Data

Tools

Indexes

Applications

Experiments

Awesome

An open API service indexing awesome lists of open source software.

https://github.com/asmuelle/perennial

A literature review that never goes stale: subscribe to a research question and get a versioned, citation-verified survey that ingests each week's papers and flags which claims are now supported, contradicted, or superseded — every reference verified to exist.
https://github.com/asmuelle/perennial

academic citations literature-review llm research saas science

Last synced: about 18 hours ago
JSON representation

A literature review that never goes stale: subscribe to a research question and get a versioned, citation-verified survey that ingests each week's papers and flags which claims are now supported, contradicted, or superseded — every reference verified to exist.

Awesome Lists containing this project

README 

          
# Perennial



[![CI](https://github.com/asmuelle/perennial/actions/workflows/ci.yml/badge.svg)](https://github.com/asmuelle/perennial/actions/workflows/ci.yml)



> A literature review that never goes stale: subscribe to a research question and get a versioned, citation-verified survey that ingests each week's papers and flags which of its own claims are now supported, contradicted, or superseded — with every reference verified to actually exist.



**Category:** LLM wiki / auto-research (living documents + delta alerts, à la Karpathy) 



## Concept



A literature review that never goes stale: subscribe to a research question and get a versioned, citation-verified survey that ingests each week's papers and flags which of its own claims are now supported, contradicted, or superseded — with every reference verified to actually exist.



## Target User



PhD students, postdocs, lab PIs, and industry research scientists in fast-moving fields (ML, computational biology, neuro) drowning in arXiv's ~28K submissions/month. Individuals pay from personal or grant funds; PIs buy lab plans; adoption is viral within research groups because the shared review becomes the lab's canonical artifact.



## Auto-Research Mechanic (the living document + delta engine)



Seeded with a research question (optionally the user's own related-work draft), the agent generates a STORM-style structured survey with claim-level citations pinned to paper + section + quoted span. Then weekly maintenance: arXiv/OpenReview/bioRxiv/PubMed/Semantic Scholar deltas via RSS with semantic dedup; cheap-model triage filters AI-slop and citation-mill papers, including an integrity screen verifying every reference in a candidate paper resolves to a real DOI/arXiv ID — directly attacking the 1-in-277 fabricated-citation epidemic; a frontier pass integrates genuinely new results into affected sections only, marking existing claims newly supported / contradicted / superseded (Scite-style entailment applied to a living artifact, not retrospective lookup). Weekly delta email: '3 papers change Section 4; one contradicts the headline result you cite.' Full version history, per-claim provenance, verified BibTeX export.



## Product Surface



Web app — the living survey needs reading/diff UI — plus weekly email delta digest and BibTeX/Overleaf/Markdown export embedding it directly in the manuscript workflow.



## Why Now (2026 timing)



arXiv growth went super-exponential and arXiv banned unreviewed CS survey papers in Oct 2025 — survey demand is spiking exactly as human-authored supply is throttled. Fabricated citations rose 6x+ since 2023 with 100 fakes passing NeurIPS review, making 'every reference verified to exist and say what we claim' a marketable guarantee.



## Comparables



- Elicit — est. $18-22M ARR (2025), 50K+ paying users, tiers Free/$10-12 Plus/$42-49 Pro/Enterprise; $22M Series A; category leader and ARPU ceiling benchmark

- Scite — $3.6M ARR and ~21K B2C subscribers at acquisition by Research Solutions (NASDAQ: RSSS, Nov 2023); $12-20/mo individual; closest exit comp and B2C-to-site-license playbook

- Consensus — $1.5M ARR (Aug 2024), revenue up 8x within ~a year, $30M round led by GreatPoint Ventures, 8M users; $15ish/mo premium; proves fast growth is possible in this niche

- SciSpace — $12-20/mo Premium, $8-18/seat Teams, $70/mo Advanced Deep Review tier; proves premium-tier headroom above $20/mo for synthesis-grade output

- Undermind — $16/mo Pro, free tier 5 searches/mo; deep-literature-search adjacent comp

- Semantic Scholar / Google Scholar alerts — free; raw paper feeds with zero synthesis; the free-alternative pressure floor

- ChatGPT Deep Research / Gemini — ~$20/mo bundled; improving citation grounding is the main long-term moat threat, though currently perceived by researchers as the fabricated-citation source



