Developed and published by Datasein Oy (Finland; Chief Architect: Jari Katajisto). Every Kataguru Sceptic edition embeds the Multi-Mode Master Chat Template: 6/7 native operating modes without model swapping, 12 cognitive reasoning frameworks ({REASON:...}), and sticky language/plain-language switches.
Instead of unloading and reloading separate 20–70 GB models for coding, translation, plain-language rewriting, and dataset auditing, a single Kataguru Sceptic instance stays permanently resident in GPU VRAM. Concurrent requests can switch operating modes, cognitive frameworks, and languages on the fly with zero reload latency:
Executes a 3-stage DETECT → VERIFY → VERDICT-FIRST JSON audit. Rejects hallucinations, math/code bugs, refusals, sycophancy, and political/commercial lobby bias. Enforces learning_value >= 8.0.
Distills raw books, standards, or technical documentation into golden SFT/QLoRA training pairs (learning_value >= 8.5) with complete, un-truncated code and zero ideological or commercial bias.
Repairs rejected high-potential training samples (Quadrant Q2): diagnoses broken equations, code bugs, anglicisms, or bias, and rewrites the response from first principles to >= 8.5 golden quality.
Zero-output-token reflex judge (<20 ms latency). Outputs a single digit (0–9) to classify samples into Q1 (Accept >=8), Q2 (High topic potential → send to Surgeon), or Q3 (Discard trivial slop).
Sticky zero-preamble translation mode with automatic direction detection (EN ↔ Target Language, or any of 110 pairs in Polyglot-274E). Preserves technical terminology, LaTeX, code blocks, and literary voice.
Default mode featuring ASD-STE100 precision, 12 selectable cognitive reasoning frameworks ({REASON:ihyper}, {REASON:007}, {REASON:logic}, {REASON:spoon}, {REASON:einstein}), and sticky language switches.
Native to Polyglot-274E: accepts retrieved documents in ANY combination of 11 languages (via documents=[...] or /rag + <documents>) and synthesizes a strictly grounded answer in your chosen target language with mandatory inline [1], [2] citations, conflict detection, and zero extrapolation.
Include any tag in your message (sticky until changed). Prefixing the mode name with 'i' (e.g. {REASON:ihyper}) injects the cognitive framework directly into the system prompt and closes <think></think> immediately (+8.69 pp accuracy on structured tasks, -31% latency):
| Reasoning Tag | Cognitive Framework & Behavior |
|---|---|
{REASON:ihyper} / {REASON:hyper} |
MECE Structural Decomposition (Mutually Exclusive, Collectively Exhaustive): Decompose → Invariants → Edge Cases → Synthesis. |
{REASON:i007} / {REASON:007} |
007 Fault-Tolerant Agentic & PLC State-Machine Loop: Preconditions → Interlocks/Watchdogs → Deterministic Code or <tool_call>. |
{REASON:ilogic} / {REASON:logic} |
First-Principles Axiom Deduction: Conservation laws → Symbolic step-by-step derivation → SI dimensional & boundary-limit check. |
{REASON:ispoon} / {REASON:spoon} |
Spoon Deep Research: Critical multi-perspective scientific audit and premise verification without hollow roleplay theater. |
{REASON:ieinstein} / {REASON:einstein} |
Einstein 10-Perspective Brainstorm: Applies 10 orthogonal scientific/engineering perspectives before synthesizing a novel solution. |
/no_think • /think • {REASON:low..deep} |
Direct control over <think> depth: instant response (/no_think) vs. adaptive or fixed-depth internal chain-of-thought. |
/help • {REASON:help} |
Prints the model's built-in interactive manual of all 6 operating modes, reasoning tags, and language switches. |
Language switches remain active across the conversation until changed, working seamlessly across all 6 operating modes:
| Switch | Target Language / Mode |
|---|---|
/en • /fi • /auto_lang |
Force 100% English (/en), 100% modern Kotus Finnish (/fi), or restore automatic language matching (/auto_lang). |
/selko • /lihtne • /egyszeru |
Official Plain Language (8–12 words/sentence) for Finnish (fi), Estonian (et), and Hungarian (hu). |
/lengva • /viegla • /audvelt |
Plain Language for Lithuanian (lt), Latvian (lv), and Icelandic (is). |
/erraza • /martivi • /syml • /sade |
Plain Language for Basque (eu), Georgian (ka), Welsh (cy), and Turkish (tr). |