| # File and field guide |
|
|
| The layout follows the original SwarmWorld dataset; its own guide applies to the trace record types. Differences are listed here. |
|
|
| ## Study folders (`studies/<run>/<study>/`) |
|
|
| | File | Meaning | |
| |---|---| |
| | `study-manifest.json` | The engine's protocol record, fixed before the run: conditions, populations, seeds, ticks, schedules, engine revision, capabilities, prompt/schema/config hashes. For the parallel runs it also lists the per-cell parts it was merged from. | |
| | `study-summary.json` | One record per episode with endpoints, held-out assays, mobility, technology ecology, model usage and the trace path (rewritten to be release-relative). | |
| | `run-meta.json` | Our run manifest: package version, served model name, server version, base URL, transport, sampling, hashes, start and finish times. | |
| | `config/*.yaml` | The rendered run configuration the engine was given (named by the hash of its content). | |
| | `*.jsonl.gz` | Authoritative event traces; `-member-NNN` files are the isolated independent-search replicates. | |
|
|
| ## Metadata |
|
|
| | File | Meaning | |
| |---|---| |
| | `metadata/studies.csv` | One row per study: model, design, counts, model usage, hashes. | |
| | `metadata/episodes.csv` | One row per episode: principal endpoints (discovery-frontier AUC, held-out resilience AUC, best artifact performance, portfolio resilience, artifacts, collaborative fraction, distance travelled), research score and milestone count, calls, errors, tokens, trace path, state digest. | |
| | `metadata/files.csv` | Size and SHA-256 of every payload file; `scripts/verify_dataset.py` checks them. | |
| | `metadata/run_commands.md` | The commands that produced each study. | |
|
|
| ## Analyses (`analyses/`) |
|
|
| | Path | Meaning | |
| |---|---| |
| | `qwen38-27b-gptq4-n50/study1-analysis/`, `study2-analysis/` | `analyze-study` outputs for the two N = 50 studies (endpoint tables, scaling and contrast CSVs, overview, portfolio and technology-ecology figures). | |
| | `qwen38-27b-gptq4-n50/mobility-4cond/` | `analyze-mobility` over both studies: movement regimes, spatial self-organisation, signatures (CSV, JSON, figures). | |
| | `qwen38-27b-gptq4-n50/trace-report-3203.json`, `trace-figures-3203/` | `trace-report` on the seed 3203 full-culture society (35 artifacts): information flow, coordination, lineage figure. | |
| | `qwen38-27b-gptq4-n50/dossiers-3203/` | `technology-dossiers`: the five best technologies of that society with briefing sheets. | |
| | `qwen38-27b-gptq4-n50/compare-n50-both.csv` | Every N = 50 episode against the original dataset's `episodes.csv`, by condition, seed and metric. | |
| | `model-registry.md`, `model-registry.csv` | Per served alias: checkpoint, quantisation, server, kernels, context, KV cache, flags, measured throughput. | |
|
|
| ## Research score and milestones |
|
|
| The engine tracks sixteen research milestones per episode; `research_score` is the fraction reached and `milestones` the count. The validity pilots are compared on this score because artifacts are rare at N = 4 and 16. |
|
|
