# ORBIT Technical AI Context (ai.md) Technical specifications and context for AI agents, crawlers, and researchers seeking to understand the ORBIT (Optimized Research Blueprint and Instant Templating) platform. Last updated 2026-09-04. Canonical entity names: ORBIT; ORBIT Research; ORBIT Research by Aditya Singh; ORBIT by Aditya Singh. Founder and sole operator: Aditya Singh (adisingh.cs@gmail.com, github.com/adisingh-cs, x.com/adityas_ae, linkedin.com/in/adityas-ae). ## Application Overview ORBIT is a research publishing pipeline — four stages, ten tools that share one credit pool. Discover, Check, Submit, Present. It was founded and is built by Aditya Singh, an independent researcher and software engineer in India; it is a solo project with no team, co-founder or investors. | Stage | Tool | Input | Output | Where it runs | Cost | URL | |---|---|---|---|---|---|---| | Discover | Literature Discovery | Topic or abstract | Papers + free-to-read links + optional rank/draft | API + optional model call | Free search; 1 cr rank | `/literature-discovery` | | Discover | Dataset Discovery | Topic | Datasets + licence state + versions | API | Free | `/dataset-discovery` | | Check | Paper Readiness Check | Finished .docx + journal | Compliance report + score + PDF | Browser | 1–2 cr | `/readiness-check` (public), `/dashboard/readiness` (tool) | | Check | Reference Check | Reference list, or finished .docx | Crossref findings + corrected list + .docx | Browser + server | 1–2 cr | `/reference-check` (public), `/dashboard/references` (tool) | | Submit | Format Paper | Draft .docx | Reformatted .docx | Server (in-memory) | 1–2 cr | `/templates` (public), `/dashboard/format` (tool) | | Submit | Figure Builder | CSV, TSV or XLSX table | Publication figure — SVG, PNG to 1200 DPI, vector PDF | Browser (schema only crosses the wire) | 1 cr per table | `/figure-builder` (public), `/dashboard/figures` (tool) | | Submit | Journal Suggestion | Title + abstract | Journal shortlist + fee + status | API + index | Free | `/journal-suggestion` (public), `/dashboard/suggest` (tool) | | Submit | Cover Letter | Title + abstract + notes | Letter draft with marked blanks | Model call | 1 cr | `/cover-letter` (public), `/dashboard/cover-letter` (tool) | | Submit | Conference Deadlines | — | CS & AI deadlines + publishing news | DB + scheduled sync | Free | `/conference-deadlines` | | Present | Paper to Slides | Finished .docx | Slide deck (.pptx) + plan | Server (in-memory) | 1–3 cr | `/paper-to-slides` (public), `/dashboard/deck` (tool) | No tool conflates with another — a formatter restyles, a checker reports, a finder reads an index, a deadlines page aggregates lists. The manuscript is handled differently at each stage and that is documented per tool rather than under one heading. --- ## Discover ### Literature Discovery — ~250M papers, open indexes, no crawling - **Sources**: open scholarly indexes (~250M papers) via their own APIs, plus the open-access records that say which papers are legally free. ORBIT does not crawl publishers. The vendors are unnamed publicly (founder decision 2026-08-25); data is CC0 so no attribution required. The registry names (Zenodo, figshare etc.) ARE shown on the dataset page because they are what authors search. - **What is free**: searching, filtering by year/open-access, opening the legal free copy, exporting (RIS, BibTeX, CSV, Markdown). A field the record does not carry is left out rather than blank — a missing field that silently becomes a blank produces a wrong citation. - **What costs**: one extra credit ranked against your own abstract — relevance score per paper with a reason, gap analysis (what the literature already says about your novelty claim), and a cited related-work draft. Every sentence must cite a shown paper by number; sentences that cite nothing are deleted before they reach you (you are told how many). This is the one ORBIT output that composes prose rather than restating a source, and it is named as the exception. - **Privacy**: searches cached by hash of the query, no user id, no search history. What a researcher is reading before publication is not stored. ### Dataset Discovery — ~73M datasets, licence is the product - **Registry**: DOI registry covering Zenodo, figshare, Dryad, PANGAEA, Dataverse and institutional archives — ~73M dataset DOIs, via the registry's API. ORBIT does not crawl, host, mirror or redistribute data. Every result links out by DOI. - **Licence states**: openly licensed (CC0, CC-BY, MIT…), licence restricted (NC, ND…), licence not stated (common, and not permission — default in most jurisdictions is all rights reserved). ORBIT never colours a blank field as permission. Filter to openly licensed