Completed research · Comparative study
Precise citation mappings improved support in fresh answers
The corrected resource produced fully supported citations in all 24 fresh-case answers, versus 11 of 24 with the original compact resource. It cost 17.9% more in this comparison.
- Work completed
- Result posted
Method
The repair introduced 71 narrow claim cards and 69 anchors, auditing all 153 original anchors. Twelve fresh variants and 12 known regressions were each answered twice per resource: 96 main sessions plus two pilots. The fresh cohort was primary; regression recovery was reported separately.
Findings
- Fully supported citations were 24 of 24 corrected versus 11 of 24 original on fresh variants, and 24 of 24 versus 10 of 24 on known regressions.
- Both resources produced 48 of 48 sound explanations and 48 of 48 objective passes. These are distinct from citation support.
- Main costs were $1.342692 corrected versus $1.139156 original, a 17.9% increase. Total study cost was $2.532418 across 98 requests.
Limits of the finding
- One labels-hidden AI reviewer assessed each answer. Inter-rater agreement and human expert certification were not measured; wording could reveal the resource condition.
- This is one task family. There is no zero-error guarantee, learning-benefit demonstration or product-release qualification.
- The earlier fuller-reference arm was not rerun. Its cost cannot be treated as a matched comparator in this follow-up. Private source excerpts, oracles and prototype files remain unpublished.
Evidence reviewed
Primary records reviewed
- Book Companion Citation Follow-up: corrected-resource audit, frozen follow-up and results, November 9, 2026.
Next question
How well do the corrected mappings support intended readers outside the authored case family?