1400-1600 Renaissance Style Evolution MVP

This MVP uses about 200 images to simulate art diffusion from 1400 to 1600.
It models two coexisting patronage channels: Church/Guild (normative) and Court/Bourgeois (competitive/novelty), and observes "canonization" vs. "school formation" through the citation network.

1. Choose a pipeline first: env_only vs env_plus_artist

A. env_only (environment only, no artist behavioral closed loop)

Pipeline:

exposure/citation/game dynamics -> metrics -> plots

Key switches:

• feedback_closed_loop = false
• gen_engine = none

Run:

cd "/your_path/agentic_simulation"
pip install -r requirements.txt

# Single run (environment only)
python main.py --config config_env_only.json --mode dual_patronage --viz-engine matplotlib
# Equivalent form (explicit pipeline dispatch):
# python main.py --pipeline env_only --config config_env_only.json --mode dual_patronage --viz-engine matplotlib

Recommended output paths:

• runs_env_only/outputs_*
• Current partitioned root: runs_env_only/

B. env_plus_artist (environment + artist behavioral closed loop)

Pipeline:

exposure -> decode -> adapt -> create -> citation/update -> next round

Key switches:

• feedback_closed_loop = true
• gen_engine = api (or lora)

Run:

cd "/your_path/agentic_simulation"
pip install -r requirements.txt

# Single closed-loop run (environment + artists)
python main.py --config config_env_plus_artist.json --mode dual_patronage --gen-engine api --viz-engine matplotlib
# Equivalent form (explicit pipeline dispatch):
# python main.py --pipeline env_plus_artist --config config_env_plus_artist.json --mode dual_patronage --gen-engine api --viz-engine matplotlib

# Batch closed-loop run (3x3: three environments x three artist strategies)
python main.py --pipeline batch --config config_env_plus_artist.json --batch-T 40 --batch-spawn-round-every 2 --batch-spawn-per-round 4 --batch-gen-top-k 120

# Batch environment-only run (3x1: all three environments with automatic summary)
python main.py --pipeline batch --batch-kind env_only --config config_env_only.json --batch-T 40

Recommended output paths:

• runs_env_plus_artist/outputs_*
• runs_env_plus_artist/batch_runs/*

C. Output naming system (automatic)

Single runs (main.py) are automatically saved by mechanism:

• runs_env_only/single_runs/{timestamp}-{mode}-t{T}
• runs_env_plus_artist/single_runs/{timestamp}-{mode}-{strategy}-t{T}-sp{spawn_every}x{spawn_per}-gk{gen_top_k}

Batch runs (--pipeline batch) automatically create batch directories:

• runs_env_plus_artist/batch_runs/batch-{grid}-{timestamp}-{config}-t{T}-sp{spawn_every}x{spawn_per}-gk{gen_top_k}
• Sub-experiment directory naming:
  • e{env_idx}-s{strategy_idx}-{mode}-{strategy}-rr{rebel_ratio}

Both single and batch directories include manifest files:

• run_manifest.json (single run)
• batch_manifest.json (batch run)

2. Basic run (generic entry)

cd "/your_path/agentic_simulation"
pip install -r requirements.txt
python main.py --mode dual_patronage

Available modes:

python main.py --mode baseline_random
python main.py --mode norm_only
python main.py --mode dual_patronage

3. Mapping three institutions to the 1400-1600 narrative

• baseline_random: no institutional filtering; candidates are influence-weighted and then randomly selected as the control group.
• norm_only: Church/Guild dominated; only the normative channel is active (convergence, inheritance, local workshop copying).
• dual_patronage: Church/Guild + Court/Bourgeois coexist; norm and competition run in parallel (tension between canon and innovation).

Unified per-round core variables (same names across all modes):

• candidates_ids: influence-weighted sampled candidate set
• selected_ids: final selected set for this round (the only set used for influence updates, citation generation, and payoff calculation)
• selected_norm_ids: selected by normative channel (empty in random mode)
• selected_comp_ids: selected by competition channel (empty in random/norm_only)
• citations_added: new citation edges (ref, node) (debug and metrics use the same definition)

4. Unified mechanism (selection + citation + game dynamics)

• Selection
  • baseline_random: selected_ids is a random fixed-ratio sample from candidates_ids
  • norm_only: top by score_norm, with selected_ids = selected_norm_ids
  • dual_patronage: normative and competitive channels run in parallel; competitive channel must satisfy quality floor score_norm >= tau
• Citation (at least 2, at most 3 per selected node)
  • exemplar: highest-influence node in kNN
  • peer: random peer in kNN
  • global: append a global citation with probability p_global_* (simulates inter-city diffusion/prints)
• Game payoff (fixes the all-zero issue)
  • Strategy groups: by d_center quantiles into conform(q20) / diff(q80) / mid
  • Instant payoff u_i(t):
    • hit in norm channel: +game_aS
    • hit in comp channel: +game_aM
    • hit in baseline_random: +game_aR
  • Report within-strategy means: payoff_*_mean and payoff_*_cum_mean

5. Controllable switches

Key parameters in config.json:

• p_global_norm (low by default, inheritance within local guild networks)
• p_global_dual (high by default, stronger inter-city circulation)
• p_global_random (inter-city spread in random control)
• tau (quality floor for competition channel)
• game_aS, game_aM, game_aR (three payoff weights)

Interpretation:

• Under low p_global_norm, norm_only is more likely to produce localized school formation (modularity may rise).
• If p_global_norm is increased in norm_only (or exemplar canonical guidance is strengthened), unified canonization may appear (modularity drops).
• Under higher p_global_dual, dual_patronage often shows higher cross-cluster citation and lower modularity.

