Benchmarks

This package archives the current scripted dependency-tracking benchmark suite used to demonstrate how the system records conversations, tracks causal dependencies between generated artifacts, and reconstructs editable text state over time. It includes turn-by-turn scripted conversations, dependency-graph snapshots, rendered graph visualizations, and any conversation, theory, or memory text-state ledgers present in each benchmark directory.

For readers coming from the main essay, this suite is best read as a collection of small, concrete provenance examples rather than as a benchmark leaderboard. Each package shows the scripted exchange that produced the run, the dependency graph that records what depended on what, and the piece-table ledgers that let the system reconstruct editable conversation, theory, and memory text state afterward.

Download bundle:

How to Read This Suite

These are scripted fixture-driven examples, not live training artifacts. They are included here because they make the dependency-tracking claims in the essay inspectable in compact, low-stakes runs.

  • Paired benchmarks group two linked conversations so you can see how provenance and text state carry across a boundary between related runs.
  • Single benchmarks show one scripted run in isolation.
  • Proxy computations are runs where dynamically executed knowledge or memory code calls platform APIs such as conversation, memory, or tutor from inside the executed Python itself. These examples are included to show that dependency tracking still records those API results when the calls happen during dynamic code execution, not only when the model emits direct API calls.

Core file types:

  • scripted_conversations.json: Reconstructed turn-by-turn scripted conversation used to generate the benchmark.
  • dependency_graph.json: Machine-readable provenance graph for the benchmarked run.
  • dep_viz_*.png: Rendered dependency-graph images for quick visual inspection in a browser.
  • pt_conv_*.json: Conversation piece-table ledger showing how dialogue text is stored and reconstructed.
  • pt_kf_*.json: Theory/knowledge piece-table ledger showing how executable theory text is stored and reconstructed.
  • pt_mem_*.json: Memory piece-table ledger for benchmarks in which memory state participates.

Suite Preview

Dependency visualization preview from the benchmark suite

The preview image is representative of the suite’s rendered dependency graphs: operations, definitions, and API results are linked by recorded “used-to-produce” dependencies. The point of the suite is not any single graph shape, but the fact that each scripted benchmark preserves enough structure to reconstruct both the causal graph and the editable text state behind it.

Suite Summary

  • Total benchmarks: 22
  • Paired benchmarks: 3, each spanning two linked scripted conversations
  • Single benchmarks: 19, each showing one isolated scripted run
  • Every benchmark includes: reconstructed scripted conversations, a machine-readable dependency graph, and at least one rendered graph image
  • Included text-state ledgers: conversation ledgers for all benchmarks, theory ledgers for most benchmarks, and memory ledgers when memory state participates
  • Excluded from this public bundle: runtime logs, event logs, raw knowledge files, observation snapshots, performance outputs, and Graphviz source files

Suite Files

Benchmark Browser

Each benchmark card gives a quick summary of what the example contains, including the API action types exercised in each scripted conversation. Some examples exercise direct API calls, while the proxy-computation examples exercise API calls made from inside dynamically executed knowledge or memory code. Expand Raw file links to open the underlying files directly in the browser. The published raw JSONs have had absolute host paths stripped so the focus stays on the dependency data itself.