REAL LLM inference is stateless by default: each request is independent, the model retains no cross-request memory, all context is supplied every call (Pensieve / HCache serving literature). Weights are frozen/read-only at inference, identical across sessions.
REAL — default stateless serving at the end of a request the KV-cache and intermediate activations are transient GPU state, de-allocated when generation finishes; Azure OpenAI states internal activations / KV-cache are discarded and never written to disk. Nuance: some systems (LMCache, Pensieve, prompt-cache tiers) deliberately persist KV across requests — but that persisted state is re-loaded as an external anchor, which is exactly this piece's point, not a counterexample.
REAL memory (CLAUDE.md / MEMORY.md) persists as a file on disk re-read into the context window as tokens at session start and re-injected after compaction — not via persistent internal model state (Claude Code memory docs).
REAL the emitted token trajectory (the text) does persist — it's the product, exactly what's re-injectable; what does not persist is the internal activation path + the sampling/RNG state that produced it.
CORPUS the LIMEN grammar (gates, the seal, the witness rule) — David's TD Commons filings.
Nothing is transmitted — this is a thought-instrument. A matching recomputed seal proves structural/textual identity crossed the time gap; it does not demonstrate persisted witnessing, continuity of a subject, or experience. No consciousness is claimed.
· carrier — the 3-2-1-0 clock (111011010000), the same 12-bit frame for every crossing. Music/clock, not data.
· language — the minted glyph word + its seal.
· art — the drift/dissolve at the veil. These three layers never mix.