Deltasci Ground

DeltaScience: The Grounding Layer (scan → gap → verify)

Purpose

Ground an AI-assisted research idea or draft against the real record, in three moves:

1. Scan — find the closest existing work (prior art) across OpenAlex, arXiv, PubMed, GitHub.
2. Gap — judge whether that space is crowded, contested, or open, and name the distinguishing angle.
3. Verify — check every citation against the source of truth: does the identifier resolve, does the metadata match, does the cited paper actually support the claim?

The one rule that governs everything: no LLM in the trust path

There are two kinds of work here, and they are not symmetric:

• Discovery (scan, gap) is yours. You are the LLM. Write the queries, judge which results are genuinely relevant, reason about where the gap is. A weak discovery pass can only make you miss a paper — it cannot make a false statement about what exists — so your judgement is welcome here.
• Trust (verify) is the engine's. A citation is real, and supports its claim, only when deltasci verify says so. Never assert from memory that a PMID/DOI is valid or that a paper supports a claim. The engine fetches the live record and decides deterministically. This is the entire point of the tool: the verdict must not depend on a model that can hallucinate agreement.

If you ever catch yourself about to write "this citation looks correct" without having run deltasci verify, stop and run it.

Prerequisites

pip install deltasci          # core engine (keyless)
pip install 'deltasci[pdf]'   # add PDF support for whole-paper input

The engine is deterministic and needs no API key. All three commands emit --json.

Inputs

Procedure

Step 1 — Frame the idea

Read the idea or the paper's title + abstract. Identify, in the field's standard vocabulary:

• the problem (e.g. "long-horizon sparse-reward credit assignment"),
• the technique (e.g. "group-based reinforcement learning for LLM agents"),
• the application / domain.

Critical: find the paper's own coined names — its method, system, or benchmark names (e.g. a made-up acronym like SkillEvo, WebArena-Lite) — and set them aside. Never search for them. No other work uses those terms, so they return nothing and poison a query. This is the single most common reason a scan finds "no prior art" for a hot area.

Step 2 — SCAN: you write the queries, the engine fetches real records

Write 3–5 search queries, most-specific first, collectively covering problem + technique + application, using canonical terms and synonyms. Then issue them with the explicit-query primitive:

deltasci scan \
  --query "llm agent reinforcement learning skill" \
  --query "long-horizon sparse reward credit assignment" \
  --query "group relative policy optimization GRPO" \
  --json

• For a non-biomedical idea, drop PubMed: --sources openalex,arxiv,github.
• --limit 20 for a wider net.

Every hit in the JSON is a real, retrieved record (title, authors, year, venue, url). Read them and rerank by genuine relevance — judge by meaning, not shared words. A paper that merely shares vocabulary but solves a different problem is not close. Do not invent or embellish any record; only use what the engine returned.

If failed_sources is non-empty, a corpus was slow/rate-limited — note it as a coverage gap (the run is incomplete, not empty).

Step 3 — GAP: reason, grounded only in the retrieved works

From the real hits, classify the space:

• CROWDED — strong, direct prior art exists; the researcher should read it before building.
• CONTESTED — adjacent work exists; a distinguishing angle is needed.
• OPEN — little direct prior art.

Ground every statement in the listed works, naming them by author and year. State what the works already cover and the one distinguishing angle the idea leaves open. Never invent a paper to fill the story.

Honesty rule on absence: you may call CROWDED or CONTESTED freely (you can't un-find a close match). But only call OPEN if the scholarly sources (OpenAlex / arXiv / PubMed) actually answered. If one failed, the space is INCONCLUSIVE — re-run, never "open." Absence of evidence from a source that didn't respond is not evidence of an open gap.

Optional deterministic cross-check (density-based, keyless):

deltasci gap --query "llm agent reinforcement learning skill" --json