How Arbiter weighs evidence.

Arbiter's target venue scope is journals appearing on at least one of four authoritative ranking lists: the FT50 (Financial Times), UTD24 (University of Texas at Dallas), ABDC (Australian Business Deans Council), and AJG (Association of Business Schools Academic Journal Guide). Approximately 3,230 distinct journals appear after ISSN de-duplication. Current indexed works and abstract coverage are reported live in section X because some list segments are still being ingested. Year-stamped editions for reproducibility: ABDC 2025, FT50 2016, UTD24 2024, AJG 2024.

SSRN preprints from authors affiliated with AACSB-accredited institutions. Working papers are important in finance and economics; they are deferred to v1.1 to allow rigorous quality filtering before inclusion. The scope decision is documented in the product roadmap.

Papers cited by five or more distinct T1 works. This tier is designed to capture foundational methodological and theoretical works published outside ranked business journals — for example, statistical methods papers in psychology or foundational economics works in non-business venues. T3 is not in the v1.0 corpus; it is planned for v1.1.

Conference proceedings, book chapters, dissertations, technical reports, and works from venues on the maintained predatory publisher exclusion list. Retracted works are hidden by default (see §vii). Exclusions can be reviewed in the Refine panel; retracted works can be surfaced for fraud-research use cases.

Arbiter runs every query against a stored full-text search column (PostgreSQL GIN index on titles, abstracts, and author metadata). The index is pre-computed at ingest time, so queries return in under one second rather than scanning each record at query time. Searches that match on abstract content rank higher than title-only matches because the abstract signals topical depth.

Papers must reach a minimum relevance score (ts_rank ≥ 0.05) to appear in results. Papers below this threshold matched only a peripheral term — a word in a passing reference, not the paper’s actual focus. The gate is applied in the database before scoring, so the result set contains only papers where the query topic is genuinely present.

The same query always returns the same paper set. Arbiter enforces this with a deterministic SQL ordering (rank descending, then paper ID ascending) combined with a composite score tiebreaker in application code. This makes search results reproducible — a researcher running the same query in three years will retrieve the same papers, assuming the corpus has not been updated. What varies: synthesis text is AI-generated and may differ in phrasing across runs; the factual claims are grounded in the same fixed paper set.

Arbiter returns up to 1,000 papers that pass the relevance gate. All of these appear in the evidence panel and receive stance classification. The synthesis model receives only the top 20 papers by composite evidence score. This keeps synthesis focused on the strongest evidence rather than diluting it with hundreds of peripherally-relevant results.

How directly the paper’s title and abstract address the specific query. Derived from ts_rank (PostgreSQL full-text relevance), normalized within the result set so that the most relevant paper in each search scores 1.0. This is the strongest signal in Balanced mode (ε = 0.40) because it answers the question “is this paper actually about what I asked?”

OpenAlex FWCI, normalized with a documented fallback ladder: FWCI → percentile → log-raw-count → 0.5 baseline. FWCI compares a paper’s citation count to other papers from the same field, year, and work type — an empirical study in management is compared to other empirical management studies, not to physics papers. Values above 10.0 are winsorized. Papers fewer than 18 months from publication receive a benefit-of-doubt floor of 0.5.

A time-decay and persistence signal. New papers (under 36 months) receive a benefit-of-doubt floor of 0.50 plus a boost for early citation velocity. Older papers are rewarded for having maintained citation velocity relative to their lifetime average — M captures lasting relevance, not raw age. A 2008 paper still being heavily cited in 2026 scores higher on M than a 2008 paper that peaked in 2012.

OpenAlex source 2yr_mean_citedness, normalized across the corpus. This is the least-weighted signal (δ = 0.05) and exists primarily to break ties between papers with otherwise similar R, F, and M scores. It accounts for the documented phenomenon that papers in higher-prestige venues within a tier receive more citations, controlling for paper quality.

J is computed from the paper’s highest tier across all ranking lists (FT50/UTD24 → 1.00; ABDC A* → 0.80; ABDC A → 0.65; ABDC B → 0.45; ABDC C → 0.25). It appears on the paper detail page so researchers can see which list a journal belongs to. J is deliberately excluded from the composite score: every paper in the corpus already passed a journal-tier filter at ingest, so using J in scoring would double-count prestige and cause high-tier journal papers with marginal relevance to outrank directly-relevant lower-tier papers.

Arbiter generates synthesis at three reading levels selectable on the answer page. Essentials uses everyday language — no research jargon, just the core finding and what it means. Detailed introduces research terminology with brief explanations and discusses how studies relate to each other. Technical uses discipline-standard vocabulary, methodological detail, and direct engagement with conflicting evidence — suitable for inclusion in a dissertation literature review.

Synthesis is generated by OpenAI’s gpt-5.4-nano model via Vercel AI Gateway. Output is structured using the Vercel AI SDK’s streamObject function with a Zod schema: { headline, groups[{ stance, narrative, papers[] }] }. The headline is one sentence summarizing the key finding and its main caveat. Groups are narrative paragraphs organized by stance — papers that support the query answer, papers that challenge it, and papers that show it depends on conditions. Each group cites specific papers inline.

The synthesis prompt receives a work-type label with each paper: empirical study, literature review, or preprint. The model is instructed to weight these differently: meta-analyses and systematic reviews are highest weight (mention first, most authoritative for existence claims); empirical studies are weighted proportionally to replication count and sample scope; theoretical papers explain mechanisms and boundary conditions but are not treated as evidence that something happens. This prevents the synthesis from overstating single-study findings.

