Most AI writing tools won’t tell you how they work. We will. This page is the technical companion to our weekly proof table: every component of the polish pipeline, what it does, and why we built it that way.
Last updated May 21, 2026 · model changelog · download as PDF
Your draft goes through five distinct stages. Each one has a job, an input, an output, and a measurable failure mode.
Your draft is scored against 14 signals known to mark machine-generated prose. The current set: low-perplexity bursts, hedging absence, formulaic transitions, nominal-phrase density, citation-style anomalies, sentence-length monotony, AI-typical bigrams (‘delve into’, ‘multifaceted’, ‘leverage’), absent failure-statements, missing first-person specifics, anchored to a 12,000-document reference distribution. Output: a per-paragraph signal-score vector.
For research papers: section detection (Abstract / Intro / Methods / Results / Discussion) via header parsing + a fine-tuned classifier. Citation extraction ((Author, Year), [12], \cite{key}, DOIs) into a citation table. Defined-variable extraction. Quantitative claims (numbers + units + p-values) flagged and protected.
Based on the document type and the signal-score vector, the polish engine activates a subset of 18 named strategies. Specificity, hedging, thesis-forward opening for personal statements; nominal-phrase reduction, claim-tightening, journal-fit verbs for research papers. Each strategy is a constrained rewrite rule with an explicit input/output contract, not a free-form rephrase.
Selected strategies are applied via a fine-tuned language model under three hard constraints: (a) protected entities pass through bit-perfect, (b) per-section polish budget (Methods is the most conservative), (c) voice fingerprint, the user’s sentence-length distribution, lexical diversity, and 4-gram profile from their prior documents. The polish model is Claude Sonnet 4.6 served via Vertex AI Model Garden plus a Rewritelyapp-trained adapter; no user data trains the adapter.
Output is re-scored on (a) the 14 detection signals, (b) the relevant rubric (23-criterion for statements, style-guide compliance for papers), (c) three detector models, GPTZero, ZeroGPT, internal ensemble. Any edit that pushes a metric backward is reverted automatically.
These are the constraints we won’t turn off, even if a user asks us to. They’re what makes Rewritelyapp safe for academic use.
95.8%
Avg detector pass rate across five detectors, 200-essay corpus, week of May 21.
+38
Median rubric-score gain on personal statements (out of 100), measured across our reference corpus.
0
Citations altered or claims fabricated. The placeholder pipeline rejects any polish that doesn’t round-trip bit-perfect.
| Base model | Claude Sonnet 4.6 via Vertex AI Model Garden + Rewritelyapp adapter v3.0 |
| Training data (adapter) | Public academic writing guides + 4,800 author-permissioned essays. No user-submitted documents. |
| Hosting | Vertex AI (europe-west4) + Rewritelyapp application layer (Cloud Run, europe-west1) |
| Latency (median) | 2.1s per 500 words polished |
| Known failure modes | Documents with mixed languages, equation-heavy LaTeX, and footnote-heavy humanities prose see degraded performance. We flag this at upload. |
| Last model update | v3.0 · May 14, 2026 · changelog |
The methodology is only convincing on paper. The diff is convincing on your own draft.
Try Rewritelyapp free →