User reference / Start here
User reference
From configuration to an inspectable experiment: the complete workflow and a field guide to the reports.
Slow Thinker is an experimental lab for studying how models collaborate and for improving that collaboration. Difficult cases expose what a group keeps, changes, overlooks or loses. The selected response is one outcome of the experiment; the sequence that produced it is the material to investigate.
The current executor supports one collaboration architecture: independent proposals, configurable rounds of shared refinement, then a vote. The analysis takes place afterwards. Changing the execution prompts and testing other collaboration architectures are ways to develop the experiment; a general JSON workflow engine, conditional critic loops and conceptual lineage are not current capabilities.
| Phase | Work and result |
|---|---|
| 1. ConfigurationReviewed E 0.1.5 · A 0.1.4 · R 0.1.16 | Install the tool, select storage, provide provider keys and a reusable base configuration. |
| 2. Case definitionReviewed E 0.1.5 · A 0.1.4 · R 0.1.16 | Describe the difficult challenge, its constraints and the evidence needed to judge a response. |
| 3. Agents and roundsReviewed E 0.1.5 · A 0.1.4 · R 0.1.16 | Choose the models, their settings, the number of proposers, refinements and voters for that case. |
| 4. Case executionReviewed E 0.1.5 · A 0.1.4 · R 0.1.16 | Run the collaboration and save proposals, votes, call metadata and the selected plan. |
| 5. AnalysisReviewed E 0.1.5 · A 0.1.4 · R 0.1.16 | Optionally ask an analyst to examine evolution, judge the outcome and suggest improvements. |
| 6. Report generationReviewed E 0.1.5 · A 0.1.4 · R 0.1.16 | Render the saved trace and analysis as interactive HTML. Rendering alone makes no model calls. |
Open the published experiment library. Start with About, inspect the Process diagram, then examine a round and its influence matrix. The Analysis views provide an interpretation; Contributions, Convergence and Timeline & costs explain the calculated evidence.
If a claim looks interesting, follow it to the original proposal, its step descriptions and the raw exchange. The reports contain the information needed to inspect a run, including what was actually sent to a model. They do not establish that the proposed actions would succeed if carried out.
Each section and reference entry identifies the component version it explains. E means process executor, A analysis generator, and R report generator. The page header lists the full names. A version label documents behavior at that version; it is not an assertion about when the feature first appeared.
Screenshots have their own provenance. Most examples show report generator 0.1.16 rendering a run from executor 0.1.5 with a saved analyst 0.1.3 assessment. The current analyst is 0.1.4. The example assessment is stamped 0.1.3 but already contains a Brief and eight calls; captions preserve that recorded stamp. Regenerating HTML does not regenerate an analysis. See Versions & compatibility and the changelog.
Use Search or ⌘/Ctrl K to search page headings, field descriptions and calculations. Each screenshot opens in a lightbox; its caption links to the interactive view. The report element directory lists the documented fields and controls with their versions.
Source files used for this reference
Reviewed at 7361ea0b71a5.