Materials, thesis, and synthesis

Draft Coach is a local writing desk. It does not humanize AI prose. It helps you keep character, a thesis, and real synthesis by separating sourced materials from joins. Analyze runs a local heuristic in your browser. Optionally, Generate can draft join sentences through your OpenAI-compatible LLM key, and Check can call Pangram 4 through your Pangram API key — showing real API fields, not a fake Human%.

Last updated 19 August 2026

1. What Pangram 4 is

Pangram 4 is Pangram Labs’ text detector, released 29 July 2026. The technical report is arXiv:2607.27183. Draft Coach is not that model. Without a Pangram API key, Analyze labels are a local heuristic only. With a key, Check calls model: pangram-4 through a same-origin proxy and displays the real headline, Mixed/Human/AI verdict, window labels, and humanizer flags — it does not imitate or optimize scores.

Segment labels are Human, AI-Assisted, or AI-Generated. At document level the short verdict is Human if at least 90% of characters are Human, AI if at least 80% are AI-Generated, and Mixed otherwise.

In their formulation, Human includes negligible AI — light copy-edit, spelling, grammar, literal translation. AI-Assisted is homogeneous co-authorship: a human idea whose final words were chosen by a model. AI-Generated is original open-ended prose whose wording is the model’s.

They define a humanizer as deliberate obfuscation meant to evade detection — synonym spinning, injected typos, casing tricks, commercial “make it human” tools. On 13 commercial humanizers they report about 98.83% still landing as AI or Mixed. The humanizer head is a side probe. It does not vote on the main Human / AI label.

Inference uses overlapping windows of register (512 tokens in the model card). Neighbours, paste versus PDF extract, and whether a sentence sits in fluent chronology or among quotes can move the same words. That is a property of the detector, not a writing method.

2. Empirical writing lesson

Long-form tests of register, not a recipe: cutting narrator wrap-ups raised the Human share. Putting fluent thesis-closing synthesis back collapsed it. Isolated fluent chronology scored 100% AI. The same facts, interrupted by quotes and captions, often did not.

That is a window-of-register observation. It is not an instruction to strip the argument until a gauge goes green.

3. What the coach flags

These are local pattern flags. They are not Pangram labels. Identical sentences can flip when neighbours change, or when the same text is pasted from a PDF instead of typed.

A keeper is material. A join is work. A suspect is often the model-shaped sentence that does the joining for you — chronology that already knows how the paragraph ends. Quoted or sourced sentences stay keepers even if they contain a wrap-up phrase: lock other people’s words; do not rewrite them to feed a flag.

4. How to use

1. Type character, thesis, and synthesis first — before the draft can argue on your behalf.
2. Paste the draft.
3. Analyze (local heuristic).
4. Lock keepers (quotes, captions, numbers you did not invent).
5. For each suspect or join: type a replacement in your voice, use Generate with your LLM key, or cut the job. Do not paraphrase the same ledger-closing job.
6. Optionally Check with your Pangram API key to see real Pangram 4 labels.
7. Export or print.

Replacing “in conclusion, the evidence clearly shows…” with a synonym of the same closing is still the same job. Either say what you take the materials to mean, or stop.

5. What this tool will not do

• No humanizer — no typo injection, synonym spin, or regenerate-until-green.
• No fake Human percentage, confidence bar, or colour that pretends to be Pangram.
• No optimization for a Human% target or Deft-style Distribution Fine-Tuning.
• No sending your API keys to Anvil — keys stay in localStorage on your device.

Generate sends draft text to your LLM vendor. Check sends text to Pangram Labs. Analyze does not leave the browser. If you need provenance labels, use Check and read Mixed as the honest co-authorship bucket — not a scoreboard to beat.

6. Commercial Human% labs

On 18 August 2026, Deft launched as a writing product built around Pangram’s detector. Their public announcement (18 Aug 2026) cited “86% of user queries” scoring fully Human on Pangram — and listed tips such as adding detail to prompts, using style controls, and noting that short sentence prompts get detected as AI. That last point is a window-of-register observation: bare, generic sentences look model-shaped because they carry no locked particulars. It matches what Draft Coach is for — lock quotes and facts, write thesis and joins yourself — not a guide to gaming a Human% figure.

Products in that category optimize for Pangram Human%. Pangram 4’s humanizer probe exists largely because of that market — synonym spin, typo injection, commercial “make it human” restyling. That is not this tool’s goal. Draft Coach will not pad your prompt, restyle prose for a detector, or show a fake Human percentage. If you need a provenance score, use Pangram 4 and read Mixed as the honest co-authorship bucket. We do not publish a recipe to hit 86%.

What we copied from Deft, and what we did not

Deft (18 August 2026) uses proprietary Distribution Fine-Tuning (DFT) on Pangram’s detector. We cannot replicate DFT. Draft Coach copied only the surface workflow: Generate/Rewrite, style presets (Analysis, Essay, Creative, Rewrite), and longer specific prompts that include particulars from your locked materials. We did not copy “optimize for 86% Human”, their model, or any detector-targeted training. Generating joins with any chat model is still AI-Assisted or AI-Generated in Pangram’s own labels — even if a later Check returns Mixed or Human.

7. Questions