4
0–100 evaluators
ARES — AI Research Excellence System
From raw idea
to published paper.
A complete operating system of prompts, skills, evaluators, and templates that turns an AI assistant into a disciplined research partner for the world's most demanding venues.
Release v3.2·
Verified 5 Nov 2026·
Engine ARIA v3.1·
License MIT·
Language plain Markdown·
Install 1 paste
ARES is a repository of forty-plus modular documents that take a general-purpose LLM and give it a research editor's discipline: adversarial peer review before submission, quantitative 0–100 scoring gates, venue-exact templates, and a strict evidence floor that refuses fabricated citations. This release re-verifies every dataset, trial identifier, and DOI against live registries — nothing here is remembered, everything was checked.
The ten-station pipeline — say the word, get the deliverable
IDEATION → LITERATURE → METHODOLOGY → run your study → DATA → WRITING → QUALITY → JOURNAL → CITATION → IMPACT
§01 · The problem
Why this system exists
Generic LLM prompting produces reviewer bait: shallow summaries, hallucinated citations, boilerplate prose.
Ask an ordinary AI assistant to "help with your paper" and you get a plausible-sounding draft with invented DOIs, generic claims, and none of the falsifiability elite reviewers demand. ARES takes the opposite approach on four fronts:
Epistemic depthvs. summaries
Every skill forces causal depth, falsifiable hypotheses, boundary-condition stress tests, and an anti-summary standard that synthesizes tension instead of listing chronology.
Floornon-negotiable
Adversarial reviewbefore submission
A five-persona panel — Methodologist, Domain Expert, Logician, Communicator, Impact Assessor — attacks the manuscript the way Reviewer 2 will, then a meta-review consolidates the verdict.
5×1reviewers
Quantitative gates0–100 rubrics
Four independent evaluators score the manuscript with weighted dimensions and explicit go/no-go thresholds. No score, no submission.
4engines
Empirical provenancereal-world only
Every example is grounded in verified public datasets, clinical trials, and DOI-registered literature. Fabricated identifiers are classified as critical flaws and eliminated — this release itself corrected six.
100%resolvable
§02 · The system, in numbers
A fact ledger
Every figure below is a verifiable property of this repository.
10
Research skills
4
End-to-end workflows
7
Venue templates
4
Worked exemplar papers
11
Audit dimensions (Skill 06)
12
Roadmap steps (v3.2)
125
Learning library files
14,784
PeerRead paper drafts
83.8%
Pairwise rank accuracy
PeerRead figures per Kang et al., NAACL 2018 (arXiv:1804.09635): 14,784 paper drafts with accept/reject decisions, of which 10.7K include full textual expert reviews. The 83.8% figure is the relative-rank engine's validated pairwise accuracy on a 100-pool ICLR 2017 benchmark.
§03 · Start here
Five steps, five minutes, zero setup
No plugins. No install scripts. Just Markdown and one paste.
Step 12 minutes
Get the repository.
git clone https://github.com/AdilShamim8/ARES.git — or press Code → Download ZIP on GitHub. No build step; it is all plain text.download
Step 22 minutes
Install ARIA. Copy the boxed prompt inside
MASTER_PROMPT.md and paste it into Claude Project Instructions (recommended) or as the first message of a new chat. This one paste activates the five-step cognitive engine for every reply.1 paste
Step 31 minute
State your topic. Be specific: "Sample-efficient graph neural networks for drug-target affinity prediction, targeting Nature Machine Intelligence." No topic yet? Ask for a high-novelty question in your field instead.
1 sentence
Step 4on demand
Activate skills like switches. Each command returns one structured deliverable — see the skill table in §05. Example:
ACTIVATE: IDEATION returns a gap matrix plus a scored novelty certificate.10 commands
Step 5iterate
Loop until the gates are green. Run the evaluators, fix what the simulated reviewers found, repeat. The pipeline in the masthead is the recommended order for a full paper.
ship it
Beginner note
If you have never used an AI assistant for research: you do not need to read all 40 files. Read this page top to bottom once, paste one prompt, and start with Step 3. The system will tell you what to do next at every stage.
