Blame for docs/EXPERIMENT_PROTOCOL.md

# Experiment Protocol: Disentangling Models from Personas in LLM Social Simulations

## Overview

This document is the canonical reproducibility narrative for the dyadic, triadic,
and four-way (tetradic) sweep experiments and the lexical/content analyses reported
in the paper. It complements the code in `sweep/`, `oasis/`, `launch/`, and
`analyses/` by explaining the experimental design decisions that are not obvious
from code alone.

---

## Simulator

We adapt the OASIS simulator (v0.2.5, https://github.com/camel-ai/oasis) to
support heterogeneous multi-model populations. The key modifications are in
`oasis/run_v9.py`:

- `hetero` mode: 2 model families, 10 agents each
- `hetero3` mode: 3 model families, 10 agents each
- Plackett-Luce feed ranking with Gumbel noise (T=5.0)
- Post-share balancer: prevents one model family from dominating posts
- Warm-up phase: all agents post once before the main loop; any run where a
  family has < 50% participation is aborted
- SIGTERM-safe checkpoint/resume for preemptible SLURM allocations

---

## Models

Five models are tested. All are served via vLLM 0.19.0 on 8× NVIDIA L40S 48 GB.
See `env/vllm_version.md` for exact launch parameters.

| Model | HF ID | TP | Port |
|-------|-------|----|------|
| Qwen | Qwen/Qwen3-32B | 2 | 8002 |
| gpt-oss | openai/gpt-oss-20b | 1 | 8003 |
| Magistral | mistralai/Magistral-Small-2509 | 2 | 8004 |
| GLM-4 | zai-org/GLM-4-32B-0414 | 2 | 8007 |
| Gemma | google/gemma-4-31B-it | 2 | 8009 |

---

## Persona initialization

Personas are sampled from the Moltbook community archive (HF dataset). The
seed database (`oasis/seed_db/hf-zero-50.sqlite`) contains 50 real agent
profiles: display names, bios, and prior post histories (up to 5 posts per
agent, ~500 chars each).

A diversity filter ensures at most 1 HK CLAW minter persona per run (a narrow
cluster that otherwise over-represents in random samples). The filter is
deterministic given the RNG seed.

Each seed value produces a different persona sample. Seeds: [42, 137, 271, 314, 999]
for dyadic runs; [42, 271, 999] for triadic runs.

---

## Rotating assignments

For **dyadic** runs: 20 agents (10 per family), two orderings per seed (A=X,B=Y
and A=Y,B=X). This gives 2 × C(5,2) × 5 = 100 runs.

For **triadic** runs: 30 agents (10 per family), 3 of 6 orderings per seed
(selected by taking permutations at indices 0, 2, 4 for maximum positional
diversity). This gives 3 × C(5,3) × 3 = 90 runs.

---

## Simulation parameters

| Parameter | Value |
|-----------|-------|
| Steps | 100 |
| Activation probability | 0.3 per agent per step |
| Feed size (K) | 20 |
| Gumbel noise temperature | T=5.0 |
| Post-share tolerance | δ=0.05 |
| Input token limit | 12,000 |
| Max output tokens | 1536 (512 for Gemma) |
| Temperature | 0.85 |
| Max tool-call attempts per turn | 6 |

---

## Metrics

### Directed pairwise engagement homophily (H)

For families X and Y:

```
H_{X→Y} = obs(X→Y) / exp(X→Y) - 1
```

where `obs(X→Y)` = fraction of all directed comment edges from X to Y, and
`exp(X→Y)` = null expectation under random mixing proportional to group sizes.

Two edge types are counted:
1. commenter → post_author
2. replier → parent_comment_author (threaded replies)

Positive H = X engages with Y more than random; negative = avoidance.
Diagonal (X=X) = in-group self-preference.

### Swap stability

For dyadic runs: mean |H_{X→Y}^(AB) - H_{X→Y}^(BA)| across seeds.
For triadic runs: mean max-|ΔH| across the 3 orderings per (triplet, seed).
Lower = more stable across persona-to-model assignment rotations.

---

## Validity criteria

A run is included in the paper analysis if:

- **Dyadic**: ≥ 18 distinct agents commented, ≥ 200 total comments
- **Triadic**: ≥ 26 distinct agents commented, ≥ 200 total comments

These thresholds filter out runs that were aborted (e.g., server failure,
warm-up failure) before completing enough interaction to yield reliable H estimates.

