FIELD REPORT R–001 5 AUGUST 2026 Verified

Prompt reuse / Persistent state

A repeated prefix skipped 2,048 prompt tokens.

Astronomical restored one complete prompt-state block on the second request while preserving generated-token parity on a Qwen3.6 model artifact with multi-token prediction weights.

1first-request miss
1repeat-request hit
2,048prompt tokens restored
Equalgenerated tokens

01 / The repeated work

Long shared prefixes should not be processed from zero every time.

Tool definitions, system instructions, repository context, and conversation history often repeat between requests. Astronomical records exact decoder state at fixed 2,048-token boundaries in a bounded persistent store. A matching future prefix can resume from that state instead of replaying every earlier token.

02 / Measured

Miss once. Publish state. Hit on the identical request.

REQUEST 01 Cold prefill

Lookup misses. The model processes the prompt and publishes one complete state block.

MISS
REQUEST 02 Restored prefix

The identical prefix matches. One block returns directly to live decoder state.

2,048 SAVED

The acceptance compares cold and restored output token IDs, not only a hit counter. This guards against a fast restore that resumes from the wrong model state.

03 / The MTP boundary

Persistent target state and speculative state are not interchangeable.

The current persistent format stores target decoder state. It does not store the shifted request history used by multi-token prediction. Restoring target state into an active speculative request would therefore combine incompatible histories.

Astronomical resolves this without disabling reuse: when persistent prompt storage is available, the request uses target-only execution. When it is unavailable, an eligible highest-logit request can use multi-token prediction normally.

INVARIANT

Never trade correct model state for speculative speed. Reuse the compatible target path instead.

04 / What it means

The cache behavior is proven. Universal speedup is not.

The result proves persistent publication, lookup, state restoration, and output parity through the real supervisor and worker path. The amount of wall-clock acceleration depends on prompt length, model, memory ceiling, existing operating-system file pages, and storage.

A representative public performance claim still requires approximately 1,000 generated tokens and explicit cold/warm methodology. This report therefore publishes tokens avoided, not a universal percentage speedup.