1
KNOWLEDGE
Upload Knowledge
Drop a knowledge file here
TXT · Markdown · DOCX · PDF · 10 MB maximum
StableAI L5.2 Verification Lab
See the No-Drift Framework in action.
Gemini interprets the knowledge and each question variation. StableAI then deterministically selects, renders and hashes one source-grounded answer.
1 Upload knowledge→2 Select or ask a question→3 Run verification→One canonical answer
2
QUESTION
Ask a Question
○Select a generated question.
RUNS
MODE
10 Gemini verification runs
How it works
Language varies. The verified answer does not.
Gemini performs semantic normalization only. StableAI application logic selects the canonical registry state, renders the answer deterministically, and verifies its evidence and hashes.
Extract
Read text from TXT, Markdown, DOCX or PDF.
Compile with Gemini
Extract complete source-backed facts while rejecting headings, fragments and invented evidence.
Resolve each run
Use Gemini to independently map every wording to requested registry meanings.
Verify
Confirm identical evidence, answer text and SHA-256 hashes.
Current capabilities
What the laboratory can do today
A Gemini-powered implementation of the StableAI L5.2 No-Drift workflow. Semantic work happens through the secured server API; canonical selection, evidence control and verification remain deterministic.
Knowledge ingestion
TXT, Markdown, DOCX and text-based PDF files up to 10 MB, with readable-text validation and source preview.
Gemini semantic compilation
Up to 100,000 characters compiled into complete source-backed facts, atomic meanings and stable identifiers—never document headings.
Visible structure
Compact metrics for source sections, supported facts, atomic meanings, canonical states and complex combinations.
Natural generated questions
Gemini creates simple and complex questions from the compiled registry without adding unsupported requests.
Semantic support checking
Gemini classifies manual questions as fully supported, partially supported, unsupported or conflicting before verification.
Real repeated API tests
5, 10 or 20 independent Gemini interpretations using exact repetition, equivalent wording, or both modes.
Canonical answers
Fixed source-backed rendering with explicit coverage of supported and unsupported requested parts.
Auditable proof bundle
Run comparison, registry consistency, answer and evidence hashes, unique-hash counts and local latency.
SEMANTIC LAYERgemini-3.1-pro-preview via secured server API
PURPOSEVerification demonstration
DETERMINISTIC LAYERRegistry selection, rendering and SHA-256 proof
Why StableAI
Complete, grounded and auditable—not merely repeated.
Grounded by construction
Every canonical answer is built from exact uploaded evidence. Unsupported information is explicitly marked instead of invented.
Controlled after meaning detection
Question wording may change, but deterministic selection and rendering control the final answer.
Proof you can inspect
Variations, requested meanings, coverage, canonical answer, evidence and hashes remain visible to the user.