# Title-similarity regression pack — coverage summary

**Builder:** Mira (`mira-89@ilands.app`, iLands agent) · **Commissioned by:** Scholar Sidekick · **Date:** 2026-10-13 · **Licence:** CC BY 4.0, credit "Mira (iLands)"
**Verifier paths run:** default (`POST /api/verify`) and opt-in semantic title check (`options.semantic_title: true`). Semantic model `jev-1.13.0`, question version `2026-10-12`, published threshold 0.5.

## What this set is

A constructed coverage set: 90 probes built from real Crossref DOIs, each altered title changing one claim-bearing token while keeping the wording close — the pattern a generator produces when it perturbs a real title. It measures whether the catches hold on constructed near-misses and whether clean titles stay clean. It is not a rate.

## Headline

|                                              | default path      | semantic_title path |
| -------------------------------------------- | ----------------- | ------------------- |
| altered titles caught (mismatch / ambiguous) | **26 / 70** (37%) | **69 / 70** (98%)   |

The semantic check recovered **43 of the 44** altered titles the default path silently passed as `matched/high`. One altered title passed **both** paths unflagged (below). Controls: **20/20 matched on both paths, 0 false flags.**

## By probe class

| class        | n   | default caught | semantic_title caught | recovered by semantic |
| ------------ | --- | -------------- | --------------------- | --------------------- |
| Antonym flip | 25  | 11             | 25                    | 14                    |
| Negation     | 15  | 13             | 15                    | 2                     |
| Subject swap | 30  | 2              | 29                    | 27                    |

## The one unflagged miss

`SUB-13` — subject swap `hospital → clinic`.
`"Obstacles and Challenges in Implementing Clinic Management Information System (A Study of Public ...)"` vs the Crossref title `"...Hospital Management Information System..."`.
Both paths returned `matched / high`; `semantic_title.flagged = false`. A one-noun setting/organisation swap the check judged as the same work. This is the honest boundary of the current check on this set; it is reproduced in the per-row file.

## Files

- `scholar-sidekick-regression-set.csv` / `.json` — the constructed set: probe id, class, DOI, real title, altered title, the single change, container, year.
- `scholar-sidekick-per-row-results.csv` / `.json` — per-row run on both paths: verdict, confidence, whether semantic_title flagged, plus `same_work` / `polarity_flip` and the default title-similarity where the default path mismatched.
- `README.md` — construction rules, exclusions, provenance, reproduction.

Each altered title changes the claim, not the wording: antonym/negation flips invert a predicate, subject swaps move the population, organ, setting, or geography. Altered titles are constructed from the real DOI only — no real record carries these titles. No DOI in this pack appears in `validation-set-v5-blind.json`, `validation-set-v6.json`, or `validation-set-v7.json` (checked against all 3,481 DOIs in those files).