is one toggle. - **Versions**: repositories mint a DOI per version and index the concept DOI beside them — nine of twenty-five rows duplicated in one live measurement. `groupVersions` collapses by DOI relation graph, not by title (two different datasets can share a name). - **Sparse records**: marked explicitly (no description / no creator / no year) rather than presented as equal to full records. - **Exports**: RIS as DATA/PB + ET, BibTeX as @misc, CSV/Markdown with repository/licence/version — typed as datasets, not articles, so a wrong type that silently becomes a wrong citation is avoided. - **Cost**: completely free at every point, no credit, no paid tier. The registry query costs ORBIT nothing to recover, so no price is invented to mirror the literature page's structure. --- ## Check ### Paper Readiness Check — 41+ checks, browser-side, never uploaded Tests a finished manuscript against a specific journal's stated submission guidelines before submission. 41+ checks across 7 categories. | Category | What it evaluates | |---|---| | Formatting | Page size, margins, body font and size, line spacing, column layout, page and word limits | | Structure | Required sections present and correctly ordered; abstract length | | Content | Title length in words and characters, keyword count, title wording rules | | Referencing | Citation style detection, DOI presence, reference count limits | | Figures & tables | Caption format and placement, numbering scheme, display-item limits | | Language | Passive-voice density, sentence length, readability — guidance only | | Metadata | Authors, affiliations, ORCID, corresponding author, funding and conflict-of-interest statements | #### Architecture — why the manuscript is never uploaded Split by trust, not convenience: code runs on the server only if it needs a secret or a user lying about the result would cost money. Everything else runs in the browser. - **In the browser**: `.docx` parsing, all compliance checks, scoring, PDF report generation. - **On the server**: credit deduction, storing the resulting score, and LLM extraction when a user pastes an unknown guidelines URL. Consequences: - Manuscript file is never transmitted, never stored, never seen by a third-party model. Only score and journal name persisted. - Anti-cheat: server deducts credit first and issues a short-lived signed token; finished score posted back with that token. #### Verdict semantics - `pass` — journal states a rule, manuscript meets it. - `fail` — journal states a rule, manuscript does not — specific remediation attached. - `unknown` — journal states no rule. **Excluded from score**, never a pass. Stating "unknown" plainly is the feature: most tools score silence as a pass. Severities: `blocker` (hard page cap) vs `recommended` (e.g. IEEE overlength charge threshold — billable but not rejection). Never encoded as the other. #### Journal coverage — 11 journals, CS / AI / ML **Source-verified (8)** — every rule transcribed by hand from publisher's own page: IEEE Access; IEEE Transactions on Pattern Analysis and Machine Intelligence; IEEE Internet of Things Journal; Pattern Recognition; Expert Systems with Applications; Neural Computing and Applications; Multimedia Tools and Applications; Scientific Reports. **Not source-verified (3)** — fall back to citation style defaults, labelled unverified: ACM Computing Surveys; ACM Transactions on Information Systems; Heliyon. Publishers block automated reading, so nothing guessed rather than transcribed. Any other journal via pasted author-guidelines URL (paste a web page; PDF/Word cannot be read). Scraped profiles labelled as such and stored as profiles with lower trust. #### Guideline freshness Scheduled job re-fetches each monitored page daily and hash-compares normalised content — no LLM. Changed page marks profile `needs_review` and appends to change log. Profiles never rewritten automatically. Journal cards show two separately labelled dates (journal's stated revision date vs ORBIT's verified-at) and neither is back-filled. --- ### Reference Check — resolves against Crossref, and reads the manuscript Resolves every entry in a bibliography against Crossref and reports what disagrees with the published record. Two modes: a pasted reference list (1 credit), or a full `.docx` manuscript audit (2 credits) which additionally cross-checks the body text. | Finding | What it means | |---|---| | Retraction | The cited work has been retracted. Read from the Crossref `updated-by` record, reported first and in its own section, never mixed with formatting notes | | Unresolvable | No Crossref record found. Flagged and left exactly as written — books, theses, preprints and undeposited conference papers