§04 · The map
How a manuscript moves through ARES
One idea enters; a submission-ready manuscript exits.
System flow — install once, execute on demand, gate before submission
┌──────────────────────────────────────────────────┐
│ YOUR IDEA / EXISTING DRAFT │
└────────────────────────┬─────────────────────────┘
│
▼
┌──────────────────────────────────────────────────┐
│ MASTER_PROMPT.md · ARIA v3.1 engine │
│ 5-step cognitive architecture runs before │
│ every response: Summit Scan → Gap Analysis → │
│ Non-Obvious Bridge → Adversarial Test → Synthesis│
└────────────────────────┬─────────────────────────┘
│
┌──────────────────────┬──────────┴─────────┬──────────────────────┐
▼ ▼ ▼ ▼
┌──────────────┐ ┌───────────────┐ ┌──────────────┐ ┌──────────────┐
│ SKILLS 01–10 │ │ WORKFLOWS ×4 │ │ TEMPLATES ×7 │ │ EXAMPLES ×4 │
│ on command │ │ 28-wk / 30-dy │ │ Nature·Sci· │ │ full worked │
│ ACTIVATE: X │ │ interdisc.· │ │ NeurIPS·IEEE │ │ papers with │
│ │ │ replication │ │ ACM·Springer │ │ real datasets │
└──────┬───────┘ └───────┬───────┘ │ arXiv │ └──────────────┘
│ │ └──────────────┘
▼ ▼
┌──────────────────────────────────────────────────────────────┐
│ EVALUATORS · 0–100 with gates │
│ Novelty ≥70 Rigor ≥75 Clarity ≥80 Impact ≥75 │
└──────────────────────────────┬───────────────────────────────┘
▼
┌──────────────────────────────────────────────────────────────┐
│ QUALITY FORTRESS · 11-dim audit + VLM + 5-panel │
│ adversarial review, calibrated on PeerRead (83.8%) │
└──────────────────────────────┬───────────────────────────────┘
▼
┌─────────────────────────────────┐
│ SUBMISSION-READY PAPER │
│ + rebuttals, cover letters, │
│ policy briefs, follow-ups │
└─────────────────────────────────┘
§05 · The ten skills
Stations on the assembly line
Each directory holds an authoritative SKILL.md plus specialized modular prompts.
01 · skills/01_ideationIdeation EngineGap typology, contrarian probes, hypothesis forge, and a scored novelty certificate with go/no-go bands.ACTIVATE: IDEATION
02 · skills/02_literature_commandLiterature CommandTemporal mapping, intellectual genealogy, consensus mapping — synthesis, never chronological summary.ACTIVATE: LITERATURE
03 · skills/03_methodology_architectMethodology ArchitectDesign selection matrices, threat-to-validity audits, statistical power plans, pre-registration language.ACTIVATE: METHODOLOGY
04 · skills/04_data_intelligenceData IntelligenceResults audits, effect-size architecture, colorblind-safe figure specs, the 30-second figure test.ACTIVATE: DATA
05 · skills/05_writing_excellenceWriting ExcellenceSeven-sentence abstracts, introduction arcs, methods precision, discussion defense, closing impact.ACTIVATE: WRITING
06 · skills/06_quality_fortressQuality FortressConference-grade review: 11-dimension audit, VLM figure inspection, PeerRead-calibrated relative rank.ACTIVATE: QUALITY
07 · skills/07_journal_strategyJournal StrategyVenue decision matrix across five dimensions, editor cover letters, reviewer-response mastery.ACTIVATE: JOURNAL
08 · skills/08_ethics_complianceEthics & ComplianceBias detection, IRB checklists, FAIR data management plans, dual-use safety review.ACTIVATE: ETHICS
09 · skills/09_citation_masteryCitation MasteryCitation cartography, balance audits, multi-format bibliography engines, hallucination screening.ACTIVATE: CITATION
10 · skills/10_impact_amplifierImpact AmplifierPress releases, policy translation memos, social threads, and the follow-up research agenda.ACTIVATE: IMPACT
§06 · The gates
Four quantitative evaluators
Weighted 0–100 rubrics with explicit submission thresholds — a manuscript that cannot pass does not ship.
