Detecting machine-written work is uncertain by nature. A tidy sentence is not proof. A detector score is not a verdict. So our process is built to slow down, gather evidence from several angles, and resist the pull of any one alarming signal.
What follows is the standard we hold ourselves to. It exists as much to protect writers from a wrong call as it does to catch genuine deception.
One signal is never enough
Flat prose, a short editing history, a high detector score — any of these alone can have an innocent explanation. We act only when independent signals converge on the same conclusion.
The detector is a last resort
We reach for Pangram only when human review has already produced strong suspicion. It confirms or complicates a case we've built — it never opens one.
High confidence, with the evidence
Nothing is passed to other editors on a hunch. When we do share, the full evidence goes with it, so each editor can weigh it and reach their own judgment.
Each stage has to earn the next. We begin with the least invasive, most human judgment and only escalate as suspicion holds up. Most reviews end early — and quietly.
The sniff test
Before any tool, an editor reads the piece as a reader. We note what the writing itself suggests: semantic flatness where nothing is at stake, uniform structure, hollow specificity, leftover prompt scaffolding, an absence of the mess and idiosyncrasy that mark a real voice. This is a first impression, recorded honestly — not a conclusion.
Every case starts hereWriter history and track record
The strongest evidence lives here, so it carries the most weight. We look at the editing history of the submission — whether it grew through real revision or arrived as a single paste — and we compare it against the writer's known work: voice, tics, recurring flaws, range. We also bring their past stories into the investigation, asking not only whether the pieces sound alike, but whether earlier work shows machine signals of its own. A pattern across the record is itself a signal.
Weighted heaviestPangram — only on strong suspicion
If, and only if, the read and the history have produced strong suspicion, we run Pangram alongside a plagiarism scan and a metadata pull. These are corroborating instruments. They can strengthen or unsettle a case already built by human judgment; they are never the reason a case exists, and never sufficient on their own.
Gated behind 01 & 02The writer's chance to answer
When the earlier stages converge on concern, we talk to the writer. This is their right, not a formality. We ask about choices made, material cut, the discard pile behind the finished piece. A genuine author can walk you through the work. The conversation can clear a case outright, and it outweighs a weak automated flag.
Can clear or confirmPangram will not be run on a first pass. It is used only after the initial review has established strong suspicion through reading and writer history. A detector score, high or low, is treated as one weak and fallible signal among several — never as proof, and never as grounds to flag a writer on its own.
An investigation that ends in uncertainty ends there. We do not circulate suspicions, scores, or half-formed cases. Sharing happens under three firm conditions:
- Only high-confidence cases are shared.The signals must converge across human judgment and the writer's history, corroborated rather than assumed. A single flag, however striking, does not clear this bar.
- Sharing always includes the evidence.When a case does go to other editors, the complete record travels with it — the read, the history and track-record findings, any tool outputs, the writer's statement, and the interview transcript.
- Each editor decides for themselves. We pass evidence, not conclusions to be rubber-stamped. Every editor reviews the record and reaches their own informed judgment on it.
We don't believe in blacklists. We do believe that when there is credible, vetted evidence of a bad actor, publications deserve visibility and transparency — the full record, not a verdict — so they can decide for themselves. Our aim is to inform that judgment, never to replace it or derail a writer's prospects.
Access to our records comes with an obligation. Publications that receive them vow not to treat them as a blacklist, and to review each case in full before acting — weighing the evidence themselves rather than reacting to a name.
Transparency only protects writers if everyone who holds the record honors it.
Membership can be revoked, but never by one hand. Oversight rests with a rotating board of three writer representatives and three publication representatives, serving staggered one-year terms so the board never turns over all at once and no single cohort holds it. Their names are public.
Where there's credible evidence that a publication has used our records as a blacklist — acting on a name without reviewing the case — the board reviews the matter with the same care we bring to any other: no single complaint is treated as proof, the evidence is examined in full, and the publication is given the chance to respond before any decision is made.
Should the board deadlock, it does not decide in haste. A tied vote is held for one week and then cast again, anonymously, so each member is free to reconsider on the merits alone. Only if the board remains tied after that second vote does a founding member decide. The founder's hand is the last resort, never the first.
Access is suspended. A single lapse is not treated as a pattern.
The same remedy again. The graduated path is deliberate, not lenient.
Three violations end membership. The goal is to protect writers, not to punish publications acting in good faith.
One packet per submission — filed privately with the rubric, evidence, attachments, writer statement, and interview transcript. This is that packet, blank. Fill it in if you like; it is here so you can see exactly what a reviewer is asked for at every stage.
0Intake
Log before evaluating, so later stages don't bias the read. Work from a copy; preserve the source file untouched.
1Sniff test — blind read
Read as a reader first, before any tool. Mark each signal, then log specifics below. The visceral read matters and is hard to fake.
Areas suspected and proof
| Area suspected | Example / quote / proof | Strength |
|---|---|---|
2Writer history and corpus
The strongest evidence. Weight it heavily. Pull the writer's past stories into the investigation — not only whether the pieces sound alike, but whether earlier work itself shows AI signals. A pattern across the track record is itself a signal.
Editing forensics for this submission
Corpus comparison against prior verified work
Track-record review — prior stories brought into the investigation
List each past piece examined, and whether it reads as human, uncertain, or shows AI signals of its own.
| Prior piece (title / date) | Reads as | Notes and proof |
|---|---|---|
3Automated detection
Run only after human judgment is recorded, so tools corroborate rather than lead. All three are weak, false-positive-prone signals — they never convict on their own.
Pangram (AI detector)
Plagiarism scan
Catches copying, not generation. Record matches and sources.
Metadata pull
Application tag, author / last-modified-by, revision count, editing time, timestamp mismatches.
4Interview and defense
Triggered when stages one through three converge on concern. The writer's right, not a formality — it can clear or confirm, and overrides a weak automated flag. Full transcript attached to this packet.
Questions covered
5Scoring and disposition
Require convergence, not any single signal. Writer history and corpus weigh heaviest; the sniff test is significant; detector and plagiarism are supporting only; the interview can clear or confirm.
6Evidence packet — attached
Attach to every review, cleared or not. This packet is filed privately with all of the below.
We never flag on a detector score alone. Detection tools carry false positives that fall hardest on non-native English writers and neurodivergent authors, and we treat that risk as our own to manage.
Presumption of innocence runs through the whole process, alongside a documented path for any writer to appeal a finding.