Dual-Calibrated AI Detection (Pangram v4 & GPTZero) · Structure-Safe Multi-Domain Humanizer · Safe-Edit Gate & Bulk v1 API
Trained on non-marketing writing (tech, academic, finance, news, health, how-to) and dynamically blends our new v2 DPO adapter with r16 while stripping robotic listicles into flowing 3–4 paragraph prose.
Instead of paying $3,000/mo for a 24/7 cloud B200 or waiting 3 minutes for serverless cold starts, our 8B model + 3 LoRA adapters sit warm in 10.1 GB of local T4 VRAM and rewrite in ~12 seconds.
Still strictly bans any made-up statistic or hallucination (0 fabrications), but stops rejecting great human rewrites just because natural phrasing compressed 1 minor setup detail (+20.5% pass yield).
Our free CPU detector (0.934 Pangram-4 AUROC, 0.929 GPTZero AUROC) ranks drafts for $0, and our 509-case repair study proved edits adding <20 words on a solid pass have a 0.0% flip rate.
Previously, our humanizer had four pain points: (1) it was trained mostly on marketing articles so it struggled on technical, academic, or finance pieces; (2) hosting a 70B model 24/7 on a B200/H200 GPU costs ~$3,000/month; (3) the fact-checker failed any draft that left out even a single minor setup detail; and (4) every small repair or rule edit required paying for brand-new GPTZero and Pangram 4 scans. Here is how Humanizer v2 solves all four:
| Goal | Old System | Humanizer v2 Solution | Measured Result |
|---|---|---|---|
| a. Multi-Domain Evasion | Skewed toward marketing copy; single LoRA adapter (`r16`) with `repeat_penalty=1.0` | Trained multi-domain SFT+DPO v2 adapters, blended with `r16`, added context-wide `repeat_penalty=1.10`, and formatted listicles into 3–4 paragraph prose | Passes both GPTZero & Pangram 4 ("Human") across Technical, Academic, Finance, Marketing, and Casual domains |
| b. Instant $0/mo Hosting | 70B cloud endpoint (cold-start delays or $3k/mo 24/7 B200 server) | Quantized `Llama-3.1-8B-Instruct` (`Q8_0`) + 3 hot-swappable LoRAs running in 10.1 GB VRAM on this server's idle NVIDIA Tesla T4 GPU | $0.00/month recurring cost, 0.0s cold start, ~12s per rewrite (23.4 tok/s) |
| c. Reader-Faithful Grading | Failed any draft missing 1 peripheral setup fact (`43.6%` false-alarm rejection rate) | Strictly requires 0 fabrications (zero made-up claims), while allowing 1–2 minor omissions when ≥85% of facts survive | +20.5 percentage points higher valid pass rate with 0 hallucinations |
| d. Local Checker & Flip-Line | Re-scanned every post-edit on paid APIs; old local detector didn't predict Pangram 4 | Built CPU dual-detector (`ModernBERT` + `Qwen-0.5B` token surprisal cone) + empirical 20-word edit gate | 0.9343 AUROC on Pangram 4, 0.9287 AUROC on GPTZero; 0.0% flip rate on safe micro-edits |
| Domain | Held-Out Test Piece | Old r16 Alone (GZ / PG4) | New Local T4 v2 / Blend (GZ / PG4) | Reader-Faithful (Recall / Fabs) | Action |
|---|
Live versioned endpoints ready for external SaaS billing, bulk domain audits, and internal agent workflows.
Base URL: https://chat.frac.tl/humanizer-v2/v1
| Endpoint | Method | Marginal Cost | Latency | What It Does | Live Test |
|---|---|---|---|---|---|
| /v1/detect | POST | $0.0000 (Local v2) | ~90ms / batch | Single or batch AI screening (up to 100 texts/call) predicting both GPTZero (0.929 AUROC) & Pangram-4 (0.934 AUROC), with optional live verification. | |
| /v1/humanize | POST | ~$0.0003 – $0.002 | ~8–14s (T4 GPU) | Prose bridge + Local T4 8B Multi-LoRA rewrite + burstiness-safe grammar polish + Local v2 / Live GPTZero+Pangram4 + Reader-Faithful check. | |
| /v1/safe-edit-check | POST | $0.0000 | ~45ms | Empirical Flip-Line gate (509-case study): verifies whether a post-humanizer edit (<20w added on a solid pass) can safely skip a paid re-scan. | |
| /v1/usage | GET | Free | ~5ms | Returns per-API-key word usage counters, active T4 GPU VRAM status, and model version metadata. |
# 1. Instant $0 Dual-Detector Screen (Single or Batch)
curl -s -X POST https://api.dropslop.ai/v1/detect \
-H "Authorization: Bearer hk_live_demo_fractl_v2" \
-H "Content-Type: application/json" \
-d '{
"texts": ["Your passage or bulk URL text here..."],
"mode": "local_v2"
}'
# 2. Multi-Domain Humanize + Grammar Polish + Audit
curl -s -X POST https://api.dropslop.ai/v1/humanize \
-H "Authorization: Bearer hk_live_demo_fractl_v2" \
-H "Content-Type: application/json" \
-d '{
"text": "Your AI draft here...",
"lora": "blend",
"passes": 1,
"mode": "verified"
}'
# 3. Safe-Edit Flip-Line Gate ($0 Re-Check Skipper)
curl -s -X POST https://api.dropslop.ai/v1/safe-edit-check \
-H "Authorization: Bearer hk_live_demo_fractl_v2" \
-H "Content-Type: application/json" \
-d '{
"original_text": "Verified humanized passage...",
"edited_text": "Verified humanized passage with a quick tweak.",
"before_gz": 0.02,
"before_pg": 0.0
}'
// Click "Run /v1/detect", "Run /v1/humanize", "Run /v1/safe-edit-check", or "Run /v1/usage" above to inspect the live JSON payload.