Novelty Scoreevaluators/NOVELTY_SCORE_RUBRIC
Conceptual (25%), methodological (25%), empirical (20%), translational utility (15%), interdisciplinary synthesis (15%). Top-tier gate for Nature / NeurIPS-grade claims.
≥ 70/100
Rigor Scoreevaluators/RIGOR_RUBRIC
Validity (25%), statistical power & inference (25%), reproducibility & open science (20%), baseline fairness (15%), boundary conditions (15%). Mandatory pre-submission audit.
≥ 75/100
Clarity Scoreevaluators/CLARITY_ANALYZER
Hourglass architecture, paragraph cohesion, cognitive-load economy, visual grammar, terminology precision, abstract calibration — six axes of communication quality.
≥ 80/100
Impact Predictionevaluators/IMPACT_PREDICTOR
24–60-month citation trajectory: theoretical longevity, tool adoption, cross-disciplinary diffusion, translational & policy utility, field-shaping robustness.
≥ 75/100
§07 · The adversarial core
Quality Fortress — peer review before peer review
Accepts LaTeX, PDF, Word, and Markdown. Calibrated on real conference review corpora.
Skill 06 simulates a senior area chair's board: a security scan (prompt-injection defense, double-blind leaks, unicode obfuscation), an 11-dimension technical audit with VLM figure inspection, and a five-persona adversarial panel whose concerns consolidate into one meta-review with a priority-ordered fix list.
| Empirical grounding | What it is | Role in the system | Verified via |
|---|---|---|---|
| PeerRead (Kang et al., NAACL 2018) | 14,784 paper drafts with accept/reject decisions from ICLR 2017, NeurIPS 2013–2017, ACL 2017; 10.7K with full expert reviews | Empirical thresholds for acceptance probability and rejection triggers | arXiv:1804.09635 |
| PeerRead 100-pool benchmark | 100 sampled ICLR 2017 pools | Validates the relative-rank engine: 83.8% pairwise ordering accuracy | skills/06 docs |
| AAAI 2026 calibration | Reviewer guidelines & decision boundaries, 40th AAAI (Singapore, Jan 20–27, 2026) | Reverse-engineered reviewer decision probes | aaai.org/conference/aaai/aaai-26 |
| OpenReview venue pools | Public review archives, ICLR / NeurIPS / ICML / CoRL, 2020–2026 | Relative rank against the live accepted-paper distribution | openreview.net |
| NeurIPS reproducibility checklist (Pineau et al., JMLR 2021) | The standardized 21-point artifact audit | Reproducibility & open-science scoring | JMLR 22 (164) |
| S2ORC + SciCite (Lo 2020; Cohan 2019) | 81M+ full-text papers with classified citation intents | Citation cartography & attribution auditing | ACL Anthology |
The 11 dimensions, compressed
[A] logic & argumentation · [B] empirical rigor & cross-table consistency · [C] writing quality · [D] citation cartography (zero hallucinated references) · [E] mathematical & notational integrity · [F] double-blind compliance · [G] venue formatting · [H] anti-AI-tone purity · [I] narrative structure · [J] red flags & puffery · [K] visual & figure aesthetics (VLM).
