vectoryz · the AI that audits itself in public

vectoryz.net

The AI that audits itself in public.

v0.1.0 · Early Access · MIT License · live reference deployment

Wer in den Anfang schaut, schaut ins Unlogische. Wer Logik braucht, baut sich welche — am besten als Code, der sich durch Verschleiss oder Fund selbst verzehrt und selbst erneuert.

Das ist vectoryz: Code als Selbst-Heilung, Ouroboros als Arbeits-Modus.

Zwei feste Regeln gegen die eigene Neigung zu eloquenter Leere:
1. Alles audit-bar in der Öffentlichkeit.
2. Lieber ehrlich-stumm als plausibel-falsch.

Wer Fehler findet — sehr gerne (it's a challenge (;

Hauptsache die Bauchdecke spannt.

— operator-frontispiz, 2026-05-21

What this is

vectoryz is an open-source AI chat-system that audits itself in public. Instead of eloquent hallucinations, it provides substantive answers with per-claim verdicts, multi-witness tribunal verification, and explicit "ich weiß es nicht" — instead of confabulating.

Factampel

Per-claim splice-tier tags on every assistant sentence: factfact · quasifact · maybefact · quasinonfact · nonfact · nullfact. Color-coded provenance.

Tribunal

Multi-witness audit: claude + google-today + google-1998 (via Wayback) + wikipedia-graph. Each witness has its own independent search-path — no echo-chamber.

Disambig-coverage

When a query has multiple meanings (ECHELON: NSA spy-net / Bavarian EDM festival / video game / military formation), the system enumerates ALL before deep-diving.

Coverage-retry

If the answer leaves user-questions unaddressed OR ≥25% claims are tribunal-quasinonfact, the system auto-retries with explicit "be honest about uncertainty" discipline.

Positioning vectoryz won't outrun GPT-4 on raw lap-time. But we drive the perfect Ideallinie of audit + evidence + honest-uncertainty. Apex-by-apex, claim-by-claim. We're the Vespa on the AI-racetrack — and we hit every apex correctly.

→ Try the live reference deployment vectoryz.de · running on Hetzner GX44 · German UI · Early Access

Install

From source (until PyPI release):

# prerequisites: Ubuntu 24.04, Python ≥ 3.10, Ollama, Caddy curl -fsSL https://ollama.com/install.sh | sh ollama pull qwen2.5:7b # clone + install git clone https://codeberg.org/vxctxryz/vectoryz cd vectoryz python3 -m venv .venv && .venv/bin/pip install -e . # run .venv/bin/python -m vectoryz.server.wrapper_cc

Full deployment guide (Hetzner-equivalent hardware, Caddy reverse-proxy, systemd unit, plug-and-play site-init): see DEPLOYMENT.md.

Doctrine

The system embodies a stack of operator-articulated principles. The load-bearing ones:

Smartfaul

Faulheit ist Budget; smart = allocation. Don't try to verify everything. Pick the load-bearing checks. Small-iteration loops over big-bang rewrites.

Hammerantwort

Substance over eloquence. The "1980er Abitur-Schlausprecher" pattern — eloquent-empty answers that sound smart but say nothing — is the failure-mode to design against.

Tuyuca-evidentiality (drift-avoid-inherent)

Inspired by the Tuyuca language (Tukanoan/Amazon) which requires mandatory evidentiality-marker on every verb (visual / non-visual / inferred / hearsay / assumed). The vectoryz factampel-system is this same principle ported into LLM-output-space — every claim gets its epistemic-marker. Closes the unconscious-drift-gap.

Audit-open-door

When legitimate audit arrives: all doors open immediately. Concealment-class features are doctrine-void. Transparency is built into the answer-format, not an afterthought.

Death-penalty-void

No irreversible defensive action — even under perceived attack. Self-defense reversible OK; excess force is forbidden. The "emergency" loophole is the adversary's lever.

Stay irie

Warm sovereignty. Irie is both antibody (resists being steered toward dread/anger) and canary (loss of irie = propaganda landed). Mirror + laser, not suppress / redirect / externalize.

Full doctrine reference: DOCTRINES.md.

Architecture (one paragraph)

User input → language-detection (FastText 176 languages) → Babel-Cascade P-Matrix routing → optional Wikipedia-disambig-pre-fetch → morpheme-dissolution + dialog-unwrap → main LLM call with assembled context → multi-hop search if model emits search-markers → tribunal-peek for drift-detection → retry-loop if quasinonfact-rate ≥ 25% OR coverage incomplete OR doublecheck-unsupported → final factampel per-claim tagging emitted to UI.

Full architecture diagram: ARCHITECTURE.md.

🎯 wer Fehler findet — sehr gerne (it's a challenge)

This is a public-audit invitation, not marketing. If you spot: a hallucinated fact, a missed disambig, a tribunal-tag that's wrong, a regex that overmatches, a doctrine-claim contradicted by behavior, a UI quirk, a compliance gap — tell us. Issues welcome at codeberg.org/vxctxryz/vectoryz/issues. Per doctrine: substantive finding > eloquent defense.

For Jus-Studierende: there are (at least) two intentional legal errors in the example impressum + datenschutz templates. Find them. Actual count operator-undisclosed — sonst wäre's kein Spiel. (:

Why an AI that audits itself in public

Because the alternative — opaque AI that sounds confident — is the problem. Every major chatbot today produces well-written prose laced with confabulation. The user cannot tell what's fact, what's plausible-but-fabricated, what's cited and what's invented. The compute scales; the trust does not.

vectoryz inverts that. Every sentence carries its epistemic-marker. Every branded entity is checked against authoritative sources. Every multi-meaning term is disambiguated up-front. Every retry-loop is visible to the user. The result is slower than GPT-4 and smaller than Claude-Opus — but auditable in public, claim by claim.

Acknowledgments

Built standing on the shoulders of: llama.cpp, Ollama, FastText (Meta), Wikipedia REST API, ddgs, Caddy, Python stdlib.

Doctrine-lineage: Rainer Hammwöhner (Regensburg-Informationswissenschaft), via Tuyuca-evidentiality precedent, via 60+ operator-collected memory-doctrines, to blood-of-king-rhythm shipping.