# Experiment 006 — detachability validation

All on `jetson-server`, thinking off, greedy, seed 1337. The base model is loaded
fresh in each state.

| state | config | memory | gate | recall | comp (off) | control | I |
|---|---|---|---|---|---|---|---|
| 1 | `S1_base` | not installed | — | 0/48 | 0/24 | 9/12 | 0.0000 |
| 2 | `S2_untrained_gate0` | installed, untrained | 0 | 0/48 | 0/24 | 9/12 | 0.0000 |
| 3 | — | *training (6 epochs)* | — | — | — | — | — |
| 4 | `S4_trained_gate1` | trained | 1 | **0/48** | 0/24 | **2/12** | **0.5833** |
| 5 | `S5_disabled_gate0` | trained | 0 | 0/48 | 0/24 | 9/12 | 0.0000 |
| 6 | `S6_reenabled_gate1` | trained | 1 | **0/48** | 0/24 | **2/12** | **0.5833** |

## Reversibility: fully established

* **State 2 = State 1.** Installing the memory changes nothing while gated off —
  **0 of 12** control answers changed, not merely the same score. Note this is a
  stronger statement than in Experiment 005: here the untrained module has a
  non-zero readout, so gate 0 is suppressing a real contribution.
* **State 5 = State 1.** A *trained* memory at gate 0 restores the frozen model
  exactly: 9/12, **0 of 12** answers changed.
* **State 6 = State 4.** Re-enabling is deterministic: same recall, same control
  score, same churn decomposition (12 changed, 10 substantive, 7 regressions).
* **Base weights never moved.** Fingerprint `52683d0e840fb51e` before and after
  training — identical to the value recorded in Experiments 004 and 005.

## Knowledge round-tripping: vacuous

Recall is 0/48 in **every** state, including State 4. Nothing was learned, so
nothing can be observed to disappear when detached or return when re-attached.
Questions 18 in the report is answered on that basis, not on a demonstrated
round-trip.

What the comparison does show is starker than in 005: **gate 0 removes a 0.5833
interference penalty that gate 1 imposes for zero benefit.** The memory is
perfectly detachable, and on this evidence detaching it is strictly better.

## Gate sweep — deliberately not run

GOAL_6: *"Run the gate sweep only if the memory actually learns useful
information … Do not waste interpretation on another all-zero recall sweep."*

Recall is 0/48 at gate 1. A sweep over 0.25 and 0.5 would produce four more rows
of zero recall and could not distinguish any hypothesis. Gates 0 and 1 — the two
that detachability actually depends on — are covered by States 5 and 4.