§08 · Provenance
Real-world datasets, collected and re-verified
Every exemplar paper is grounded in public data you can download today. All identifiers below were re-resolved on 30 October 2026.
| Domain | Exemplar study | Verified datasets & registries | Primary sources (resolvable) |
|---|---|---|---|
| Computer Science | MENTOR — metacognitive self-regulation for LLM reasoning | GSM8K · SVAMP · StrategyQA · ProofWriter · BIG-bench Hard | arXiv:2111.03797 · arXiv:2103.07191 · 10.1162/tacl_a_00370 · 10.18653/v1/2021.findings-acl.317 · arXiv:2210.09261 |
| Biomedical | CRISPR editing landscapes in sickle-cell HSCs | CTX001 trial (Phase 1/2/3, Vertex) · GUIDE-seq · BCL11A +58 erythroid enhancer | NCT03745287 · 10.1038/nbt.3117 · Frangoul 2021 10.1056/NEJMoa2031054 · Frangoul 2024 10.1056/NEJMoa2309676 |
| Social Science | Social media & democratic deliberation, N = 1,500 three-experiment design | 1,124-paper PRISMA corpus · Discourse Quality Index · pre-registered panel | Bail 2018 10.1073/pnas.1804840115 · Vosoughi 2018 10.1126/science.aap9559 · Steenbergen 2003 10.1057/palgrave.cep.6110002 |
| Interdisciplinary | BioPhysiFold — biophysically-informed foundation models for drug discovery | ChEMBL Release 37 (2.92M compounds) · ZINC20 · PubChem · PDBbind v2020 · MoleculeNet | 10.1093/nar/gky1075 · 10.1021/acs.jcim.0c00675 · 10.1093/nar/gkaa971 · 10.1021/acs.accounts.6b00491 · 10.1039/c7sc02664a |
Verification policy — how this release was built
DOIs are resolved against doi.org and Crossref; trial IDs against the ClinicalTrials.gov API; database versions against live vendor APIs (ChEMBL 37, PDBbind 2025); journal metrics against the 2025 JCR release. Unregistered identifiers are classified as critical flaws and corrected or removed — v3.0 fixed six of them. Illustrative placeholders inside examples are explicitly labeled so they can never masquerade as real registrations.
§09 · Learn the craft
The writing roadmap: how to write a research paper, end to end
One manual built from the courses, videos, and books · a 12-step schedule on top of it · every step has hands-on work and an exit test.
The library runs on two documents. THE WRITING ROADMAP is the manual — how to actually write each part of the paper, with every rule traced to the course, video, or book it comes from. The 12-step program (below) is the schedule — eight weeks part-time, one exit test per step.
| The manual — five parts | What it teaches | Its sources |
|---|---|---|
| Part 0 · Orientation | Where every teaching comes from; three ways to use the guide | the whole library |
| Part 1 · Before you write | Reading notes, the RQ ladder, the gap statement | Ref-n-Write video #4 · RQ guide · Swales & Feak |
| Part 2 · Academic writing style | Clutter, verbs, sentences and paragraphs, story, moves→phrases, ESL edge, drafting order | Stanford Units 1–4 · Schimel · Glasman-Deal · Phrasebank · Wallwork |
| Part 3 · Section-by-section | All 10 manuscript parts: video → Wiley module → pattern → style rules → pitfalls → work | the 9-video playlist · all 13 section guides |
| Part 4 · Revision | Three passes — structure, clarity, surface — plus the gates | Stanford Unit 4 · Duke · Purdue OWL |
| Part 5 · Submission | Authorship, journal choice, cover letter, packet, peer review | Wiley's five non-section modules · École Polytechnique · Nature courses |
The schedule — the 12-step program. Each step file teaches one idea, assigns real work (exercises and fill-in templates, not just reading), shows a worked example, and ends with an exit test that gates the next step. Steps 5–9 link each writing step to its full treatment in the manual. Click any step below to open it on GitHub.
Phase 1
Foundations
Week 1
Before you write a sentence: install the professional's tools, learn the paper's anatomy by dissecting one, build an evidence pile, and forge a research question that passes four hard tests.
1
Install Zotero and the Manchester Phrasebank; save your first three papers.
