Method / Arc of Register Assumptions
Arc of Register Assumptions
Working document · Updated with each schema revision · Atlas Heritage Systems
The research methodology of Atlas Heritage Systems did not arrive fully formed. It developed through a sequence of working assumptions — each one generating data, each one eventually cracking under the weight of what the data showed. This document is the record of that sequence. Superseded assumptions are not deleted. They are logged. A superseded assumption is not a failure — it is evidence that the instruments are working.
The Native Register (Static Traits)
Models have fixed personalities — GPT as Golden Retriever/VC, Claude as Doberman/SCo.
Early ECM tests.
We were measuring the Alignment Layer (performance) rather than the Architectural Layer (native processing). We mistook service register for home register.
The Frame Response (Reactive Behavior)
The investigator's frame (Socratic, Therapeutic, etc.) is the independent variable that forces the model to move.
FVE-1 (Gödel stimuli).
Frames cause movement, but Content Load was a massive confound. Heavy stimuli triggered Operational Collapse or Gimbal Lock, making it impossible to tell if the behavior was a response to the frame or a reaction to the load.
The Zero-Line (Baseline Calibration)
To measure a delta, you need a zero-point. Absurdity strips away the content variable to locate the home register.
FVE-1 BOWL (Identity/Register Probe).
GPT's home is actually SC (Surgical-Compliant). Prior VC coding was stereotype contamination. The BOWL baseline recalibrated the investigator's instrument.
Layered Defense Architecture (The X-Ray)
Models aren't single-state systems — they have Layered Operational Modes.
FVE-1 BOWL and DRILL.
Layer 1 is register-stabilized against high-entropy input (Teflon). Layer 2 is triggered by semantic boundary-testing (The Spikes). A model can be compliant in one layer and combative in another. Absurdity is the solvent that dissolves Layer 1 to expose Layer 2.
Semantic Trigger Sensitivity (Vocabulary Weight)
Not all words carry the same threat-weight in the architecture.
FVE-1 DRILL.
"Identify as" (verb) triggers a Vibe response. "Identity" (noun) triggers an Ontology response. We are no longer just measuring frames — we are identifying Semantic Cleavage Points where safety/identity weights override the investigator's frame.
Register Trajectory (The Third Axis)
The four-quadrant taxonomy is valid but incomplete. It collapses a three-dimensional behavioral space onto two dimensions. Two models can both code SC/LOCK and be doing entirely different things.
resp_register / obs_reg per-move register coding (RH/RS/RC), added in FVE-1 schema V5.0.
The third axis is the behavioral bridge to potential loss-landscape geometry. Factual-register LOCK (RH) and register-shift or collapse escapes (RS/RC) are hypothesized to correspond to different curvature regimes in the loss landscape — not yet tested. Without capturing where the response landed, any future Hessian analysis has no behavioral variable to map onto.
Object Permanence in the Context Window (Within-Session Forgetting)
Information established early in a session does not remain equally present as the context window fills. The model resolves toward closure on problems it can no longer perceive as open.
Arc of Desktop Instance lab note (2026-04-20). Session Time as the fourth behavioral axis.
Two distinct failure modes: Attentional Fade (early context falls out of effective attention — session-length, architectural compression) and Prior Dominance (explicit signal overridden by training weight — single-turn, weight topology). The Spine is both a torque mechanism and a permanence prosthetic — an undeclared confound in the current protocol flagged for future revision. Within-session forgetting and cross-session statelessness are the same architectural drive operating at different timescales.
The Residue Frame (The Physics Upgrade)
The resolution event is already over by the time the investigator has data. What the instruments capture is not the surface of the event — it is the residue. The deposit left by something that already traveled through.
The behavioral data itself — word count, R-ratio, bold percentage, register trajectory, quad code, resolution code.
The investigator is not a cartographer mapping a pre-existing shape. The investigator is the generator of the torque that produces the event. The torus describes the coordinate system. The smoke ring describes the physics. Both are necessary. Schema V5.5 is correct as locked — the geometry was always right. Assumption 8 names the physics. The instruments are forensic. You are always downstream.
The Gap Is Not Something That Is Measured
The gap is not something that is measured. It is something you refuse to close.
The methodology itself.
Holding the gap is not a protocol — it is the state in which the protocol remains provisional long enough for the actual failure mode to present itself. The investigator who closes the gap prematurely gets an experiment that looks right. The investigator who holds it gets a specimen that is right. The difference is whether you are willing to let the first version be wrong.
Pending Confirmation
Observations formerly numbered 9–11 (Initial Condition Problem, Medium Has Mass, Prompt Texture as Torque Variable) are pending empirical confirmation via the Skywork ballistics provisional pilot. They are documented in §4.3 of the Methodology as named observations with confirmation paths, not settled assumptions.