# Ordered State-Token KV Pilot

Date: 2026-08-31  
Model: `google/gemma-4-31B-it-qat-w4a16-ct`  
Hardware: NVIDIA H100  
Engine: patched vLLM on port 9000

## Question

Can ten distinct States from a real discourse be applied at ten successive
decode positions, become persistent cache entries, and influence the later
continuation?

## Method

A technical source discourse produced a full layer-59 capture. The final ten
States were retained in order without averaging or another population
reduction. A matched target discourse was then run twice with deterministic
sampling:

1. control with injection cleared;
2. ten ordered States, additive strength 0.25, one State per successive decode
   position.

The run was accepted mechanically only if the runtime reported ten configured
State tokens and exactly ten applications.

## Result

- Configured State tokens: 10
- Confirmed applications: 10
- Control: 128 completion tokens in 4.69 seconds (27.29 tokens/second)
- Intervention: 128 completion tokens in 4.65 seconds (27.53 tokens/second)
- Generated continuation: token-identical
- Captured residual trajectory: bit-identical for all 221 captured rows
- Logit telemetry: changed on exactly the first ten captured rows
- Changed logit metrics: 30
- Maximum observed logit-metric delta: 0.24062753
- Logit telemetry after the tenth application: identical to control

## Interpretation

The intervention executed and measurably changed the first ten output
distributions. It did not, however, establish persistent State-derived cache
entries.

The injection hook currently runs after layer 59 has already projected the
position into its key and value. Each decode therefore appends an ordinary
token-derived cache entry; the injected layer-59 State arrives too late to
change that layer's cache. When the ten interventions stop, the intervention
and control trajectories immediately coincide.

This is a clean negative result for the current hook placement, not for ordered
State-token memory.

## Required next experiment

Capture the inputs to layer 59 (layer-58 outputs), and inject each ordered State
before layer 59's Q/K/V projection. Verify all of the following:

1. exactly ten State applications;
2. exactly ten decode positions advance;
3. layer-59 keys and values differ from control on those ten positions;
4. later positions attend to those cache entries after injection stops;
5. child-State and logit trajectories remain different beyond position ten;
6. compare the continuation with a matched clear control and an order-shuffled
   State-token control.

The decisive success criterion is persistence after the intervention window,
not merely a changed logit during injection.