You produce: 3 papers in a synced reference library
1 h
2
Dissect one real published paper: label all nine parts, find the gap sentence.
You produce: An annotated paper + a word-count ledger
45 min
3
Three-pass reading with a note template: design, finding, limitation — per paper.
You produce: Five reading notes (your first evidence pile)
90 min
4
The topic-to-RQ ladder, the four tests, and the three-sentence gap statement.
You produce: One RQ that passes focused / answerable / falsifiable / sized
90 min
Phase 2
Write the paper
Weeks 2–5
Each manuscript section gets its own step, in the order professionals draft: review, methods, results, then the argument sections, then the miniature (abstract, title) last.
5
The synthesis matrix: 8–10 papers compared by theme, then two synthesis paragraphs.
You produce: A completed matrix + paragraphs that compare, not list
2–3 h
6
Draft to the reproducibility contract; audit materials, procedure, parameters, statistics, ethics.
You produce: Methods a competent stranger could rerun
2 h
7
Finding → evidence → pointer, one figure built to checklist, zero bare numbers.
You produce: Three results paragraphs + one audited figure
2–3 h
8
The introduction funnel on your gap statement; the six-move discussion; strike overclaims.
You produce: A funnel intro + a discussion that survives hostile reading
3 h
9
Three-sentence conclusion, five-move abstract within the venue cap, six-point title check.
You produce: A standalone abstract + a title that survives the search-result test
2 h
Phase 3
Finish and submit
Weeks 6–8
Revision as three disciplined passes, a venue choice you can defend, a cover letter that argues fit — and a rehearsed answer to your harshest reviewer.
10
Three passes — structure, clarity, surface — plus the read-aloud test and the ≥10% cut.
You produce: A measurably shorter, cleaner manuscript
3 h
11
Three-venue comparison worksheet; the four-paragraph cover letter that argues fit.
You produce: A defensible venue choice + a letter for one journal only
2 h
12
The submission packet checklist; rehearse the reviewer response; plan the waiting months.
You produce: A complete packet + a rehearsed point-by-point response
2 h
How the roadmap was built — and how to trust it
Every course, video, book, and tool the steps point to lives in the 125-file reference library behind the roadmap (research_paper_writing/), and every external fact — ratings, module counts, prices, channel statistics — was read from the live web on 5 November 2026 and is documented in the library's verification audit trail (09_SOURCE_VERIFICATION/). Practice examples inside the steps are labeled hypothetical: they demonstrate craft, never report findings.
| Reference shelf | Files | Open it when you need… |
|---|---|---|
| 01_COURSES/ | 25 | A full course — Stanford's free Writing in the Sciences (4.9★, 9,859 reviews), École Polytechnique, Nature Masterclasses, Wiley's 14-module guide, CDC/TEPHINET… |
| 02_VIDEO_COURSES/ | 19 | The anchor Ref-n-Write 9-video playlist with study notes, plus six expert channels |
| 03_BOOKS/ | 8 | The canonical bookshelf — Gastel & Day 9th ed., Schimel, Wallwork, Glasman-Deal… |
| 04_FREE_GUIDES/ | 8 | Free references — Purdue OWL, Duke, the Manchester Phrasebank, IEEE Author Center |
| 05_MENTORSHIP/ | 5 | A human to read your draft — AuthorAID (free), Grad Coach, Nature editor workshops |
| 06_TOOLS/ | 6 | Software while you write — Zotero, Overleaf, Paperpal, SciSpace + comparison matrix |
| 07_SECTION_GUIDES/ | 13 | Deeper written treatment of one manuscript section |
| 08_LEARNING_PATHS/ | 7 | A sequence by career stage (undergraduate → PhD, ESL, clinical, CS) or the 30-day sprint |
§10 · Timelines & formats
Workflows and venue templates
Four pipelines for four ambitions; seven blueprints for seven publishing houses.
