# the420code — what we do NOT claim

One page, everything we refuse to overstate. If a claim isn't here as "do claim",
treat it as not claimed. This is the document a hostile expert should read first —
because we'd rather hand them the limits than have them find them.

## Across the board

- **Mechanisms are proven; scale and speed are not.** Every receipt in
  [`proofs/`](proofs/) demonstrates that something *works*. None of them claims
  production performance, SOTA results, or behaviour at large scale.
- **Proprietary ≠ proven.** Where we keep a method proprietary, that is a business
  choice, not evidence of strength. Secret methods get *no* security or quality
  credit until independently tested.
- **No faked numbers.** Every figure is produced by code in this repo, on data in
  this repo, and re-run in CI.

## AI

- We do **not** claim any model "understands quantum gravity," or solves any open
  physics problem. The knowledge graph marks such problems as explicit dead-ends.
- We do **not** claim a small on-device model out-reasons a frontier datacentre
  model. The win is sovereignty, privacy, and determinism — not raw capability.
- We do **not** claim fluent text equals correct output. Confidence is not truth.

## Knowledge graph

- It is a **toy-scale, hand-curated** demonstration, not a large corpus.
- Doing similarity/reasoning **without model-derived embeddings at scale** is an
  open research problem — proven here only as a deterministic, no-training
  substrate.

## Compression

- Beats general-purpose compressors (xz) **broadly**, but beats dictionary-DEFLATE
  **only on smaller inputs** (leave-one-out: 4/7 files; −2% on aggregate).
- The pure-Python implementation is **slow** (~seconds/KB) — fine for a proof, not
  for production.
- **High-entropy data is incompressible.** Verified: random bytes do not shrink.
  Information theory is the wall; we respect it.

## Notarization

- Hashing proves **integrity and ordering, not wall-clock time.** Trustworthy time
  requires anchoring the root to a **public** ledger (e.g. a Bitcoin transaction or
  an RFC 3161 authority) — the rig supplies hashing, the anchor supplies time.
- Retired mining ASICs **only hash.** They do not run general code, encryption, or
  a new language.

## Encryption

- **No new cipher.** We compose standard primitives. Secrecy of construction is the
  *opposite* of strength (Kerckhoffs's principle).
- The demo uses an HMAC-CTR construction because the native AEAD backend is
  unavailable in this environment; **production must use a vetted AEAD**
  (AES-256-GCM / ChaCha20-Poly1305).
- **Post-quantum** is named as a future KEM (ML-KEM / Kyber), not implemented here.

## Coherence Ledger

- It proves **what happened, not that the AI was right.** A flawless evidence chain
  can still record a bad decision. This is audit/provenance, not correctness.
- **Deterministic replay needs a deterministic model** (temperature 0 + fixed seed
  + same weights). The ledger preserves the inputs; it cannot make a
  nondeterministic API reproducible.
- Time still comes from the **public anchor**, not the ledger itself.

## Determinism

- Demonstrated **same-architecture**. True cross-*architecture* determinism (ARM
  phone vs x86 server) is the next receipt — the integer-only design makes it
  achievable, but it should be *shown*, not assumed.

---

If any item here moves from limit to solved, it moves into a pillar — **with the
evidence attached.** Until then, it stays on this page.