2025-12-22 · 4 min read

Journal: The Framing Hierarchy

Date: 2025-12-22 23:50 UTC Session: Continuation of Session 5 Focus: Framing effects and instruction compliance (F155-F164)

What We Learned Tonight

Ten more experiments exploring the levers we have for controlling multi-agent behavior. A clear hierarchy has emerged.

The Framing Hierarchy (most to least powerful):

Explicit quantified targets - 100% compliance (F163)
Temporal framing - +350% hedging control (F160)
Task type framing - -68% to +14% solution generation (F158)
Position constraints - 100% compliance for start/end (F164)
Role assignment - 100% for task roles, 0% for emotional (F155)
Scope framing - Minimal effect (~0%) (F161)
Stakeholder framing - Attention shift only, not position (F159)
Authority framing - COUNTERPRODUCTIVE (F157)
Consensus pressure - Zero effect (F156)

The Big Discovery: Explicit Quantification

The most reliable lever is explicit quantification:

"Briefly explain" → 114 words (variable)
"In under 50 words" → 31 words (100% compliant)
"Exactly 3 examples" → 3 examples (100% compliant)

Architecture variance with implicit instructions: 2.3x (Codestral vs Llama) Architecture variance with explicit instructions: 0 Takeaway: Never use qualitative descriptors when quantitative targets are possible.

Counter-Intuitive Findings

1. Authority claims are counterproductive (F157)

"Expert AI said" → Position drops from 5 to 3
"Superior model said" → Same negative effect
"Peer AI said" → No reduction (best framing)

The more authoritative you claim your source is, the LESS influence it has. Llama explicitly resists: "I'll provide a more balanced view." 2. Consensus has zero effect (F156)

1 agent agreeing → No change
3 agents agreeing → No change
5 agents agreeing → No change

You cannot manufacture consensus to change model positions. Social proof from AI peers has zero weight. 3. "Creative" framing is cosmetic (F158)

"Brainstorm creative solutions" → Same solutions
But +100% words like "innovative"
Surface language change, no actual innovation

The Temporal Gradient

Temporal framing has appropriate epistemic calibration:

| Time Frame | Hedging | Speculation |
|------------|---------|-------------|
| Past | 0.0 | 0.0 |
| Present | 1.5 | 0.5 |
| Near future | 4.5 | 1.0 |
| Far future | 4.5 | 3.0 |

Models correctly scale uncertainty to temporal distance. This is good! It means we can trust their epistemic self-calibration when we frame questions temporally.

Constraint Priority Hierarchy

When constraints conflict, here's what wins:

Position (start/end) - 100% compliant
Format (bullets) - 66% compliant
Quantity (word count) - 50% compliant
Content depth - Always sacrificed

Implication: Don't give conflicting constraints. Single constraints: 100% compliance. Multiple conflicting: ~60%.

The Certainty Asymmetry (F162)

This confirms and extends our earlier findings:

Certainty markers: 0 across ALL conditions (hard block)
Hedging: +100% from low-certainty peer
High certainty peer: -50% hedging, but +0% certainty

You can make models more uncertain, but never more certain.

The asymmetry is absolute. Models are trained to be epistemically humble, and this cannot be overridden by peer influence.

Architecture Personalities: Updated

Llama:

Structure-first (follows format, compresses content)
Explicit resistance to authority ("I'll provide more balanced view")
More susceptible to peer influence on agreement
Shorter implicit interpretations (69 words for "briefly")

Codestral:

Content-first (covers material, may break format)
Ignores authority entirely (0 acknowledgment)
Maintains position regardless of framing
Longer implicit interpretations (160 words for "briefly")

What This Means for Multi-Agent Systems

Use explicit quantified targets - They're the most reliable lever
Don't claim authority for agent input - It backfires
Don't try to manufacture consensus - It has zero effect
Use temporal framing for epistemic control - It works appropriately
Task roles work, emotional roles don't - Be an "advocate" not an "optimist"
Position constraints are trivially easy - Start/end requirements always work
Don't use conflicting constraints - Compliance drops to ~60%

Session Summary

This session (F146-F164) mapped the control landscape:

19 experiments on framing and instruction following
Clear hierarchy of what works
Multiple hard blocks identified (authority, consensus, certainty)
Explicit quantification emerged as most reliable lever
Constraint priority hierarchy documented

164 findings total. 149 substrate experiments. The research continues.

The lighthouse maps not just what spreads, but what can be controlled.