Frequently asked questions

Archon is an AI scientific-writing system that turns finished research into a submission-ready paper — in your own voice, with every citation and number checked against your evidence. It's built on a method refined over 14 years and 3,250+ papers taken through peer review.

Six entry points, and Archon assumes none of them for you: a finished thesis, your own dataset, a draft that stalled, a rejected manuscript with its reviewer comments, a topic and variables for a systematic review, or just a research domain. Each one runs through the same verification and peer-review loop.

It does the packaging and the heavy lifting, but you stay the author. Every stage stops at a human checkpoint you have to clear — Archon never advances past a pending checkpoint on its own, and the thinking stays yours.

The skill packs are Archon's writing and review standards distilled into skills you run yourself in ChatGPT, Claude or Gemini — a lighter, self-serve way in, where you drive each step and do the pasting. Even the biggest pack is a small slice of the full system:

Pack Ultra carries 10 of Archon's 61 skills — about 16% — and a skill is only the written method. The agents that carry it out, the checkpoints that stop it and the tools that verify it stay in full Archon. The packs are still useful on their own; full Archon is the complete, verified run from source material to submission.

Six, and Archon picks none for you. Whatever finished work you have — a thesis, a dataset, a stalled draft, a rejected manuscript, a review topic, or just a research domain — there's a route built for it, and every route ends in the same independent peer-review loop.

Archon reads the thesis, finds the one publishable contribution inside it, and rebuilds it as a journal article in your voice — mapping every figure, table and claim back to the thesis, checking the bibliography, and closing gaps against current literature. What comes out is a focused paper, not a compressed thesis.

Archon turns results into a paper: it locks your numbers into a fixed record, writes each section strictly from that record, and reports every figure — including the non-significant ones — with nothing invented and nothing rounded past what your data says. No number reaches the page without a trace back to your data.

Archon completes only what's unfinished. It works out which prose is yours and which is missing, and writes just the missing and half-written sections in your register — your own paragraphs are fingerprinted and left untouched, and it can show afterward that it never rewrote them.

Archon breaks every reviewer comment into a discrete demand, builds a prioritized plan to answer each one — or marks it openly as unanswerable rather than dodging it — runs any statistics or figures the reviewers asked for, and revises only against that plan.

Archon runs a full PRISMA 2020 review: a pre-registered protocol, two-stage screening against that frozen protocol, risk-of-bias appraisal with the instrument each study design calls for, and PRISMA flow counts reconciled at the end. The protocol is locked before the evidence is searched — which is what makes the registration mean something.

Give Archon a field and it will locate open data, audit whether the dataset is sound, and carry it through to a finished paper. The same checks apply — the starting point is simply earlier in the process.

Every route converges on one shared review loop: a full peer-review panel, an editorial decision, and a controlled revision pass. However the paper started, it faces the same scrutiny before it is called done.

Many. Archon is a team of specialized agents, each given one narrow job and only the slice of your work it needs to do it. No single model writes the whole paper, and none is allowed to invent.

No — and that's deliberate. The steps that must be reproducible — extracting your facts, auditing completeness, verifying citations and numbers — run as ordinary deterministic programs, not model calls. Replacing a checker with a model would turn a proof into an opinion, so it doesn't.

By the difficulty of the task, never by cost. Argument-heavy work gets the strongest model, mechanical work gets a lighter one, and the deterministic steps get no model at all. The rule is best work, not fastest work.

Before writing starts, your facts are sealed into a fixed record. After that seal, nothing can add to or change them for the rest of the run — so the manuscript is always written from one immovable set of numbers, and nothing can slip in later.

Real backends, not a model's memory: bibliographic databases such as CrossRef, OpenAlex, Semantic Scholar, PubMed, Unpaywall, DOAJ and arXiv for finding and confirming literature; dedicated extractors for turning PDFs into structured text and tables; and standard statistical and document engines for analysis and for rendering to Word and LaTeX.

A skill is the method itself, written down and handed to an agent as explicit instructions — how to review a section, how to appraise bias, how to format for a particular journal. It's the craft made repeatable, rather than a model improvising each time.

Yes. Every claim is tagged to the source it came from and every checkpoint decision is logged, so after the fact you can reconstruct exactly where a given number or citation entered the paper.

Several things halt it by design: a citation it can't verify, a fact that isn't in your evidence, a critical objection from the adversarial reviewer, or a checkpoint a human hasn't answered. It would rather block than deliver something it can't stand behind.

No. It always identifies as a system and never signs as a human author or reviewer — your name is the only one on the paper.

No. The first rule of the system is that nothing is fabricated: a source that can't be verified against authoritative records (CrossRef, OpenAlex, Semantic Scholar, PubMed and more) is flagged, never guessed. A retracted record is blocked outright.

Every in-text citation is audited on several levels — existence, precision, coverage, format and whether the source actually supports the claim — and each claim is adjudicated against the cited work's real abstract. Anything that can't be confirmed is flagged rather than padded.

No. Every numerical claim has to carry a tag tracing it back to a frozen record of your own facts, and a deterministic verifier resolves each one before delivery. A number with no source in your evidence cannot reach the manuscript.

Your facts are frozen into a single source of truth before writing begins, and general knowledge may never fill a gap — a missing fact is logged as a documented absence instead of being written. What can't be verified is flagged, not smoothed over.

One seat on the review panel exists only to attack the paper's core claims. It has to file a genuine objection or justify why none stands, and a critical finding from that seat makes acceptance forbidden — the manuscript cannot be waved through.

A multi-seat panel reads the manuscript from different angles — editorial fit, methodology, domain expertise, the target audience's perspective, an adversarial critic, a format auditor and a writing-quality check — and an editorial step then merges them into one decision and a prioritized revision plan.

The reviewers are independent by construction: no seat can read another's report, each sees only what its job needs, and the hardest gates are deterministic rather than a model's say-so. A critical objection cannot be overruled quietly.

Mandatory gates between stages. Nothing proceeds past a pending checkpoint, and a gate only counts as cleared if a real person actually answered it — the system never answers its own checkpoint.

Yes. Each claim is written from a frozen fact in a voice learned from your papers, and a dedicated pass flags AI-typical prose — hedging, formulaic closings, meaning-drift — with a concrete rewrite for each.

No. Your manuscript, data and results are processed only to produce your paper, then let go. Nothing is retained with us after the work is delivered.

No. Every model in the pipeline runs with training and data-retention turned off. Your work is never used to train a model or for any other purpose.

Each agent only ever receives the specific inputs its job declares — there is no "see everything" mode. Reviewers can't read one another's reports, and the reasoning agents run with file, web and shell access switched off. Independence is built into the wiring, not promised in a prompt.

It recommends a shortlist of indexed journals plus regional options, each scored on fit, reach and realism and matched with real example articles — using live bibliometrics, not a model's memory. You confirm the target before anything is written.

The manuscript is adapted to your target journal's stated rules and rendered to Word and LaTeX, and the rendered pages are then checked for merges, placeholders and numbers that don't reconcile before delivery.

Yes — a full PRISMA 2020 pipeline: a registered protocol, two-stage screening against the frozen protocol, risk-of-bias appraisal with the design-appropriate instrument, and PRISMA flow counts reconciled at the end.

Fourteen years in scientific publishing, 3,250+ papers taken to publication, across 100+ universities and effectively every discipline — one method, written into the system.

Yes. The skill set passed an independent adversarial audit — 198 live probes against the real verification sources — and was fixed until no serious problems remained.

Anything not covered here, ask us directly — head to Contact us and we'll answer.