| Workflow | Timeline | Use when | |
|---|---|---|---|
| Full Paper Lifecycle | 28 weeks · 8 phases | You have a topic and want the maximum-quality pipeline | comprehensive |
| Rapid Publication | 30 days | A breakthrough needs a priority claim — without sacrificing rigor | compressed |
| Interdisciplinary Innovation | stage-gated | The work bridges two fields and must satisfy reviewers from both | cross-domain |
| Replication & Extension | stage-gated | Replicating prior work with pre-registration and equivalence testing | pre-registered |
| Template | Venue | Current limit (checked Oct 2026) |
|---|---|---|
| Nature_Family.md | Nature & Nature Portfolio | Article ≈ 2,500 words (6 pages) to 4,300 words (8 pages); Methods ≤ 3,000; summary ≤ 200 |
| Science.md | Science (AAAS) | Research Article ~4,500 words; Report ~2,500 words |
| NeurIPS_ICML_ICLR.md | Top ML conferences | 9-page main text (NeurIPS 2026), proofs & error bars, ethics statement |
| IEEE.md | IEEE Transactions | Two-column, formal math, complexity analysis |
| ACM.md | ACM SIGCONF / TOCS | CCS concepts, artifact badges |
| Springer.md | LNCS & Springer journals | Monograph & journal standards |
| arXiv.md | arXiv preprints | Citation-velocity-first preprint architecture |
Journal quick reference (verified 2025 JCR release, June 2026): Nature 56.1 · Science 45.8 (2024) · NEJM 84.5 · Cell 45.1 · PNAS 9.5 · Nature Machine Intelligence 29.8. Official conference acceptance rates, latest cycles: NeurIPS 24.5% (2025) · ICML 26.9% (2025) · ICLR 32% (2025). Full table in SETUP.md.
§11 · Release history
Release history
Newest first. All facts on this page were re-verified against live registries on 5 November 2026.
v3.25 Nov 2026 · current
The beginner-first roadmap release. The Research Paper Writing library is reorganized around a 12-step roadmap (START_HERE → toolkit → anatomy → reading → research question → six writing steps → editing → venue → submission), each step with hands-on exercises, fill-in templates, practice examples, and an exit test that gates the next step. The library grows to 125 files and adds THE WRITING ROADMAP — the end-to-end manual (five parts: preparation, academic writing style, all ten manuscript sections, revision, submission) in which every rule is traced to the verified course, video, or book it comes from; the README leads with it. This page gains the roadmap section, mobile stacked tables, a scrollable phone nav, a back-to-top control, and scrollspy navigation.
3.2this release
v3.15 Nov 2026
Research Paper Writing learning library added — 110 new files across ten folders: 25 verified course profiles, the Ref-n-Write 9-video playlist with study notes, 8 canonical books, 8 free guides, 5 mentorship programs, 6 tools, 13 manuscript-section guides, and 7 career-stage learning paths. Every course rating, module count, price, and channel statistic was read from the live web on 5 November 2026 and is documented in the library's verification audit trail. Main README gains a dedicated section and the architecture tree.
3.1superseded
v3.030 Oct 2026
Repository renamed CRES → ARES to match the new GitHub home. 40+ files rebranded; six unregistered/mismatched DOIs corrected; clinical trial ID fixed to
NCT03745287; ChEMBL updated to Release 37; PeerRead identifier corrected to arXiv:1804.09635; the unverifiable "Poldrack" attribution replaced with the verified AgentReview framework (Jin et al. 2024, arXiv:2406.12708); journal metrics refreshed to the 2025 JCR release; Nature word-limit guidance aligned with the current author guide; broken skill cross-references repaired; new beginner-first README and this page.3.0superseded
v2.4September 2026
VLM multimodal figure review, multi-format manuscript intake (LaTeX / PDF / DOCX / MD), PeerRead 100-pool benchmark, conference-grade Quality Fortress upgrade.
2.4superseded
v2.0earlier 2026
The ten-skill modular architecture, four quantitative evaluators, four production workflows, seven venue templates.
2.0superseded