Glass-box scholarly evidence — how we used Scite without fabricating it.

A general-purpose LLM will happily invent a citation. It will produce DOIs that look plausible but resolve to nothing, attribute findings to authors who never wrote them, and do this with confidence calibrated to your tone, not to truth. For background reading, this is annoying. For an evidence layer that other people will quote, it is disqualifying.

Scite’s MCP server is the cheapest available fix. Every paper returned by mcp__scite__search_literature is a real record with a Smart Citation graph attached (supporting / contrasting / mentioning), an editorialNotices field that exposes retractions and expressions of concern, and — for open-access works — full-text excerpts. The tool itself enforces the rule that the agent must not paraphrase its own training data.

The gap, in practice, is between the MCP’s affordances and what a one-shot user actually gets out of it. A single agent firing one query produces a shallow sweep. The system described below is the answer to that gap.

The pipeline is two Claude Code skills and one subagent. /scite-preflight writes a mission folder to disk — README with strategy, one editable markdown prompt per agent, a synthesis instruction file, and an empty results directory. Nothing executes. The commander reads it, edits anything that looks wrong, and only then invokes /scite-research, which launches 1–10 parallel workers and synthesises a tiered evidence report.

The split exists because planning errors and execution errors fail differently. A vague key claim is cheap to fix in markdown; the same vagueness costs five wasted MCP calls if it ships into an agent run. Fusing the two phases would force the commander to carry both mental loads at once. Splitting them makes each phase inspectable on its own terms.

Because the mission folder lives on disk, partial runs resume cleanly. /scite-research checks results/ before launching; any agent with an existing agent-NN-results.json is skipped. A rate-limit hit or a crashed session does not require starting from zero — which matters more than it sounds, as the Phase 2 incident showed.

On 2026-04-14, Batch 1 of the human-history mission ran ten parallel agents against eras 01–10 and consumed roughly 150 of the account’s 250 monthly MCP calls. Batch 2 was launched optimistically — and seven of its ten agents (eras 14–20) hit the wall on their first call:

The right reading is not “Scite’s default limit is too low” (it is appropriately scoped for solo exploratory work). The right reading is that a multi-agent batch pipeline must treat the monthly call budget as a first-class design constraint, not as something to discover empirically.

On 2026-04-16 the pipeline was hardened. A Phase 2.5 step now sits between planning and execution: the preflight skill loads ~/.scite-quota.json, computes the projected cost of the mission as N_agents × 7 + 3 (five searches plus one adaptive buffer plus three synthesis-side section reads), compares it against remaining budget, and emits a verdict.

The result is that the rate limit stopped being an after-the-fact obstacle and became a pre-launch parameter. When the next mission is planned, the budget block tells the commander whether it fits this cycle or needs to wait for the reset on the first of the month.

Scite is narrow on purpose. It indexes peer-reviewed scientific literature with a citation-classification graph; it does not cover preprints comprehensively, monographs, grey literature, news, or domain-specific corpora. The project leans on a small set of complementary tools, each chosen for what Scite cannot do.

The pattern that emerged: Scite is reserved for the moment when a claim is about to be cited in something that someone might quote. Everything before that is allowed to be cheaper, faster, and more speculative — provided it stops short of the citation boundary. Crossing that boundary without an MCP call is a project-level rule.

Finding A. The dois[] batch parameter is case-sensitive. Discovered by Agent 02 of the Mediterranean-diet validation, then patched in the worker skill: lowercase every DOI before passing it to dois=[]; preserve the original case in output JSON. Verified against the human-history Phase 3 mission with no regressions on archaeology / classics DOIs.

Finding B.Pre-2000 humanities monographs are systematically under-indexed in Scite’s Smart Citation graph. They appear as in-text citations inside other papers but are not retrievable as Scite records of their own. The list is not obscure: Snodgrass Archaic Greece (1980), Hansen The Athenian Democracy in the Age of Demosthenes (1991), de Polignac Cults, Territory, and the Origins of the Greek City-State (1995), Thapar Aśoka and the Decline of the Mauryas (1961). These are universally-cited foundational works in their subfields.

The implication is structural: the worker’s tier rules had to grow a humanities branch. In humanities, support_ratio is null for most papers (the field does not generate contrasting Smart Citations the way biomedicine does), and the Tier 1 rule is citing_publications_count > 30 on a journal article rather than a support-ratio threshold. A monograph citation-proxy provision is on the future-work list — for now, the field gap is documented in the output, not papered over.

• Field-asymmetric coverage. Pre-2000 humanities monographs are under-indexed (see Finding B). Biomedical and natural-science coverage is strong; archaeology and classics are weaker. Research plans should price this in.
• 250-call default ceiling. Adequate for 3–5 standard missions; tight for the full 50-era project. The budget check prevents surprises but does not raise the ceiling.
• Reproducibility against a living index. Scite is updated continuously. The method is reproducible; specific DOI lists carry a timestamp. Re-running a mission six months later may surface different papers — that is correct behaviour, not drift.
• The sharpened-claim pattern is manual. Replacing “first democracy” with “the three-feature Cleisthenic combination” is a commander-side framing decision. The system supports the pattern but does not automate it.
• Synthesis cost. The Opus synthesis step is the most expensive call in the pipeline. For rapid exploration a Sonnet synthesis is acceptable; for anything that will be quoted, Opus stays.

This page is a compressed reading. The authoritative artefacts are checked into the repository and unchanged from the original 2026-04-16 sprint:

Every metric on this page traces to a specific file in that inventory. Nothing is asserted that is not retrievable. Read the full whitepaper if you want the math behind a number; this page is here to make the system legible to the people who will quote it.

← Back to methodology