Note: The H-matrix figure scripts (`analyses/plot_matrices.py`) apply only
`nu ≥ 26` for triadic runs (no nc threshold), matching the original data
selection used for the paper's H tables. The variance decomposition scripts
(the variance scripts) apply `nc ≥ 200` for both
dyadic and triadic, which excludes one additional triadic run with nc=169.
The paper's stated 182-run pool (100 dyadic + 82 triadic) was computed at an
earlier snapshot; the current pool is 100 dyadic + 89 triadic = 189 runs (see
`data/MANIFEST.md`).

---

## Run identification convention

Database file names encode the run configuration:

```
sweep_{modelA[:3]}_{modelB[:3]}_s{seed}.db       # dyadic
sweep_tri_{A[:3]}_{B[:3]}_{C[:3]}_s{seed}.db     # triadic
```

Each `.db` file has a corresponding `.metadata.json` sidecar with the
agent-to-model mapping, and `.metadata.json` + `_chats.jsonl` + `_reasoning.jsonl`
side-car files.

---

## Variance decomposition methodology

The variance decomposition (Appendix D) uses **comments per post** (CPP =
in-degree / number of posts) as the dependent variable, restricted to agents with
at least two posts. Raw in-degree conflates per-post appeal with posting *volume*
(an agent that posts more mechanically receives more comments); CPP removes that
exposure confound. Earlier drafts used raw in-degree, under which persona appears
to dominate two-way mixtures purely as a posting-volume artifact.

| Script | What it produces |
|--------|-----------------|
| `analyses/variance_cpp.py` | Hierarchical OLS R² decomposition + leave-one-model-out, on CPP (Tables D.1–D.2) |
| `analyses/absolute_variance.py` | Variance components across dyadic / triadic / four-way regimes |
| `analyses/plot_fig_variance_overview.py` | Figure: proportional CPP variance (model vs persona) + per-model H by mixture size |

Model selection is the agent's own base model plus the unordered partner-pair
(the other models present). In four-way runs all four models are present in every
run, so the partner composition is constant and model selection reduces to
own-model identity.

---

## Reproducibility notes

1. **Stochastic elements**: Two sources of randomness per run, both controlled
   by `--rng-seed`:
   - Persona selection (which agents from the seed DB)
   - Group allocation (which agents go to which model family)
   - Feed noise is per-agent Gumbel, not globally seeded; exact reproduction of
     feed orderings would require seeding each model call, which is not done.

2. **Model-side non-determinism**: vLLM with temperature=0.85 is stochastic.
   Running the same configuration twice will produce different comments. The
   paper reports means ± std across seeds and rotations; individual runs are
   not expected to be identical.

3. **Checkpoint/resume**: Partial DBs can be resumed with `--resume`. The
   `_checkpoint.json` sidecar records the exact last completed step.

4. **Warm-up ordering**: Agents post in a seeded-random order during warm-up,
   ensuring the initial post distribution across model families is
   approximately uniform by construction.

---

## Four-way (tetradic) runs

The four-way sweep uses the pool {gpt-oss, Qwen, Magistral, GLM-4}, 40 agents
(10 per family), 100 steps, post-share balancer on, activation prob 0.3
(`run_v9.py --mode hetero4`). Two seeds (42, 271) each cover 12 distinct
model->block rotations (a Latin-square-style set), for 24 valid runs.

- **Validity:** >= 200 comments and all four models present.
- **Run identification:** `tetradic/tetra_s{seed}_r{rotation}.db`.
- **Metrics:** same directed engagement homophily (H) as dyadic/triadic; for the
  variance decomposition and content analyses, CPP as above.

Driver: `launch/run_tetradic.sh`. Per-model H over the 24 runs and the four-way
H matrix/heatmap: `analyses/tetradic_recompute_H.py`, `analyses/tetradic_results.py`.

---

## Content analysis

Finding 3 analyzes the text agents produce. Build the corpus once with
`analyses/extract_texts.py` (-> `data/texts.jsonl`: one row per post/comment with
its model and persona). Then:

- `analyses/lexical_table.py` — TF-IDF (word 1–2-grams, English stop words removed,
  + char 3–5-grams) -> logistic regression recovering model / persona (Table 1).
- `analyses/svd_axes.py` — agent-level TF-IDF -> truncated SVD (30 components); the
  first axis (document language) is dropped, leaving 29 style axes. Reports per-axis
  top terms from V^T.
- `analyses/content_engagement.py` — decomposes CPP variance into length / style /
  model / persona blocks (single + leave-one-block-out unique R²).
- `analyses/style_geometry.py` — per-model style-direction centroids, their split-half
  stability, and cosines to the engagement direction (linear probe on style).
- `analyses/cluster_model_persona.py`, `analyses/length_regimes.py` — supporting
  checks (does style cluster by model vs persona; length/volume confound by regime).