The decimal point has a sales job now. Four places after it and suddenly the procurement committee hears the voice of God. Somebody get that number a lanyard.
On 16 October 2026, Pangram published a new accuracy explainer. It reports an internal false-positive rate of 0.0041% and a headline false-negative rate of 0.3396% for Pangram 4. These are the vendor's benchmark results. The post explains the existing detector; it does not announce Pangram 5, a new TextVeil trial, or the discovery of a machine that reads your diary backwards through time.
Ask what the percentage divides
A false-positive rate starts with documents known to be human. It asks how many received the wrong label. The question a flagged writer wants answered starts somewhere else: given this flag, how likely is it wrong?
Those are different conditional questions. You cannot turn one into the other by copying the number into a disciplinary email and adding a stern salutation. To make that conversion, you also need the mixture of human and generated documents being checked, the detector's behavior on both, and evidence that the benchmark resembles that setting.
Here is the useful thought experiment: keep a detector fixed, then change the pile of documents it sees. A pile containing almost no generated writing and a pile dominated by generated writing can produce very different mixtures of correct and incorrect flags. The machine did not change. The denominator did. This is basic counting, and apparently it needs a bouncer.
It does not follow that Pangram is useless. It follows that an administrator cannot staple a benchmark to an individual accusation and call the staple peer review.
Read the thing the API actually returns
Pangram's model card distinguishes its fields. Document fractions describe the share of characters assigned to each class. Its continuous AI-involvement score is not the probability of the displayed label. Its local confidence category is not a calibrated probability estimate. The humanizer flag is another output.
That matters before anybody writes a sentence beginning with "the detector says you are 90 percent robot." A fraction of classified characters is not a percentage of your soul repossessed by Anthropic.
If a dashboard compresses several fields into one triumphant badge, ask what disappeared. Ask which field drives the badge, which threshold produced it, and which model version supplied the result. A prettier interface can still throw away the part you needed to understand.
The footnotes belong beside the sales pitch
The October explainer says Pangram targets prose of at least 50 words, mostly complete sentences. It lists code, markup, translations, reference lists, and highly formatted material among its less reliable or out-of-scope cases. It recommends checking PDF parsing and using raw text or DOCX where available. Read those limitations here.
So keep the material actually sent for a scan. If an extraction mangled the paragraph, a screenshot of the original PDF does not explain what the classifier received. If you clipped a passage, keep that passage too. Otherwise the supposed evidence turns into a pub argument where everybody remembers a different sentence.
The commercial incentive deserves the same daylight. Menlo advertises a $100 million Anthology Fund with Anthropic. The application layer has capital behind it. Our opinion: a market that profits from machine writing does not get to make the writer's questions look unreasonable. Ask them. Ask with your shoes on the desk if necessary.
Keep your voice. Keep the receipt.
TextVeil's job is to humanize the draft and strip the machine residue. Put the server-room cadence through the mosh pit. Use Restyler to give the writing an actual attitude. Keep the original, read the edit, and scan with your own Pangram key when you want that feedback. Pangram bills those scans separately.
This dispatch ran no detector experiment and announces no new pass rate. We are not borrowing Pangram's numbers, flipping them upside down, and selling them back as a TextVeil victory banner. That particular bullshit already has enough distribution.
Make the paragraph sound awake. Make the number explain itself. Neither job requires surrendering your em dash to a man holding a dashboard.
Bring the draft to TextVeil · Our earlier watermark dispatch