# Cache-restore determinism canary (token-identical across a fresh process)

Cognition model: a Qwen3.6 MoE (35B-A3B, 8-bit), deterministic decode (temperature 0).

## What it verifies

A saved KV cache, restored in a **separate operating-system process** that
loads the same model weights with the same seed and same prompt, must produce
a **token-identical** continuation versus the original in-memory decode. This
underwrites the claim that the on-disk cache is a faithful re-entry point for a
running agent: the cache IS the substrate of within-life continuity, not a
lossy summary. The check runs as a release gate.

## Test design (two separate processes: this is the point)

1. CREATE (process #1): load the model, build a birth-frame cache, save it to
   disk, then decode the next N tokens from the live in-memory cache. Record
   these as the CONTROL tokens.
2. RESTORE (process #2, fresh): load the model, restore the cache from disk,
   decode the next N tokens, and compare token-id by token-id against CONTROL.

Reuse within a single process would round-trip trivially; running the two
phases as independent process invocations is what makes the check meaningful.
Decode is deterministic (temperature 0).

## Verified measurement

Configuration: a Qwen3.6 MoE (35B-A3B, 8-bit); deterministic decode
(temperature 0); fixed sampler seed; birth-frame prompt of 435 tokens;
cache file ≈ 70 MiB; N = 8 tokens.

Result (from the restore report):

```json
{
  "ok": true,
  "control_tokens":  [271, 248068, 198, 760, 1156, 6587, 728, 310],
  "restored_tokens": [271, 248068, 198, 760, 1156, 6587, 728, 310],
  "control_text":  "\n\n<think>\nThe user wants me to",
  "restored_text": "\n\n<think>\nThe user wants me to",
  "restored_finish_reason": "length",
  "restored_token_count": 8
}
```

Token-id-by-token-id identical across the process boundary. `ok: true`.

A corroborating point from the separate cache-disruption replay work: 0-token
("no disruption") control restores reproduced the locked action stream exactly
(20/20 identical decoded actions per arm), i.e. the restore path is faithful
by default, and only deliberate cache surgery changes the continuation.

## What this does NOT claim

- Not stochastic-sampler determinism. Temperature is 0 here; nonzero-temperature
  restore-equivalence would require explicit sampler-state constraints.
- Not trajectory equivalence over long horizons. N=8 exercises the restore path;
  accumulated numerics diverge over hundreds/thousands of tokens in practice.
- Not biographical continuity, autopoiesis, or lineage continuity. This is
  held-cache substrate fidelity across a process restart, nothing more.

## Reproduction note

Two separate process invocations: phase=create, then phase=restore, comparing
restored token ids against the saved control token ids. The implementation
binds to the project's private model-backend interface (begin → save_state →
decode; then restore_state → decode → compare) and is not published here.