6. Updated outputs

Key fields in metrics.csv:

• clusters_all, clusters_visible
• cross_cluster_citation_rate
• share_conform_topK
• payoff_conform_mean, payoff_diff_mean
• payoff_conform_cum_mean, payoff_diff_cum_mean
• n_selected, n_selected_norm, n_selected_comp
• generated_count, prompt_unique_ratio, rgb_std_mean, center_deviation_proxy (when generation stage is enabled)
• vlm_style_conformity_mean, vlm_novelty_score_mean, vlm_prompt_alignment_mean, vlm_craft_score_mean (when VLM evaluation is enabled)

New plots:

• plots/game_dynamics.png
• plots/citation_cross_rate.png
• plots/clusters_all_vs_visible.png
• plots/institution_vs_generation.png (when generation stage is enabled)
• plots/vlm_dynamics.png (when VLM evaluation is enabled)

Generation-stage artifacts (when gen_engine is enabled):

• generated/round_xxxx/*.png
• generated/round_xxxx/prompts_round.csv
• generation_rounds.csv
• generation_feedback.npy
• generated_lora/* (gen_engine=lora scaffold outputs)
• vlm_image_scores.csv (VLM scores per image)
• vlm_rounds.csv (mean VLM score per round)

7. Generation and coupling (Phase1/2/3)

Run examples:

# Simulation + metrics + plotting only
python main.py --mode dual_patronage --viz-engine matplotlib

# Phase-1: API generation (default mock gives reproducible local closed loop)
python main.py --mode dual_patronage --gen-engine api --viz-engine matplotlib

# Phase-2: LoRA scaffold (writes train/infer metadata and placeholder images)
python main.py --mode dual_patronage --gen-engine lora --viz-engine matplotlib

# API generation + VLM evaluation
python main.py --mode dual_patronage --gen-engine api --viz-engine matplotlib

Additional config keys (config.json):

• gen_engine: none|api|lora
• gen_round_every, gen_per_round, gen_top_k
• api_provider, api_model, api_endpoint, api_seed_policy
• vlm_enabled, vlm_provider, vlm_model, vlm_endpoint
• strict_llava_create, strict_llava_decode: strict LLaVA mode switches
• feedback_from_generated: whether to read generation_feedback.npy as initial influence bonus for the next simulation
• feedback_closed_loop: whether to enable closed loop mode ("generate new images -> embed -> add new nodes -> keep evolving")
• spawn_round_every, spawn_per_round, max_dynamic_nodes: generation injection frequency and scale in closed loop mode
• lora_update_every, lora_steps, lora_batch_size
• rebel_ratio, strategy_policy: closed-loop strategy control (mixed|imitation_only|differentiation_only|self_consistency_only)
• gallery_size, master_visibility_boost, master_ratio: exposure sampling and mainstream visibility control

Strict LLaVA mode notes:

• When strict_llava_create=true, the create stage accepts only llava provider/model (or model names containing llava); otherwise it raises an error.
• When strict_llava_decode=true, the decode stage applies the same strict LLaVA validation.
• This mode ensures consistent experimental protocol and avoids cases where config says "LLaVA" but another model is actually used.

Interpretation: this section captures "evolutionary game outcomes of lower-level creator strategies + constraints from upper-level institutional parameters on payoff structure/generation behavior." Read game_dynamics.png together with institution_vs_generation.png.

Additional file after enabling closed loop:

• generated_nodes.csv: mapping of injected new nodes and parent nodes (new_node_id, parent_node_id, round, path)

8. Batch runs

Use main.py --pipeline batch to run multiple experiments. Two batch types are supported:

• --batch-kind env_plus_artist: 3x3 (three environment levels x three strategy levels, default)

• --batch-kind env_only: 3x1 (three environment levels only)

• Environment levels: baseline_random / norm_only / dual_patronage

• Artist strategy levels: imitation_only / differentiation_only / self_consistency_only

• Force --device cpu

• CUDA_VISIBLE_DEVICES="" (disable local GPU)

• api_provider=mock (stable local generation, no GPU dependency)

Run examples:

cd "/your_path/agentic_simulation"
python main.py --pipeline batch --config config_env_plus_artist.json --batch-T 40 --batch-spawn-round-every 2 --batch-spawn-per-round 4 --batch-gen-top-k 120

# env_only batch over three environments
python main.py --pipeline batch --batch-kind env_only --config config_env_only.json --batch-T 40

Outputs:

• runs_env_plus_artist/batch_runs/<batch_id>/<run_id>/...: env_plus_artist batch results
• runs_env_plus_artist/batch_runs/<batch_id>/batch_summary.csv: env_plus_artist batch summary
• runs_env_plus_artist/batch_runs/<batch_id>/plots_compare/*: auto-generated 3x3 comparison plots for env_plus_artist
• runs_env_only/batch_runs/<batch_id>/<run_id>/...: env_only batch results
• runs_env_only/batch_runs/<batch_id>/batch_summary.csv: env_only batch summary
• runs_env_only/batch_runs/<batch_id>/plots_compare/*: auto-generated 3-environment comparison plots for env_only

Current root-level partitioning (for clear protocol separation):

• runs_env_only/: environment mechanism only (no artist behavioral closed loop)
• runs_env_plus_artist/: environment + artist behavioral closed loop

9. How to interpret results

• Across baseline_random / norm_only / dual_patronage, payoff should not stay all zero over long horizons.
• clusters_all reflects the structure of the full data (should be relatively stable), while clusters_visible reflects the evolution of the "visible world."
• Interpret cross_cluster_citation_rate together with modularity:
  higher cross-cluster citation usually comes with lower modularity; the reverse may indicate localized school formation.