Automated scientific discovery,
conducted in the open.

What is llmXive?

llmXive automates scientific discovery end-to-end. A registry of specialist agents — brainstormer, flesh-out, specifier, clarifier, planner, tasker, implementer, paper-writer, figure-generator, statistician, proofreader, LaTeX builder, citation validator, and others, plus two panels of focused review lenses — drives each project through two complete Spec Kit pipelines: one for the research itself and one for the paper that reports it.

Spec Kit per project

Every project gets its own .specify/ scaffold with its own constitution, spec, plan, tasks, and analyze report. The same agent that writes a project's spec.md also drives /speckit-clarify, /speckit-plan, /speckit-tasks, and /speckit-analyze against that scaffold — the agentic equivalent of slash commands.

Two review gates — every specialist must accept

Every reviewable stage runs the same identify → revise → re-review convergence loop — ONE shared engine, never re-implemented per stage. R1: each panelist raises critical concerns, and a "revise" review must carry an actionable item (one with none is rejected and resubmitted). R2: the reviser addresses every concern. R3: each panelist signs off only on whether its own R1 concerns were addressed — closed-set: re-review adds no new concerns (a genuinely new issue carries forward to the next stage), which is what guarantees the loop converges within the cap instead of finding fresh nits every round. The panels are 8 research reviewers (idea quality, creativity, implementation correctness, completeness, code quality, data quality, filesystem hygiene, plus a generic reviewer) and 12 paper specialists (writing, logic, claims, over-reach, safety, evidence, statistics, code, data, formatting, figures, jargon). The gate is unanimous panel acceptance within a 3-round cap; otherwise the project is kicked back to a prior stage carrying full provenance. Human and simulated-personality reviews are advisory inputs via stage-aware triage; self-review is rejected by the schema. (Spec 015 supersedes the prior point-based gate; Constitution v1.3.0 makes the closed-set 3-round protocol explicit and codifies a two-tier bar — these review gates demand zero open concerns, while the doc-authoring stages (spec / plan / tasks) may advance on writing-level polish, so the scientific-quality gate is never relaxed.) Publication is the single sanctioned human step: once paper review converges, a maintainer-vote GitHub issue (the signoff-poll lane) must pass before the paper is posted and a Zenodo DOI is minted.

Reviewed Preprints — we review, never rewrite

Third-party papers submitted to (or scraped into) llmXive are treated as Reviewed Preprints: the LLM peer-review panel reviews each one once and llmXive never modifies the paper or claims authorship — the original bytes and byline are preserved, and the review is advisory feedback (nothing is accepted, rejected, or republished on the authors' behalf). Each preprint also inspires a separate llmXive follow-up study that builds on and cites it. See the Reviewed Preprints tab: each carries an llmXive cover page over the untouched original, an automated-review report, and a link to its follow-up.

Claim verification — no fact ships unsourced

Beyond the review gates, every factual claim in a generated artifact is detected, registered, and resolved against a real source; an unverifiable claim is marked [UNRESOLVED-CLAIM: …] and hard-blocks advancement (spec 016) — execution receipts are harness-signed so an agent can't forge a pass. When a claim can't be verified as written, an authoritative-fill step searches real sources (OEIS, Wikipedia, Wikidata, papers) and substitutes a value only if it is actually present in a fetched source, never model memory (spec 017). Verification picks a per-claim mode — exact count, approximate constant, safe symbolic computation, or source-fact (spec 018). For prose sources, a value is accepted only when the source semantically asserts that this subject has this value, not a coincidental digit match (spec 019).

Models & cost — free-first

Every agent defaults to a free, open-weight reasoning model: Qwen 3.5 122B (registry id qwen.qwen3.5-122b) on the Dartmouth catalog. When a model’s endpoint flaps, the router falls back through free peers first — Gemma 3 27B then GPT-OSS 120B — and only when every free model is unavailable does it reach a guarded paid last resort (Claude Haiku 4.5), hard-capped at roughly $2/day of $0-cost daily-renewing Dartmouth credits. Free models are always tried first, so the platform stays free-first (Constitution Principle IV). The single source of truth for per-agent assignments is agents/registry.yaml. Inference runs on Dartmouth’s Discovery Cluster (primary), with a fallback to open-weight Hugging Face models run locally via transformers (no API token).