The synthesis model sees only titles, abstracts, authors, years, and work-type labels — not full paper text. It does not access the internet, follow citations, or reason about papers it was not given. The synthesis is a structured summary of the abstracts of the top 20 papers by composite evidence score. It is a starting point for literature engagement, not a substitute for reading primary sources.

Every paper is classified into one of three stances relative to the search query. “Supports”: the paper’s findings are consistent with or favor the claim. “Challenges”: the findings contradict or question the claim. “Depends on…”: the relationship is contingent on a moderating variable, context, or condition. Display labels are "Supports" / "Challenges" / "Depends on…"

Stance is classified by a separate model (OpenAI gpt-4.1-nano) in a dedicated request pipeline — it is not done by the synthesis model. Papers are batched 50 per request using generateObject with a Zod schema for structured output. The classifier receives each paper’s title, abstract, and the original search query. Papers without usable abstracts (missing, too short, or Cloudflare-blocked) may not receive a classification; this is why the stance counts sometimes show “N of M classified” when they differ.

Business research findings almost always come with conditions: industry, country, firm size, managerial context, time period. The classifier prompt instructs the model to favor conditional classification when findings are contingent, and the majority of papers in a result set will typically fall into this category. This is not a failure mode — it accurately reflects the nature of management research, which rarely produces unconditional universal claims.

Standard search engines return what you ask for. Arbiter’s synthesis explicitly instructs the model to find and name contradictory evidence. Seeing “6 Challenge” next to “119 Support” is the product working as intended, not a bug. Most novice literature reviews are unreliable because the researcher only found papers that agreed with their hypothesis. Arbiter fights this confirmation bias by making disagreement the same size and the same color as agreement.

Every ranking list edition is archived as a versioned data file with its effective date. Current editions: FT50 (2016), UTD24 (2024), ABDC (2025), AJG (2024). When a new edition is released, Arbiter ingests it as a new version; the prior version remains active until the user or institution migrates explicitly.

A paper’s ranking list membership is recorded at ingest time. If a journal is removed from a list in a subsequent edition, papers already indexed retain their historical membership. A researcher who needs “FT50 as of 2023” can specify that corpus explicitly.

Every citation export includes the methodology version and the list editions effective at query time. This supports the emerging norm in management research of including database-search parameters in methods appendices, supporting full query reproducibility in published work.

Cross-checked against the Retraction Watch database. Retracted papers are hidden from default search results; a toggle in the Refine panel surfaces them for researchers who study retraction itself. When shown, retracted papers display a danger-red badge and are excluded from composite evidence scoring entirely.

Papers under an active Expression of Concern are visible by default but carry a red EoC badge and a −0.25 penalty applied to their composite evidence weight. The penalty is shown in the weight breakdown panel on the paper detail screen — not silently applied. Researchers can see exactly why a score is lower than expected.

Venues from publishers on a maintained exclusion list — derived from documented predatory publisher indices, updated quarterly — are excluded from T1 ranking eligibility. They may still appear as T3 citation neighbors if cited sufficiently by T1 works, but carry a visible venue-quality warning and receive the minimum V score (0.05).

Incorrect authors, journal, year, or DOI. These are metadata errors typically traceable to source databases (OpenAlex, Crossref) and are corrected upstream and re-ingested.

The paper has been retracted but Arbiter has not yet received the Retraction Watch update. These flags are prioritized for same-day resolution.

A researcher believes the composite weight is materially incorrect — for example, due to a known citation-cartel inclusion or a venue reclassification not yet in the current list edition. These are logged for the quarterly methodology review.

A user identifies a venue not yet on the exclusion list that shows characteristics of predatory publishing. These are reviewed by the methodology team and escalated to the publisher exclusion process if warranted.

Arbiter indexes titles, abstracts, keywords, and author metadata. It does not search the full text of papers. For works where the abstract is absent — common for pre-2000 publications — a “no abstract available” indicator is shown. Researchers requiring full-text search should supplement with publisher databases.

Arbiter does not use author h-index, institutional affiliation, or career-stage signals in evidence weighting. This is a deliberate methodological choice to avoid amplifying citation-cartel behavior. The composite weight is a paper signal, not an author signal.

OpenAlex coverage for papers published before 1990 is substantially less complete than for recent work. Arbiter inherits this limitation. Researchers conducting historical literature reviews should supplement with JSTOR and publisher archives for pre-1990 coverage.

The AI synthesis is a language model output based on retrieved abstracts. The same query may produce slightly different phrasing across runs because the model operates at non-zero temperature. The underlying paper set is fixed and deterministic; the prose summary is not. Synthesis is a starting point for literature engagement, not a substitute for reading primary sources. Citations are verifiable; the synthesis framing should be treated as one reasonable interpretation of the evidence.

T3 citation neighbors — works cited by five or more ranked-journal papers — are not in the v1.0 corpus. This means some important methodological papers (statistics papers in psychology, foundational economics works) may not appear even when directly relevant. Researchers should supplement with discipline-specific databases when foundational methodology coverage matters.

Arbiter. (2026). Arbiter evidence methodology, v1.0. Opus Vita. https://arbiter.ac/methodology/v1.0

“Literature evidence was retrieved and synthesized using Arbiter (v1.0; Arbiter, 2026), which indexes journals on the FT50, UTD24, ABDC, and AJG ranking lists and applies a composite evidence weight combining query relevance (R), field-weighted citation impact (F), recency momentum (M), and venue quality (V). The Balanced mode (ε=0.40, β=0.40, γ=0.15, δ=0.05) was used. Retracted papers were excluded by default.”