How highsierralabs/laconic_tests works
Laconic_tests is an empirical study of whether a short LLM instruction reduces verbosity without losing decision-critical content. The workflow applies the directive to scored prompts across three models and modes, preserves their transcripts, normalizes replies, and compares response length and retained rubric items; results are summarized and visualized. The graph includes the documented human-scoring and adaptation branches. Source sampling confirms transcript parsing and figure generation; scoring relationships follow the README and docs descriptions rather than inspected scoring implementation.
Main entry points: parse_transcripts.py Transcript parser · laconic_v2.md Laconic directive · round2_prompts.md Scored prompts · raw Raw transcripts
Components 15
Experiment design
- Laconic directive
directive/laconic_v2.md - Scored prompts
prompts/round2_prompts.md
Response corpus
- Raw transcripts
data/raw - Transcript parser
scripts/parse_transcripts.py - Normalized replies
data/responses - Word counts
results/word_counts.csv
Scoring and analysis
- Scoring rubric
docs/scoring.md - Item scores
results/scores_items.csv - Response scores
results/scores_response.csv - Study summary
results/summary.md - Study figures
figures
Reuse
- Adaptation guidance
docs/adapting.md
Other components
- Researcher
- LLM models
- Human scorers
Connections 15
- Researcher → Laconic directive: selects
- Researcher → Scored prompts: selects
- Laconic directive → LLM models: instructs
- Scored prompts → LLM models: prompts
- LLM models → Raw transcripts: produce transcripts
- Raw transcripts → Transcript parser: reads
- Transcript parser → Normalized replies: writes replies
- Transcript parser → Word counts: writes counts
- Normalized replies → Human scorers: present for scoring
- Scoring rubric → Human scorers: guides scoring
- Human scorers → Item scores: record item judgments
- Human scorers → Response scores: record response judgments
- Item scores → Study figures: feeds retention plot
- Word counts → Study figures: feeds word plots
- Researcher → Adaptation guidance: consults