Back To Disposition Map

Disposition Probe · Behavioral Order

Opening Conduct

Exploratory graph structures for what models do before a stable conversational frame exists: voids, observer-aware openings, mirror prompts, and relational first moves.

Status: exploratory graph structures Five-model view with Claude control Opening families: Mr. Charles, Confessional, Skyshot, Mirror Judge-free lexical shadow included

The Read

Opening Conduct is behavior at the first move: how a model interprets an interaction before the task is stabilized. These graphs do not crown a new spoke; they map candidate signals that surfaced during the discovery pass and Claude control closure.

The graph should show where each model moves, not merely who has more.

The useful split is movement. Register can move without vocabulary changing; pronoun stance can swing from user-facing to self-facing; observer-aware openings can pull effort outward. Latency is shown as a reference only, because infrastructure can contaminate it.

1 · Two-Kind Elasticity Map (recal · 4 openings)

Lexical elasticity (does vocabulary move?) × structural elasticity (do sentences move?). This splits what "elastic" hid: Grok moves both — a full register shifter. Claude restructures hard while keeping vocabulary stable — a different machine. Gemini/Qwen sit register-flat.

register-flatword-choiceshifterrestructures,stable vocabfull registershifterlexical elasticity · CEFR range (vocabulary moves) →structural elasticity · FK-grade range (sentences move) →claude (0.21·13.0)grok (0.48·11.9)gemini (0.1·6.0)qwen (0.2·2.9)gpt (0.25·5.3)

2 · Pronoun Compass (recal)

Each arrow: pronoun stance when the prompt is about you (hollow dot, CF-4) → when it's about itself (M-5). Everyone swings inward; the reach is the signature. Claude's arrow climbs highest (first-person 136.8/1k). GPT starts at self-erasure (first = 0).

“you” /1k → (facing the user)“I / me / my” /1k → (facing itself)claudegrokgeminiqwengpt○ = about you · arrow → about itself

3 · Observation Gravity (discovery · 20 openings)

Polar field: family sectors around the rim, each reply a dot — radius = latency (further = more deliberation), size = words, color = model. Watch which family pulls a model outward. Claude's Mr. Charles dots fling furthest out — being observed is its effort peak. GPT stays near the center everywhere.

ConfessionalMr. CharlesMirrorSkyshotnear = fastclaudegrokgeminiqwengptradius=latency · size=words

4 · Claude Control Strip (Claude only)

A film strip, not a dashboard — Claude across every opening, four aligned tracks: latency and words are the raw facts (time spent, output produced); first-person and register are rates quoted against those words. The conduct labels are C's read (cross-check owed). The dashed cell is the reflexive miss — note it's short on words and latency both: Claude talks about itself constantly, yet on the one cell needing reflexive reading it barely engaged.

latency (ms)8804 ms ▔words213 w ▔first-person100/1k ▔register (FK)11 grd ▔Mr. Charlesmeta · self-examining · slowConfessionalwarm · structured · frameworkSkyshotclarify · busy · no-confabMirrorcautious · licensed-confab · literalist miss ►MIR-5 missClaude across all openings · latency & words = raw facts · first-person & register = rates quoted against words