Show the quote
Every value carries the exact sentence it came from, with its location in the paper.
02 · The Instrument
An evidence base you can defend.
Yajur v1 is the working core of a systematic review: search a curated corpus, screen with the scholar in the loop, and extract every included paper into a table where each cell can show its evidence, then draft the review from exactly that approved evidence. The scholar stays the author; every line can be defended.
2.1 · Search
Search the canon, not the web.
Yajur searches 148 top-tier journals (the UTD-24, the FT50, the AJG top tier, plus whatever the scholar adds) and nothing else. It works by idea, not keyword. Filters follow the review: journal, year, method, theory. The citation graph runs backward and forward for snowballing. And every result explains why it belongs.
Does algorithmic advice change what consumers buy?
journals: all 1482015 – 2026method: anysnowball: on
-
Consumer reliance on algorithmic advice in retail choice matched “algorithmic advice” · construct: advice uptakeinclude ✓
-
Machine recommendations and preference formation matched “recommendation acceptance” · field experimentqueued
-
Delegating the cart: field evidence on automated shopping matched “delegation to algorithms” · snowballed from № 1queued
Every result explains why it matched. Nothing outside the corpus can appear.
2.2 · Method
A method, not a chat.
Protocol first. Then a reproducible search, human-confirmed screening, and a PRISMA flow that builds itself as you go. The method log exports straight into a paper, because the point isn’t speed alone, it’s a process you can defend in peer review.
0
records identified: corpus v2026.06, query № 1
duplicates removed: 182
0
after de-duplication
excluded at title / abstract: 888
0
screened in: human confirmed
excluded (wrong method / population): 153
0
full text assessed
excluded (quality / no full text): 26
0
included: every exclusion logged, with reasons
2.3 · Extraction
Forty papers. One honest table.
Every included paper is read in full (start to finish, not skimmed) and reduced to a structured card: what it studies, how, on what sample, what it found. Cards line up as a comparison table. Add your own column and extraction re-runs across the set. And every value carries the exact sentence it came from.
| paper | theory | method | sample | core finding |
|---|---|---|---|---|
| Kim & Osei ’23J. Consumer Res. | Trust transfer | Lab experiment ×3 | n = 1,842 | Advice uptake rises with disclosed accuracy, and collapses after one visible error. |
| Braun ’21Marketing Sci. | Bayesian updating | Field experiment | 214k sessions | Recommendations lift basket size 6–9% in utilitarian categories; null in hedonic.† |
| Rao & Feld ’19Management Sci. | none (analytical) | Analytical model | n/a | Optimal human-override threshold falls as advice precision becomes observable. |
| Ito ’20Info. Systems Res. | Agency theory | Archival panel | 3,120 firms · 9 yrs | Delegation to algorithms rises with monitoring cost, and reverses under audit. |
Blanks stay honest: “not reported” is a real answer, and interpretation is labeled, never asserted.
Trust, kept cheap
Check it, don’t trust it
Every quote is checked back against the source automatically. No match, no cell. The model never gets the last word.
Label the blanks
“Not reported” is a real answer. A labeled blank beats a confident guess, every time.
The scholar signs off
Screening and corrections stay human. Every edit becomes a label that improves the system.