Affine-structure choice
Each trial presents a target LUXVAR token and asks the user to choose between a published orbit-near answer and a morphology-oriented distractor.
VARZIN Cognitive CAPTCHA.
Structure as a challenge surface.
A tested local Streamlit prototype and hosted FastAPI generation service that turn controlled Aff(ℤ₁₂) structure and LUXVAR challenge pairs into an anti-automation research demonstrator. Commercial API access is distributed through RapidAPI, with the execution origin hosted on Render; production-grade bot resistance remains a separate, not-yet-established claim.
Evidence boundary
The public challenge construction is a reproducible prototype. It is not presented as a production-proven security barrier because current evidence does not establish deployment-grade bot-resistance metrics.
What the prototype actually does.
Aff(ℤ₁₂) + Core-30
The original generator builds matched challenge triplets from a designed Core-30 state representation and the 48 affine operators over ℤ₁₂.
Local MVP + hosted API
The corrected local UI is a Python/Streamlit app on port 7860. A separate FastAPI generation engine is deployed commercially through RapidAPI, with its origin hosted on Render.
Tested before being added to the site.
Generator executes
The published Cognitive Prior/CAPTCHA suite executed successfully and generated 10 challenge trials. All generated answer distances were finite and every selected distractor had a greater published orbit distance than its paired answer.
Legacy UI bug found
The older local Streamlit demo rendered correctly, but its “Proceed to Next Challenge” state was lost on rerun. This prevented reliable multi-challenge progression. The original file was left intact for provenance.
Verified prototype corrected
The original commercial MVP source was backed up and corrected in place, including persistent answer state and portable dataset loading. Automated Streamlit testing completed all 10 packaged challenges with score 10/10 and zero application exceptions. A replacement distribution package was assembled as MVP v1.0.2.
Local endpoint confirmed
The corrected original prototype was launched successfully at http://127.0.0.1:7860 and returned HTTP 200. The implementation found on the machine is Streamlit, not Gradio; the port remains configured as 7860.
The product has a live hosted API path as well as the local MVP.
RapidAPI Hub
Commercial subscriptions and consumer authentication are handled through varzin-cognitive-engine.p.rapidapi.com.
Render-hosted FastAPI
The execution origin is hosted on Render. Billable generation calls are designed to accept trusted RapidAPI-proxy traffic rather than unrestricted direct-origin use.
Zenodo DOI 10.5281/zenodo.22036769
The Level-1 software record remains the open academic/reproducibility reference. Hosted SaaS delivery and academic code availability are separate distribution layers.
Current API contract
GET /api/v1/generate
The live FastAPI contract uses query parameters modulus and samples. Older project material showing this endpoint as POST is superseded by the live OpenAPI contract.
curl --request GET \
--url 'https://varzin-cognitive-engine.p.rapidapi.com/api/v1/generate?modulus=12&samples=5' \
--header 'X-RapidAPI-Host: varzin-cognitive-engine.p.rapidapi.com' \
--header 'X-RapidAPI-Key: YOUR_RAPIDAPI_KEY'Why the current product claim is narrower than the old mockup.
The dedicated frozen-model audit reports near-chance recovery of the designed axes for the tested Mistral 7B and Llama 3 8B protocol. That is useful motivation for a challenge design, but later projection/training experiments show that targeted structure can be recovered after intervention. Therefore the current evidence does not support a claim that language models are inherently unable to solve the challenge.
Supported now
- A reproducible 10-trial challenge generator exists.
- A local browser UI can serve and score those challenges.
- The challenge design distinguishes a published orbit-near option from a morphology-oriented distractor.
- The underlying Track D frozen-model result is negative for the tested models and protocol.
Not yet established
- Human pass rate versus modern automated systems.
- Production false-positive and false-negative rates.
- Operational security and production service reliability.
- Generalization beyond the published prototype challenge family.
The public prototype is a demonstrator, not a production security claim.
The published challenge dataset is intentionally inspectable for reproducibility. A commercial deployment would need a larger private challenge family, ordinary web-security controls, and independent human-versus-automation evaluation before effectiveness claims are made.
Explore the challenge idea.
This browser exercise is intentionally transparent and is not an authentication gate. It demonstrates the interaction model using public prototype mappings.
Challenge
Select the published orbit-near token for:
ELŪZ–MAHAR
A credible route from experiment to SaaS.
Verified local MVP v1.0.2
Original paid-share source corrected, 10-trial dataset packaged relative to the app, scoring flow verified end-to-end, and unsupported production-security metrics removed.
Independent evaluation
Measure human usability and automated-system performance on a larger unseen challenge family.
Commercial service
Only after evaluation should product-level security metrics, service guarantees, pricing, or enterprise claims be published.