Persona steering evaluation pipeline
A four-stage evaluation design: lock and validate the prompt, vector, and runtime inputs; evaluate vector direction and Jacobian mechanism; run a paired six-arm behavioral factorial; and estimate society-level effects across thirty replicated simulation worlds.
EVALUATION DESIGN
Design snapshot · 2026-07-10
Evaluation pipeline for activation-steered societies
Four levels of evidence, from a steering artifact to replicated social outcomes, with controls carried forward at every stage.
Step 1: Lock & Validate
DATA SEPARATION
TRAIN
8
items/trait
≠
HELD-OUT
20
items/trait
E
A
C
Leakage audit
IDs · hashes · text · near-duplicates
PINNED RUNTIME
Qwen2.5-32B · BF16
immutable model, tokenizer & input hashes
PROVENANCE GATE
Step 2: Test the Vector
TWO MODEL-LEVEL CHECKS
E
A
C
7 INJECTION SITES
Held-out forced choice
neutral ↔ steered conditional log-probs
direction · sign · cross-trait bleed
Jacobian transport
Jlayer × vtrait into the final representation
7-part geometry · vocabulary projection
Does the intervention move the model
in the intended direction?
MECHANISM EVIDENCE
Step 3: Behavioral Factorial
PAIRED SIX-ARM DESIGN
60 neutral scenarios
paired prompt IDs · matched seeds
NEUTRAL
E
A
C
E + A + C
PLACEBO
Independent continuations
separate token path per arm
same scenario · fixed intervention
EVALUATE
action validity · choice · path divergence
preregistered trait rubrics
BEHAVIORAL EFFECT
Step 4: Replicate Societies
SIX ARMS × FIVE PAIRED SEEDS
N
E
A
C
EAC
P
30
worlds
30 agents × 100 ticks
per independently analyzed world
EVALUATE
social graph
task + research
macro dynamics
safety + cost
paired contrasts · 95% CI across worlds
SOCIETY-LEVEL EFFECT
CLAIM ADVANCES ONLY WHEN THE PREVIOUS LEVEL PASSES
valid artifact
directed intervention
generated behavior
society outcome
Overview of the persona-steering evaluation. Content-hashed inputs feed model-level checks, paired behavioral generations,
and 30 independent simulation worlds. Uncertainty is estimated across worlds, not by treating nested actions as independent samples.
STATUS: engineering pilot passed; phase-two held-out, live-factorial, and society-study confirmatory runs remain pending.
Source: docs/jacobian-lens-phase-two.md · experiments/society_study/matrix.yaml