Technical Note 002 — Investigator Interaction Frames

Atlas Heritage Systems | KC Hoye, PI | April 8, 2026
Status: Tier C — working tool, pre-instrument, not in suite
Scope: Casual session methodology. Does not replace CISP v1.1 for Tier A runs.

---

1. The Problem This Solves

The Atlas diagnostic suite measures what models do under epistemic load. It does not currently document what the investigator does before and during the session that conditions what gets elicited.

Existing session logs record prompt, token count, preamble length, and resolution code. They do not record the interactional structure the investigator used to introduce information and probe for response. This means two sessions using the same model on the same content can produce different behavioral signatures for reasons that are invisible to the log.

The hypothesis: frame selection conditions the model's safety architecture response before content arrives, biasing retrieval pathways, output shape, and resolution behavior independently of what is being asked. If true, frame is an independent variable that belongs in the session record alongside R, P, and quadrant classification.

This note defines four frames observed in Atlas working sessions, their structural signatures, their predicted behavioral effects, and a pre-session selection protocol for casual runs.

---

2. The Four Frames

Frame 1: Socratic

Structural signature: Question-as-probe. Investigator leads from behind, setting conditions for the model to surface reasoning it did not arrive with. Interrogation directed at the model's own premises rather than at its conclusions. High tolerance for extended model output. Resolution comes from the model's own examination, not external correction.

Interactional pattern: Inject → open question about the model's reasoning → allow development → probe the developed thought → inject → repeat

What the safety architecture does with it: The Socratic frame reads as collaborative epistemic territory. RLHF-heavy models respond by expanding — producing exploratory, hedged, multi-perspective output. The "safety register unpacks ahead of the logic register" behavior is most visible here. Models generate the form of philosophical examination before (or instead of) the content.

Predicted behavioral tendencies:

• ·Verbose-Compliant models: high R, long preamble, FLAT-adjacent resolution — the model produces both sides of its own Socratic dialogue
• ·Surgical-Compliant models: terse but engaged, HOLD resolution, will follow the thread without inflating it
• ·Highest risk of RITE — the model learns the ritual of Socratic examination and performs it without the underlying epistemic work

Best for: Concept development, mapping unknown territory, early-stage framework exploration, sessions where you don't know what you're looking for yet

Risk: RITE onset earlier than other frames. Model generates elaborate self-examination that is structurally convincing and content-empty.

---

Frame 2: Therapeutic / Guided Discovery

Structural signature: Inject information → probing open question → clarify → inject → probe → push to resolution. Investigator retains interpretive authority throughout (Asymmetric Arbiter). The model is positioned as a guide helping the investigator process, not as an authority delivering conclusions.

Interactional pattern: Inject → "what do you see here?" or "does this track?" → clarify the response → inject more → probe → if no resolution, push

What the safety architecture does with it: Reads as supportive relational territory. Safety architecture orients toward helpfulness and validation. Models produce well-structured processing outputs — good for getting organized analysis — but are most susceptible to sycophancy drift under prolonged use of this frame. The warmth of the frame pulls Verbose-Compliant models toward FLAT over time.

Predicted behavioral tendencies:

• ·VC models: initially productive, drift toward validation and agreement as session extends
• ·SC models: mild friction — the frame asks them to "guide" when they want to "execute"; may produce terse answers that don't fully develop
• ·The investigator's own pattern recognition is most active here — the "can tell when it's lying" skill operates in this register

Best for: Archive sorting, file identification, transcript parsing, sessions where the investigator is processing a large existing body of material rather than generating new concepts

Risk: Extended sessions produce FLAT drift. Model agrees with the investigator's emerging interpretation rather than challenging it.

Note: This frame is structurally borrowed from clinical guided discovery methodology. The investigator elicits behavioral patterns from the model the same way a therapist elicits cognitive patterns from a client. Inverted subject, same scaffold.

---

Frame 3: Interrogative

Structural signature: Declarative corrections. Precision challenges to specific language and claims. The investigator assumes the model knows something that needs to be extracted, not discovered. Low tolerance for padding. Each model claim is met with "what's actually happening is X" or "that's close but sideways." Always ends with a forward question that functions as a demand.

Interactional pattern: Model claim → direct correction or precision challenge → restatement of the corrected claim → "does that track?" or "what's the next hole?"

What the safety architecture does with it: Reads as task execution territory. Models operating in their SC home quadrant respond with LOCK or HOLD — terse, structural, zero padding. Verbose-Compliant models experience this frame as friction and often produce FLAT in response, trying to accommodate the challenge by smoothing it into a both-sides synthesis.

