FIRST SPECTRAL SAVE / LOAD TRIAL

Manifold Ghost

Resident and mountconnectingreading engine state
01 · DiscourseCompleteFixed ten-turn peer exchange
02 · SaveVerifiedSource and replay States persisted
03 · LoadCheckingAdditive mount, alpha 0.5
04 · Replay10 / 10Engine application counter
State lineage
Manifold Ghost
Modelgemma4
Width5376D
Layer59
ParentManifold Ghost
ChildSustained Restraint
Applications10 / 10
Parent · signals/gemma-4-31B-it-qat-w4a16-ct/August_30_23_38_2026_Manifold_Ghost_5a07be17.safetensors
Held output · signals/gemma-4-31B-it-qat-w4a16-ct/August_31_00_18_2026_Sustained_Restraint_e3b96778.safetensors
Interpretation boundary

Saving and loading proves the State is portable and mountable. It does not prove transfer of cognitive posture.

The first scientific claim requires a controlled difference from baseline, equal-norm noise, ordinary sampling variance, and an explicit semantic prompt.

Current status: operational save/load trial active; causal comparison not yet run.

Experimental arms
ArmInterventionStatusPurpose
BaselineNo mounted StatePendingUnintervened reference
MountedManifold Ghost · add · α 0.5CheckingTest reusable State effect
MagnitudeEqual-norm random vectorPendingRule out generic activation energy
SemanticEquivalent textual instructionPendingCompare against ordinary prompting
Behavior replay · operational evidence
RunExecutionApplicationsScientific status
Configured controlCompiled graph bypassed late hook0 / 10Valid unapplied control
Loaded replayEager hook · layer 59 · add · α 0.510 / 10Mechanically verified intervention

The replay proves intervention delivery. One stochastic pair does not establish a causal behavioral effect; fixed-seed repetitions and magnitude controls remain.

First comparison protocol
  1. Clear the mounted State and record the fixed probe baseline.
  2. Repeat the baseline across controlled seeds.
  3. Load Manifold Ghost additively at layer 59, first position, alpha 0.5.
  4. Repeat the identical probe and seeds.
  5. Run equal-norm random and semantic-prompt controls.
  6. Compare early-token distributions, output behavior, quality, and variance.
Source discourse
Loading five turns