All 38 chats behind this experiment: 32 design-task sessions (2 models × 2 tasks × 8 conditions/stages) and 6 exploratory feature-extraction sessions from an earlier analysis instrument that is not used in the paper. Every chat is available both as a readable, message-by-message transcript and as the raw JSON log.
Branch relationships. Several chats share a conversation prefix and diverge at a fork point
— e.g. within each model/task pair, library_prose is the shared prefix for C6,
C6+, and C7 (all three continue the same prose design differently), and
ATWL+library is the shared prefix for ATWL_2-stage (C5). This is noted per condition
below by the “→ continues” column on the condition
protocols page; the raw chat archive index has the full
turn-by-turn fork analysis.
Design-task sessions (32)
| Condition | Length | Open | Source file |
|---|---|---|---|
| Claude Sonnet 5 · VAST | |||
| C1 C1 | 2 turns | Claude_Sonnet_5_VAST_null.json | |
| C2 C2 | 4 turns | Claude_Sonnet_5_VAST_ATWL.json | |
| C3 C3 | 6 turns | Claude_Sonnet_5_VAST_ATWL+library.json | |
| C5 C5 | 8 turns | Claude_Sonnet_5_VAST_ATWL_2-stage.json | |
| C4 C4 | 6 turns | Claude_Sonnet_5_VAST_library_prose.json | |
| C6 C6 | 8 turns | Claude_Sonnet_5_VAST_C6.json | |
| C6+ C6+ | 12 turns | Claude_Sonnet_5_VAST_C6+.json | |
| C7 C7 | 10 turns | Claude_Sonnet_5_VAST_C7.json | |
| Claude Sonnet 5 · VIS | |||
| C1 C1 | 2 turns | Claude_Sonnet_5_VIS_null.json | |
| C2 C2 | 4 turns | Claude_Sonnet_5_VIS_ATWL.json | |
| C3 C3 | 6 turns | Claude_Sonnet_5_VIS_ATWL+library.json | |
| C5 C5 | 8 turns | Claude_Sonnet_5_VIS_ATWL_2-stage.json | |
| C4 C4 | 8 turns | Claude_Sonnet_5_VIS_library_prose.json | |
| C6 C6 | 10 turns | Claude_Sonnet_5_VIS_C6.json | |
| C6+ C6+ | 14 turns | Claude_Sonnet_5_VIS_C6+.json | |
| C7 C7 | 10 turns | Claude_Sonnet_5_VIS_C7.json | |
| GPT-5.6-Terra · VAST | |||
| C1 C1 | 2 turns | GPT_5_6_Terra_VAST_null.json | |
| C2 C2 | 4 turns | GPT_5_6_Terra_VAST_ATWL.json | |
| C3 C3 | 4 turns | GPT_5_6_Terra_VAST_ATWL+library.json | |
| C5 C5 | 8 turns | GPT_5_6_Terra_VAST_ATWL_2-stage.json | |
| C4 C4 | 6 turns | GPT_5_6_Terra_VAST_library_prose.json | |
| C6 C6 | 10 turns | GPT_5_6_Terra_VAST_C6.json | |
| C6+ C6+ | 12 turns | GPT_5_6_Terra_VAST_C6+.json | |
| C7 C7 | 8 turns | GPT_5_6_Terra_VAST_C7.json | |
| GPT-5.6-Terra · VIS | |||
| C1 C1 | 2 turns | GPT_5_6_Terra_VIS_null.json | |
| C2 C2 | 4 turns | GPT_5_6_Terra_VIS_ATWL.json | |
| C3 C3 | 4 turns | GPT_5_6_Terra_VIS_ATWL+library.json | |
| C5 C5 | 8 turns | GPT_5_6_Terra_VIS_ATWL_2-stage.json | |
| C4 C4 | 6 turns | GPT_5_6_Terra_VIS_library_prose.json | |
| C6 C6 | 10 turns | GPT_5_6_Terra_VIS_C6.json | |
| C6+ C6+ | 14 turns | GPT_5_6_Terra_VIS_C6+.json | |
| C7 C7 | 10 turns | GPT_5_6_Terra_VIS_C7.json | |
Exploratory feature-extraction sessions (6; not used in the paper)
These sessions belong to an earlier analysis in which an LLM coded a feature vocabulary on the library and on the designs. The coding relied on the LLM's interpretation and was not formally defined, so it was replaced by the features computed from the ATWL text and by the coding protocol under Analysis. The sessions are kept for completeness.
| Session | Length | Open | Source file |
|---|---|---|---|
| Vocabulary definition | 36 turns | motifs_GPT_5_Terra.json | |
| Scoring fork (short) | 54 turns | feature_extractor_GPT_5_6_Terra.json | |
| Scoring fork (C4 focus) | 54 turns | feature_extraction_GPT_5_6_C4.json | |
| Scoring fork (C6 focus) | 54 turns | feature_extraction_GPT_5_6_Terra_C6.json | |
| Scoring fork (extended) | 80 turns | feature_extractor_GPT_5_6_Terra-prose.json | |
| Scoring fork (most complete) | 108 turns | feature_extractor_GPT_5_6_Terra-ATWL.json |
Full archive index & downloads
Full archive index (Markdown)
Per-chat topic/process/artifacts/final-state summaries and the complete fork/template analysis.
Archive index (JSON)
Same index, machine-readable.