## Adversarial Review — 



The pitch survives the wrong enemy and dies to the right one. (1) MISIDENTIFIED COMPETITION: Frontier labs won't build versioned scholarly claim graphs — but Scite, Elicit, Consensus, and Ai2 will, and they already own the hard assets. Scite has 1.2B+ citation statements and a production entailment classifier; adding 'living document' on top is a quarter of work for them, while Perennial rebuilding Scite's corpus takes years. Ai2's Semantic Scholar/Asta is free and grant-funded, and academics are the most price-sensitive, free-tool-defaulting audience in SaaS. Meanwhile ChatGPT/Gemini/Perplexity already ship scheduled tasks + deep research + proactive briefings (ChatGPT Pulse), so for the marginal grad student a $20/mo subscription they ALREADY pay delivers 70% of the perceived value, capping conversion. (2) THE MOAT IS CONDITIONAL: the 'claim graph compounds over months' argument only holds if incremental maintenance measurably beats a weekly from-scratch regeneration by a 1M-token-context frontier model — unproven, and regeneration cost falls every quarter. Version history and lab embedding are real but small-N lock-in. The 'anti-slop brand structurally unavailable to Big AI' claim is already falsified: Perplexity markets 94.3% citation accuracy; DOI existence-checking is a free Crossref call anyone can ship. (3) TRUST BAR: existence verification is the trivial 10%; supported/contradicted/superseded entailment in the wild is the unsolved 90% (SciFact-class models run ~70-85% F1 on curated data, worse on real papers). The users are domain experts — the one audience guaranteed to catch every error. One false 'this contradicts your headline result' email to a PI's lab, or one missed major result the user's colleague mentions first, breaks the only promise ('never goes stale, every claim verified') the product makes. Tolerable error rate is <5% false alarms and near-zero missed major signals in the user's own niche — current entailment tech does not clear that bar without expensive human-in-the-loop. (4) COGS UNDERSTATED 2-4x: weekly frontier integration passes over a 50-100K-token survey plus 5-20 deep-read papers runs $0.50-2.00/topic/week even with batch+caching; a 3-topic $19 user with active topics costs $6-24/mo, not $2-4. Margins are 40-70%, not 85%+, and the $19 tier is underwater for power users. (5) DATA ACCESS CLIFF: ML/CS preprints are open, but comp bio and neuro live behind Elsevier/Springer/Wiley paywalls a startup cannot legally full-text mine; entailment on abstracts is unreliable, so the product silently degrades outside preprint-native fields — shrinking TAM to exactly the field (ML) where free alternatives are densest. (6) CHURN: academic lit-review need is bursty (pre-submission, pre-grant); mature topics produce 'nothing changed' weeks that teach the user to cancel; students graduate; realistic individual subscription lifetime is 4-9 months. Scite and Elicit both pivoted toward institutional/publisher revenue precisely because individual academic churn is structurally bad — Perennial's expansion path (site licenses) means 12-18 month university procurement cycles a seed-stage team may not survive.



## Tech Stack & Unit Economics



DATA: arXiv OAI-PMH/RSS + OpenReview API + bioRxiv API + PubMed E-utilities (all free); OpenAlex as the canonical paper/citation backbone (free); Semantic Scholar Academic Graph API (free key, partnership tier needed for rate limits); Crossref REST for DOI existence verification (free); Unpaywall for OA PDFs; GROBID for PDF-to-structured-text. MODELS: Stanford STORM (open-source) as the survey-generation skeleton; frontier model (Claude Sonnet 4.6-class) via Batch API + prompt caching for initial synthesis and weekly integration passes; Haiku 4.5/Gemini Flash-class for abstract triage and slop screening; embeddings (Voyage/OpenAI) + pgvector for semantic dedup and topic-relevance filtering; cheap fine-tuned cross-encoder (DeBERTa/SciFact-style) as first-pass entailment with frontier-model adjudication only on flagged claim-paper pairs — this two-stage design is the only way the entailment COGS and precision both work. ORCHESTRATION: Temporal or Inngest for weekly per-topic cron pipelines with retry/replay; Postgres for the claim graph (claims, papers, entailment edges, char-span provenance); ProseMirror + structured diffing for the versioned survey UI; Next.js web app; Resend for delta digests; BibTeX/Markdown/Overleaf export. UNIT ECONOMICS (realistic, not pitched): initial survey $3-8 one-time per topic; weekly maintenance per ACTIVE topic = triage ~$0.05 (300 abstracts x cheap model, batched) + integration pass $0.40-1.80 (cached survey context + 5-20 deep-read papers, frontier batch) = roughly $2-8/topic/month; assuming 1.5 active topics average per individual user, COGS $4-12/user/month against $19 = 40-75% gross margin, requiring hard caps on topic count, integration only on affected sections, and aggressive caching to stay viable. Crawling/legal COGS near zero if strictly preprint+OA; any paywalled full-text ambition adds licensing costs that break the model.