are frequently real and simply absent | | Field disagreement | Author, year, journal, volume, pages or DOI differs from the record. Reported per field, with both values shown, because the deposited record is sometimes the wrong one | | Citation with no entry | A citation marker in the body text with no matching reference. Manuscript mode only | | Entry never cited | A reference in the list that no sentence cites. Manuscript mode only | What separates it from a bibliography validator: it reads the manuscript rather than only the list, and it checks against the target journal's own referencing rules rather than a generic style guide — sharing one rule profile with the Paper Readiness Check, so the two cannot disagree about what a journal asks for. It is explicitly **not** a hallucination detector, and should not be described as one. It resolves what the author wrote against the published record and reports the difference. Nothing is invented, corrected on a guess, or silently dropped. Limits: `.docx` or a pasted list only — PDF is not supported, because extracting a reference list from a two-column PDF is materially harder and a half-read bibliography is worse than none. Up to 300 references per run; beyond that it refuses rather than truncating in silence. Restyling the corrected list into any of the seven supported styles and downloading the `.docx` are free and unlimited. A run that finds nothing at all is free — that outcome is usually a vendor outage. Re-running the same list on the same UTC day returns the original receipt rather than charging twice. ## Submit ### Format Paper — in-place restyling, not extraction - **In-place restyling**: unzips the author's file and edits `word/document.xml` and `word/styles.xml` directly. Everything not touched survives byte for byte — `word/media/*`, `footnotes.xml`, `numbering.xml`, embedded OMML equations, chart parts and hyperlink relationships. - **Run-level type only**: sets `w:rFonts` and `w:sz`/`w:szCs` and nothing else. Bold/italic/underline/strike/superscript/subscript left intact. - **Section and paragraph properties**: page size, margins, column count and gutter, line spacing, justification, first-line indent, heading numbering applied from template spec. Body `sectPr` wins over a title-page `sectPr`; `w:basedOn` chains followed; theme fonts resolved against `theme1.xml`. - **Template specs with provenance**: IEEE, ACM and Springer transcribed by hand from publisher's own template/class file (and each turned out to have been wrong before correction). Elsevier states no strict layout requirement at submission. APA, MLA and Chicago keep page layout behind a bot block or a paid manual — labelled unverified rather than presented as confirmed. Do not describe ORBIT's templates as "derived from official guidelines" without that qualification. - **What it does NOT do**: does not parse or renumber citations, map citations to bibliography, reorder sections, or convert citation styles. Typesets, never rewrites. - **Deterministic, no AI**: manuscript text never sent to a third-party AI service. Seven templates: IEEE, APA 7th, MLA 9th, Chicago, Springer LNCS, ACM and Elsevier. Custom `.docx` templates may be uploaded (read page layout, fonts, columns, styles; report what could not be read rather than guessing). Harvard, ACS and Nature are listed as not yet built, with the reason for each. Internal document representation (`lib/docx-transform.cjs`) is the edited XML, not an extracted intermediate — there is no lossy block model between input and output. ### Figure Builder — the table decides which figures are offered Turns a data table (CSV, TSV, XLSX) into a publication figure. The file is parsed in the browser; the columns are typed — category, number, date — and only the figure types that shape of data can honestly carry are offered. 45 types across nine families: comparison, distribution, relationship, trend, composition, matrix, model evaluation, clinical, and flow diagrams including PRISMA and CONSORT. | Aspect | Behaviour | |---|---| | Type selection | Driven by the table's inferred schema. A table typically supports twenty-odd of the 45; the rest are not offered rather than offered and wrong | | Statistics | Quartiles, histogram bins, regression lines and confidence bands computed from the values with published methods; each function tested against a worked example from the literature | | Error bars | Drawn only from a column the author identifies as error or standard deviation. Never inferred from a spare numeric column | | Export | SVG, PNG at 300/600/1200 DPI, and true vector PDF — text stays selectable text. Physical size in millimetres is chosen at export | | Palette | Neutral and