Empirical validation agenda for the technical report (tracking)

## What / Where

Tracking issue for the empirical validation the FLoRA **technical report** needs. Many thresholds and coverage assumptions across the pipeline are currently unvalidated (several are self-described as such in code comments and in `CLAUDE.md`). This issue collects the measurements to run.

## Why it matters

The report will make quantitative claims about recall, precision, and outcome balance. Right now those rest on hand-tuned thresholds and untested assumptions. Each item below is a place a number could be materially wrong.

## Agenda

**Stage 1 — recall**
- Run `search/sensitivity_check.py` to estimate keyword recall, **split by success vs. failure phrasing**.
- Test whether OpenAlex `.search` actually honors **quoted phrases** (if not, phrase precision assumptions are wrong).

**Stage 2 — false-rejection rate**
- Hand-label a sample of the never-reviewed **no-phrase bucket** (~2.17M rows) to estimate the false-rejection (recall-loss) rate. (Related: the Stage-2 second-pass issue.)

**Stage 3 — linking precision & thresholds**
- Per-`link_method` precision on a **gold sample**, plus threshold sensitivity sweeps for the values currently hard-coded and self-described as unvalidated:
- citation scoring cutoffs (4.0 / 2.0)
- same-author Jaccard floor (0.05)
- `doi_verify` Jaccards (`VERIFY_TITLE_JACCARD 0.5`, `RESOLVE_TITLE_JACCARD 0.7`, `TITLE_ONLY_JACCARD 0.6`, `TITLE_ONLY_GAP 1.5`)
- `find_all_candidates()` **empty-rate**: `shared/openalex_client.py:284–299` returns `[]` whenever there is no `openalex_id_r` or no indexed references — this bounds how many rows the rule-based linker can even attempt. Measure the empty-rate to bound rule-based linking coverage.

**Stage 3 — outcome**
- Outcome **gold sample** (~150–300 rows), stratified by decision path, including the bare-`"replicated"` → `success` keyword bucket (which is a low-signal shortcut worth auditing).
- **Abstract-only vs. full-text** outcome comparison — how often does full text change the coded outcome vs. abstract alone?

Found in the 2026-07-16 end-to-end pipeline review.

Start with CLAUDE.md and search/sensitivity_check.py, then inspect shared/openalex_client.py:284–299 and the listed threshold assumptions. Run the sensitivity check and design the gold-sample measurements across recall, linking, and outcome stages; done means the agenda has recorded measurements for the stated coverage, precision, threshold, and outcome questions.

Written by the indexing model from the issue text.