AI Trial — Our Models, Live
Two open-weights models, served directly from the SigPulse testbed and routed through our own gateway (rate-limited, quota-tracked). No third-party inference provider is involved.
Access
Guests can try everything right away — 10 actions per day
(chat, diagrams, web search, video briefs and the workshop all share the same 10), no signup needed.
Want more? Email
intel@sigpulse.com
for a personal key — 50 actions/day and higher tool quotas.
The ⚖️ patent desk carries its own quota: guests 3 turns/day, personal keys 10.
The 🧹 style pass carries its own quota too: guests 9 / personal keys 45 shared between scans and rewrites. The scan is local and deterministic; the rewrite runs on a cloud model (GLM) — same disclosure as the patent desk.
Conversations are not stored on our servers.
checking backend status…
What it can do →
Ground rules
- Models can be wrong. Outputs are not editorial content and don't represent SigPulse's reporting.
- Bonsai (default): fast, literal, zero-fabrication profile. Q38 deep mode: stronger prose, but its host runs nightly video production — it may go offline without notice.
- Daily quota applies (guests 10 actions/day by IP, keys 50; UTC reset). Abuse gets a key revoked, not an argument.
- ⚖️ Patent desk: drafting aid and bibliographic search only — not legal advice; have a licensed patent agent review anything before filing. It runs on a cloud model (GLM) via our workstation agent stack — separate from the two open-weights models above — with its own quota (guests 3 / keys 10 turns per day). Explanations follow Simplified Technical English (ASD-STE100) principles in English replies (short active sentences, one idea each, consistent terms); Chinese replies apply the same clarity rules in natural Chinese. Pre-filing confidentiality note: your text is processed by a cloud model; for inventions where pre-filing confidentiality is critical, share only an outline here, or work on the full text after filing.
- 🧹 De-AI style pass: the scan is a deterministic pattern check (no model sees your text; rules cover English formally (narrative tells pending); Chinese covers formal clichés plus cross-model density invariants (onomatopoeia/exclamation, stacked-formula, simile, moral endings, analyst-jargon, tech-tutorial formulas)). The rewrite step runs on the chat model you pick. Scan quota: guests 5 / keys 30 per day, separate from the shared action pool.
- 🔥 Hot board / 📡 AI chain: live rankings and AI-supply-chain news pushed hourly from our workstation radar (top-N per source; quick comments pre-generated by a local model; region tags are language-inferred and advisory only). The Douyin board is fetched server-side from an open aggregator API (60s) instead — secondhand by a few minutes. YouTube boards (US/JP/KR) come from the official YouTube Data API within its free quota; Google Trends boards (US/JP/KR) come from Google's official RSS feed — hourly refresh, secondhand. Country tags on international boards reflect the chart/geo parameter (ISO-2). Browsing these boards runs no model and costs no quota.
- 🎥 Video check (视频验货): internal beta — personal key required, no guest access. Two engines: quick runs on our local open-weights model qwen3.8-27b (subtitles + top comments + metadata, extraction-grade initial verdicts); deep runs on a cloud model (GLM) via our workstation agent and adds live search cross-checks — final-ruling quality, slower, serialized. YouTube pages are fetched through a rate-limited proxy, so some videos degrade to "comments+data only" — the report says so in its limits. Reports are AI-generated reference, not fact adjudication: verdicts are heuristic, evidence links are not manually re-checked. Quota: quick 20 / deep 5 per key per day, UTC reset; same video+depth within 24h is served from cache and costs nothing. Optional transcript injection: paste a transcript into the request and the report is based on it — marked as user-provided (unverified), never entering the shared cache.
- We log key suffix, model, timing and token counts only — never conversation content.