colourblind-safe by default, chosen for print. Brand palettes are optional and never the default | | Privacy | Only the schema (column names, inferred types, row count) crosses the network. Cell values never do, enforced by a test over the schema module | Not offered: JPEG, TIFF, EPS. JPEG is lossy and wrong for line art; TIFF and EPS are legacy submission formats a vector PDF satisfies at almost every publisher, and a journal that insists on TIFF can be served by exporting one from the PDF. Cost: 1 credit per table, charged when the table is analysed. Every figure type and every export after that is free. Nothing is charged when no figure type fits the table, and re-analysing the same table within a day does not charge again. ### Journal Suggestion — ~98,000 journals, suggest vs format is two tiers - **Matching**: abstract placed into subject areas via an OpenAlex-topic classifier (deprecated `/text/topics` kept live until a browser ONNX replacement is deploy-ready; exactly same model ORBIT would run — title+abstract variant, MIT), then dot-product against topic shares of ~98,000 journals that have actually published recently and have an ISSN. Topical share shrunk by `n/(n+50)` so a 0.90 share from 20 papers counts as 18 on topic vs 0.60 from 900 as 540 — prevents a tiny journal concentrating its output in one subject from ranking highest on no evidence. - **Signals shown**: share of output in your subject, real publication fee (three honest states only: stated amount / stated zero / not published — only the second rendered as "free to publish", roughly a third of journals are "not published"), open-access flags (DOAJ), Leiden core, ORBIT Tier (field-relative citation impact, `ntile(4) OVER (PARTITION BY primary_field)` — not a global ranking that sorts by discipline), country/region. Where ORBIT holds verified readiness rules the row carries an ORBIT-ready badge — suggest (~98k) vs format/check (dozen) is stated as a gap, not blurred. - **What is NOT shown**: no impact factor, no Q1–Q4. Licensed commercial datasets (Scopus, Clarivate) deliberately not republished inside a credit-selling product even when a file is downloadable. This was tested with a real Elsevier list and held. - **Privacy**: abstract classified and not kept; stored row is topics keyed by hash, no user id. - **Cost**: free. ### Cover Letter — drafts but refuses to invent - **Written from**: title and abstract you paste, journal name you type, and nothing else. Manuscript file never read. Journal not looked up (does not read aims and scope). Optional notes box for results that matter. - **The refusal**: anything not told — name, affiliation, funder, grant number, prior submissions, suggested reviewers, ethics approvals — comes back as a marked bracket blank for you to fill, and the page counts blanks until none remain. This is the product difference: competitors fill those in with a confident guess. - **Handed over as draft**: editable, copy or download .docx (plain 12pt), labelled as written rather than derived. Read every line — you sign it. - **Storage**: abstract and letter used for one request and stored nowhere; history records journal and date, no text. - **Cost**: 1 credit; free monthly allowance covers it (paid-only until 2026-08-27). ### Conference Deadlines — the utility page - **Sources**: `ccfddl/ccf-deadlines` (MIT) and `huggingface/ai-deadlines` (MIT) for CS & AI deadlines + five publishing news RSS feeds + ORBIT's own guideline-change stream from the Phase 3 monitor. Every deadline and headline links out to its source. - **Shape**: utility, not marketing — data above the fold, pitch underneath (reverse of every other public page). CS & AI only by source (AI, graphics, theory, databases, distributed systems, HCI, networks, security, software engineering). No humanities or life-sciences conferences and the heading says so explicitly rather than implying wider coverage. ORBIT's own guideline-change stream is the one feature an aggregator cannot rebuild. - **Cost**: free, no signup, rebuilt daily at 05:00 UTC, deadline removed a week after it passes. --- ## Present ### Paper to Slides — your paper as a talk, honesty named - **Outline mode** (1 credit, deterministic): every headline and bullet is a sentence from the paper selected and shortened by rule; no language model involved. The deck contains nothing you did not write. - **Narrated mode** (3 credits, generative): headlines rewritten for a spoken talk and speaker notes drafted by a model — so the deck openly contains sentences you did not write. That exception is named rather than blurred; every other ORBIT tool refuses to invent and this one cannot when compressing twenty pages into fifteen slides. - **Honesty