Spaces:
Paused
Paused
Commit History
Yesterday
-
Theses over 40,000 words reach 0%: the fast engine's limits grow with the length (time cap 30 s per 1,000 words, at least 10 min; a round waits 0.08 s per flagged sentence, at least 90 s; 2 extra rounds per 10,000 words); a wallet run's deadline follows the same cap. A flat 10-minute cap and 90-second rounds dropped most flagged sentences of a long thesis each round
unknown +1: Claude Opus 5.5Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>c7ef1c5
6 days ago
-
Faster long documents: Smart Humanize uses the fast engine at any length (FAST_MAX_WORDS default 0; a 22,000-word part took 13 min on the sentence-by-sentence engine, the fast engine reached 0% in about 4); report-guided runs at 0% stop when a round rewords fewer than 3 (or 5%) of the highlighted sentences left (5 of 12 minutes went on taking 80 to 72); wallet usual cost of long fast runs from live measurements; pricing table rows all on the fast engine
unknown +1: Claude Opus 5.5Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>d14dbe0 -
Sentences with citation fields (Zotero / Mendeley / EndNote), links or superscript citations inside are now rewritten around them: the language model sees markers ([REF1] ...), a version is used only if it keeps every marker once and in order, and its words go into the plain text around the untouched fields (citations stay live); choosing, the faithfulness check and the report read the real text. In a thesis a third of the highlighted sentences were locked before (FAST_AROUND_FIELDS=0 = old behaviour)
unknown +1: Claude Opus 5.5Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>343161f -
Report-guided runs: highlighted sentences reworded round by round (more new wording each round) until little original wording is left (AI report 35%, similarity 15%); AI-report estimate from how our detector reads those sentences in the finished document; report table records each rewrite; stop after one round without progress at 0%
unknown +1: Claude Opus 5.5Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>c740ef2 -
Report-guided Smart Humanize through the fast engine: the sentences an uploaded checker report highlights are rewritten by the language model many at once (was 25 min for a 22,000-word part), then what our detector flags until 0%; report table kept
unknown +1: Claude Opus 5.5Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>5ef9b97
11 days ago
-
Pay as you go only: prepaid wallet with PhonePe top-ups; every run reserved before it starts, capped (Groq + deadline) and charged its actual price with a 300% margin after GST and fees; admin profit report
unknown +1: Claude Opus 5.5Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>89d4f87 -
Fast engine: never uses a rewrite the faithfulness check did not approve; spending cap scaled to length; stops at 0%
unknown +1: Claude Opus 5.5- A version is used only if the faithfulness check said it keeps the meaning. When the check could not run (the flat $1 job cap stopped it in a 5,500-word test), sentences keep their wording instead of taking unchecked text. - groqai.set_cap: per-job spending cap; the fast engine allows $0.60 per 1,000 words (at least GROQ_JOB_BUDGET). - Once the document is at 0%, one more round at most (for leftover stock phrases). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>f66d9ce -
0% first: rounds continue until the document is at 0% (10-minute safety cap); documents not at 0% are never charged
unknown +1: Claude Opus 5.5- fast.py: no one-minute stop any more: rounds go on until 0% on our detector, until 3 rounds bring no improvement, or the FAST_BUDGET cap (600 s); later rounds give stubborn sentences more versions and the strongest rewrite. The progress line shows the current score every round. - app.py: a document that did not reach 0% is delivered with the remaining sentences highlighted, the customer is told, and it is not charged (no pay-per-use payment, no plan credit). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>45c0ee6 -
Groq only: the fast engine rewrites with four Groq models; no other provider is ever called
unknown +1: Claude Opus 5.5- fast.py: default generators qwen3.8-27b, qwen3.6-27b, gpt-oss-20b, gpt-oss-120b (250k tokens/min each on the paid key); 24 sentences per request (the instructions were most of the tokens sent). - groqai.py: GROQ_ONLY (default on) - the OpenAI fallback of the proofread is off even if an OpenAI key is set. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>3089e27 -
Fast engine: four rewriting models (each its own rate limit), 2 versions per model in round 1, smaller requests, rounds use the remaining time
unknown +1: Claude Opus 5.5Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>fd0351c -
One-minute deadline for the fast engine; half-precision detector on the GPU when it gives the same verdicts
unknown +1: Claude Opus 5.5- fast.py: the whole job ends within FAST_BUDGET (52 s) whatever the length: language-model requests still running at the deadline are dropped (their sentences keep their best version), rounds are planned from measured times, and long documents' candidates are pre-ranked on the GPU too. - detector.py: half precision on the GPU (about 2x faster), switched on at start-up only if it changes no verdict and no score by more than 0.05 on the bundled samples; logged either way. DETECTOR_HALF=0 = never. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>439ec70
12 days ago
-
Fast engine on CPU about a third faster: versions pre-ranked by the two small detector models
unknown +1: Claude Opus 5.5On a CPU the large DeBERTa detector was about 2/3 of the detector time. Every version is now ranked by the two small models (on its own), the fluency check runs on the shortlist, and only the 2 best per sentence go to the full three-model detector; the document's AI score is still always the full detector's. Benchmark (4 AI paragraphs, laptop CPU): 100% -> 0% in 3 rounds in 218 s (was 320-350 s). int8 quantisation of the detector was tested and rejected: it changed 8 of 77 verdicts (devtools/bench_int8_detector.py). Capping versions per model at 2 was tested and rejected: it stalled at 23%. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>bc66a36 -
Fast engine: a word that opened the author's sentence is lower-cased when a rewrite moves it inside the sentence
unknown +1: Claude Opus 5.5Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>93991f7 -
Fast engine: stronger first round, rounds 2+ judged in context, never builds on a worse round, stock AI phrases always rewritten
unknown +1: Claude Opus 5.5Benchmark (devtools/bench_fast.py, 4 AI paragraphs, 3-model detector): 100% -> 0% in 3 rounds and 100% -> 6% in 4 (was 4 rounds for 0%, first round 100% -> 79%). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>0cb108c -
Fast engine: round 1 replaces every flagged sentence; larger Gemini batches with fallback for refused ones
unknown +1: Claude Opus 5.5Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>f85f16b -
Fast engine: versions from several models at once (Gemini 3.5 Flash-Lite + Qwen 3.8 by default, gpt-oss-20b as fallback), checked inside the worker threads; generator benchmark
unknown +1: Claude Opus 5.5Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>01f5654 -
Fast engine: every flagged sentence replaced at once (versions judged on their own), 16 sentences per request, keep-every-detail prompt applied in the worker threads, step timings
unknown +1: Claude Opus 5.5Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>d46df30 -
Fast engine self-test on the Space (FAST_SELFTEST=1): timings for ~1.8k, 5.5k and 15k words in the log
unknown +1: Claude Opus 5.5Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>4d0e2bc -
Fast engine (opt-in with ?engine=fast) and one-step document upload
unknown +1: Claude Opus 5.5- fast.py: the whole document at once. The detector reads it once; every flagged sentence is rewritten in parallel by the language model (keep-every-detail prompt); versions must keep terms / numbers / citations, the meaning (embeddings + an independent faithfulness reading) and read as fluent English; the detector picks the most human version in context. Rounds repeat on what is still flagged within a time budget (FAST_BUDGET, 50 s). Live engine unchanged unless SMART_ENGINE=fast or the page is opened with ?engine=fast. - Smart Humanize opens on the Word upload, and dropping a file starts the run; the separate Humanizer tab is hidden. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>4a7a056