Literature Screening Assistant

Import a Zotero-exported BibTeX file, configure the rules you care about, and run an AI triage pass powered by your chosen provider one record at a time.

Model & API keys

Choose a provider and supply the matching API key. Keys are stored locally in your browser and sent only with triage requests.

Provider

Switch providers any time; runs will use the active selection.

Model

Choose from supported OpenRouter models. Stored locally for future sessions.

Supports reasoning. ~2,000,000 prompt characters, 8,192 max output tokens.

OpenRouter API key

Create a key in OpenRouter settings, then paste it here.

Reasoning effort

Higher reasoning effort typically yields better screening quality, but may use more tokens and time.

Data policy override

Leave this set to the account default to honor your OpenRouter privacy mode. Provide a policy string to override it for triage requests.

Screening Criteria

Human-readable

Paste or type the criteria exactly as you would share them with a reviewer. We derive keyword heuristics for deterministic checks and feed the full text to the LLM triage step.

Inclusion criteria

1. Adults only – learners are adults (higher ed, workplace, community, immigration) or adult subgroup is clearly analyzable.
2. Speaking performance target – study targets L2 English speaking or a real-time production analog (speaking, oral proficiency, conversation, fluency, CAF, CEFR/ACTFL/OPIc ratings, pronunciation intelligibility, speech rate, articulation rate, pause metrics, response latency, role-play performance).
3. Behavioral-science mechanism present – intervention explicitly implements a skill-acquisition or adherence mechanism (spacing, interleaving, variable practice, retrieval/production practice, scaffolding, feedback timing/bandwidth/focus, adaptive difficulty/mastery criteria, implementation intentions, habit formation, reminders/prompts, commitment devices, incentives/reinforcement, goal/plan prompts, progress feedback, social accountability).
4. Performance outcome, not vibes – reports change in speaking performance or validated proxy for automaticity/transfer (CAF on novel tasks, latency, error decay, CEFR/ACTFL level change, intelligibility ratings). Attitudes/engagement alone do not qualify.
5. Design provides evidence – empirical primary study (RCT, quasi-experimental, pre-post with documented practice dose, field/classroom deployment) or a systematic/scoping review that maps mechanisms to outcomes with methods. Mixed methods allowed if performance outcomes are reported.
6. Scope fit – published 2015–present, or pre-2015 only if foundational mechanism directly applied to speaking design. ESL/EAL/EFL context or language-general mechanism with explicit mapping to speaking.
7. Technology neutrality – AI use is optional; include human-delivered mechanisms if portable to AI-assisted speaking systems.

Exclusion criteria

E01. Not adult population or adult data not separable.
E02. No speaking or real-time production outcome (knowledge tests, attitudes, usage only).
E03. No behavioral mechanism (tech/tool use without an explicit skill/adherence mechanism).
E04. Opinion/theory only; no data or evaluative pathway.
E05. AI/system paper with zero learner outcomes.
E06. Gray/industry report without transparent methods or outcomes.
E07. Out of timeframe and not a foundational mechanism applied to speaking.
E08. Mixed population where adult/speaking subset cannot be disaggregated.
E09. No comparator and practice dose uncontrolled/unstated in a way that prevents interpreting performance change.
E10. Insufficient intervention detail to implement the mechanism (cannot tell what was spaced/interleaved, how feedback was timed, what the adherence device was).
E11. Language/skill mismatch with no explicit mapping to speaking (e.g., reading/listening only, motor-skill analogs with no L2 transfer argument).
E12. Duplicate publication of an included study (retain the most complete version).