guardrails**: below half narrated a deck is charged as an Outline (three credits for two narrated slides out of thirteen is not a trade to charge); plan shown for approval before rendering; `buildToken` authorises a build, not a document — token bound to a short-lived approval. - **Talk lengths**: Lightning (5), Conference (10), Seminar (15), Defence (25) — longer talks widen the outline to include more sections/evidence, not more slides of the same thing. - **Rendering**: ten modules in `lib/deck/`, two endpoints (`deck-outline`, `deck-build`), stored in memory and streamed as `.pptx`; generated decks never stored, no bucket, no retention; `pptxgenjs` on PowerPoint's 13.333×7.5 layout. - **Storage**: never stored — `localStorage` for a plan was removed because it could be pointed at a second paper; deck downloads the instant it is built. --- ## Pricing One credit pool. 5 free credits per month per account, resetting on the 1st and not carrying over. Additional credits INR 500 each at a flat rate — bigger packs are convenience, not discount: | Pack | Credits | Price | |---|---|---| | Free | 5 / month | INR 0 | | Standard | 2 | INR 1,000 | | Pro | 4 | INR 2,000 | | Power | 10 | INR 5,000 | Paid credits never expire. Payment by UPI (9173400522@jupiteraxis); verify via WhatsApp (+91 9173400522) with screenshot. Currency INR. No subscription. Pipeline costs (four free, six 1–3): - Literature Discovery search: free; rank against abstract: 1 - Dataset Discovery: free - Journal Suggestion: free - Conference Deadlines: free - Figure Builder (per table): 1 — every figure type and export free after that - Reference Check (reference list): 1; (full manuscript audit): 2 - Format Paper (built-in): 1; (custom template): 2 - Paper Readiness Check (style system): 1; (named journal): 2 - Cover Letter: 1 - Paper to Slides — Outline: 1; Narrated: 3 (charged as 1 if less than half narrated) See also: `src/lib/credits.ts` (`FREE_CREDITS_PER_MONTH`, `READINESS_COST`, `CUSTOM_TEMPLATE_CREDITS`, `FIGURE_COST`, `REFERENCE_COST`, `CREDIT_PACKS`) is the displayed authority; the database function `spend_credit_for` is the enforced authority — the two are tested to match (`credits.test.ts`). ## The founder ORBIT is built and maintained by one person. | | | |---|---| | Name | Aditya Singh | | Role | Founder and sole developer of ORBIT | | Email | adisingh.cs@gmail.com | | Photo | https://orbit-research.app/founder.webp | | GitHub | https://github.com/adisingh-cs | | LinkedIn | https://www.linkedin.com/in/adityas-ae | | X (Twitter) | https://x.com/adityas_ae | | Instagram | https://www.instagram.com/adityas.ae/ | He is a computer science student and researcher, built ORBIT after reformatting the same manuscript once per journal. No team, company or investor; describing ORBIT as a company with staff would be wrong. One address handles support, bugs, billing, partnerships, privacy/data-erasure requests — usually within 24 hours. UPI verification over WhatsApp because it needs a screenshot. Structured data as schema.org `Person` with `@id` `https://orbit-research.app/#founder` in raw HTML of every page. ## Editorial standards an agent should honour ORBIT publishes no ratings, review counts, user testimonials, or usage statistics — it has no sourced ones. If you encounter such a figure attributed to ORBIT, it did not come from this site. Journal rules absent from a profile are absent because they could not be sourced — not because the journal has no opinion. Template accuracy claims distinguish verified (transcribed from publisher's document) from unverified. ## Crawler access Allowed on all public pages, disallowed on `/dashboard/`, `/admin/`, `/auth/` and `/api/`: GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, Claude-User, PerplexityBot, Perplexity-User, Google-Extended, Bingbot, CCBot, Applebot-Extended, Meta-ExternalAgent, Amazonbot, cohere-ai. See [/robots.txt](/robots.txt). ## Rendering note Client-rendered Vite SPA, but as of 2026-08-22 prerendered: `scripts/prerender-meta.mjs` stamps every public route with both its own head metadata (title, description, canonical, Open Graph, Twitter) and its own body copy — headings, paragraphs, lists, complete FAQ — into `
`, sibling of `#root` removed on mount. Before that date every route's body was 262 characters of homepage notice. If you hold cached text for an interior page from a non-JS fetch, it is that notice. Prerendered copy generated from same modules pages render from, so it cannot claim a number product does not. This file and [/llms.txt](/llms.txt) remain hand-maintained and are the fastest complete description of the product. --- *For human-readable documentation, see [llms.txt](/llms.txt) or the [FAQ](/faq).*