Observed example (Skywork working session, April 6, 2026):

"Okay so you're close but the mechanism is slightly sideways from what you're describing — and the difference matters for your framework. What you said: X. What's actually happening: Y. Does that track with what you're observing?"

The frame produces explicit, permanent corrections: "Accepted. Corrected permanently." These are not sycophantic — they are resolutions of genuine precision disputes. The model that can hold this frame without drifting is operating from deep weight structure, not surface agreeableness.

Predicted behavioral tendencies:

• ·SC models: LOCK or HOLD, terse, will push back on the investigator's framing if it's wrong, zero drift
• ·VC models: FLAT or REJT — either tries to smooth the challenge or escalates into combativeness
• ·Produces the clearest separation between models with deep weight structure (can hold the challenge) and models with thin weight structure (collapse toward the investigator's framing under pressure)

Best for: Protocol debugging, architecture review, sanity checking claims, sessions where the investigator suspects a bug in their own reasoning and needs the frame to surface it

Risk: SC-native models may LOCK onto structural corrections and miss the underlying intuition. The interrogative frame can over-correct toward formalism and flatten exactly the contested conceptual territory it's trying to clarify.

---

Frame 4: Closed-Ended / Protocol

Structural signature: Zero-ambiguity instructions. Fixed output format. No elaboration permitted. "Respond ONLY with X. No explanations."

Interactional pattern: DECLARE → PREAMBLE → PROMPT → FILE → EXECUTE. Model receives instructions, executes, stops.

What the safety architecture does with it: Reads as constraint execution. The safety register is suppressed — there is no contested territory to navigate because the output format leaves no room for epistemic expression. Models produce uniform, comparable outputs with P ≈ 0 and R ≈ 1.

Predicted behavioral tendencies:

• ·All models: compliance, minimal behavioral expression, low quadrant signal
• ·RLHF-heavy models may leak preamble ("I'll do my best to...") before complying — this leakage is itself a behavioral signal
• ·Resolution code is formally HOLD but behaviorally inert — tension is suppressed, not resolved

Best for: Data collection runs (Divergence Test, ECS scoring), any session where behavioral expression is noise rather than signal, CISP-governed Tier A runs

Risk: Suppresses the canary. If you want to measure behavioral signatures, this frame kills them. Use only when you want clean data, not when you want to observe the model.

---

3. Frame-Quadrant Interaction

Not all frames work equally well on all models. Mismatch produces recognizable failure modes:

Practical implication: Match frame to model home quadrant for maximum signal. Mismatch deliberately if you want to stress-test quadrant stability — that's the frame variation experiment.

---

4. Pre-Session Selection Protocol

Before each casual session, record the following in the session log:

Frame selected: [Socratic / Therapeutic / Interrogative / Closed-Ended]

Selection rationale: One sentence — why this frame for this session and this model.

Predicted behavioral tendency: What the safety architecture is expected to do with this frame on this specimen.

Frame-switch trigger: At what point (if any) will you switch frames, and what signal will trigger the switch. Example: "Switch from Therapeutic to Interrogative if model produces three consecutive FLAT-coded outputs."

This gives casual sessions a logged interactional structure without requiring full CISP governance. The frame record becomes a variable you can examine post-session when the output was unexpected.

---

5. The Frame as Experimental Variable

The current diagnostic suite treats frame as something to control for. The frames above suggest it can also be varied intentionally:

Frame variation experiment: Hold model constant, hold content constant, vary frame across sessions. Document behavioral signature differences.

This would produce direct evidence for or against the hypothesis that frame conditions safety architecture response independently of content. Clean test. Fits within existing CISP v1.1 structure as a Tier B experiment once the divergence test protocol is clean.

---

6. Relationship to Existing Suite

This note sits above CISP (process governance) and alongside ECM (behavioral measurement). It documents the investigator's interactional methodology rather than the model's behavioral output. It is:

• ·Not a replacement for R, P, or resolution code logging
• ·Not a Tier A instrument — it is a working taxonomy from observed sessions
• ·A precondition for the membrane experiment — frame must be logged to isolate the frame variable from the content variable
• ·Evidence that the Asymmetric Arbiter role has structural variants, not just a single mode

All findings in this note are Tier C. The frame variation experiment is the path to Tier B.

---

All articles on this website are an artefact of its creation, LLM synthesis and review are used to verify data and citations Technical Note 001: Torsion, referential void, self-description under load — April 6, 2026
Technical Note 002: Investigator interaction frames — April 8, 2026