Constrained operations
Two fixed datasets, two fictional assets, and whole percentages from 0 to 100. Unknown fields, invalid inputs, arbitrary code, and model-selected URLs are rejected. The Python adapter validates the inputs again.
ENGINEERING NOTES / THE PORTFOLIO ITSELFMake the boundaries
Make the boundaries
part of the design.
A language model interprets the question. QuantCore owns the calculation. The application decides what each is allowed to do.
01 / INTERPRET
Language to settings
The model proposes two allocations and one of two datasets. Its response must match a strict schema and pass server validation.
02 / COMPUTE
Settings to Python
A fixed adapter runs QuantCore’s original numerical modules in a browser worker. Visitor and model text are never executed as code.
03 / EXPLAIN
Results to context
The server supplies reference results generated by the same pinned Python source. The model cannot substitute visitor-provided numbers.
Small surface. Explicit limits.
One spending boundary
The tour and lab share one persistent lifetime allowance. An atomic database reservation happens before every AI call, including each stage of a comparison. Failed calls keep their reservation; requests are not automatically retried. Missing storage stops paid calls.
Server-side secrets
The OpenAI credential stays on the server. Requests have size and time bounds, browser-origin checks, and a shared hourly limit. The site stores a keyed visitor hash for rate limiting, rather than the raw IP or conversation.
Browser protections
Document scripts use a per-response nonce. Separate content policies constrain the page and Python worker. Framing is blocked, content-type sniffing is disabled, and unused camera, microphone, location, and payment permissions are denied.
Inspectable computation
Four numerical modules are pinned to a QuantCore commit and copied without modification. Source and dataset hashes identify what runs. Numerical checks compare all supported allocations with the reference results used by the AI.
Useful failure modes
Calculations remain available when the AI allowance is exhausted. The Python worker has a deadline and can be restarted. AI output is rendered as text, and old explanations retain the settings they describe.
Follow the implementation.
Pinned QuantCore source ↗Source & dataset hashes ↗Fixed Python adapter ↗API boundary tests ↗Numerical verification ↗
These are implemented safeguards, not a claim of a security audit. Origin checks do not authenticate scripted clients, and a public endpoint can still be abused. The Python runtime is a pinned external dependency. AI explanations can be wrong; the calculations and sample data stay available for inspection.
WHY BUILD IT THIS WAY?QuantCore’s value is that the math can move between applications. This experiment puts that idea into practice while keeping language generation separate from numerical truth.
Try the experimentAsk, calculate, explore ↗