Completed research · Comparative study
A compact companion reduced tokens, with a citation tradeoff
Compact reference packaging used 62.4% fewer input tokens and cost 49.8% less in the main comparison. Some citations became only partly supported.
- Work completed
- Result posted
Method
Twenty-four authored cases in one asset-permission and documentation family were answered twice with each resource: 96 main sessions plus two pilots. A fuller manuscript reference was compared with a compact synthesis and guided worksheet under matched facts, instructions, model settings and a one-response budget.
Findings
- Main input tokens were 850,476 with the fuller reference and 319,452 with the compact resource. Main usage costs were $2.133412 and $1.070784, respectively.
- Separate AI review judged 48 of 48 explanations sound in each arm. Citations were fully supported in 48 of 48 fuller-reference answers and 36 of 48 compact answers; 12 compact answers were partly supported.
- The complete study used 98 paid requests and cost $3.271520 at the recorded historical usage rates.
Limits of the finding
- This is one bounded task family with AI assessment, not a human learning or intended-reader study.
- The resource change bundles selection, synthesis and worksheet organization. Preparation and maintenance costs are excluded.
- These observations establish neither improved reasoning nor demand, broad transfer, book necessity or publication readiness. Manuscript extracts, hidden answer keys and the companion prototype are not released by this result.
Evidence reviewed
Primary records reviewed
- Book Companion API Experiment: frozen protocol, results and retained assessment records, November 9, 2026.
Next question
Do intended readers complete real tasks more successfully using a corrected companion?