type + receipts-first frontmatter
(title/description/tags/timestamp/resource + sp_status/sp_gate/sp_commit/sp_repro), validated by
okf_validate.py. The commit history is itself a tiered, content-addressed store
(HISTORY.md via okf_history.py). The same discipline that gates the code gates the
docs: a claim you cannot trace to a command does not ship. The tooling and the full
MEM-OKF anti-rebuild store live in the lattice repo
(shannon-prime-lattice, tools/, papers/SP-OKF-PROFILE.md + papers/MEMORY-OKF-PROFILE.md).
This repo is the public face. The implementation is in the
main project:
shannon-prime-lattice (umbrella +
STATE + contracts + the OKF tooling), shannon-prime-system
(the math core), and shannon-prime-system-engine
(the inference engine: the served chat, the recall head, the gates). The architecture ground truth
is the lattice repo's papers/RFC-XBAR-auditable-latent-crossbar.md.