Jean-Sebastien Beaulieu

Posted on Aug 31

Day 5 of WebMCP-QCG: I Stopped Adding and Started Proving

#devjournal #webdev #ai #quantum

Day 5 began with one of the most dangerous moments in software development:

the product worked.

That is exactly when my brain becomes a hackathon vending machine. I press one successful feature and six new ideas fall out. More providers. More languages. More panels. Better graphics. Deeper memory. A smarter agent. Another integration that will only take twenty minutes.

This time, I stopped.

WebMCP-QCG already had a working foundation. My job had changed. I needed to turn it into a product I could demonstrate, measure, defend, and finish.

Day 5 became the day I stopped collecting possibilities and started collecting proof.

The gate had to earn every next action

WebMCP-QCG operates before a costly or consequential quantum operation.

A quantum artefact enters the browser. QCG computes its real SHA-256 digest, identifies its capabilities, evaluates the intended operation, explains one recommendation, asks for a human decision when authority is required, and produces an evidence receipt.

Its four quantum tools cover four responsibilities:

inspect the experiment;
evaluate the proposed call;
run a bounded local simulation;
export the evidence report.

The decision engine can recommend reuse, rejection, recompilation, local simulation, or readiness for a later external workflow. Readiness describes evidence. Authority belongs to the person operating the gate.

An agent can recommend. The application can validate. I decide.

The feature freeze became part of that architecture. It redirected the remaining effort toward accessibility, recovery, testing, interface coherence, documentation, deployment, and receipts.

For once, my backlog did not get to eat my release alive.

The interface had to become mine

The seasonal interface helped me discover the product, but the application needed a durable technical identity. The four seasons now belong to the editorial series: Autumn for emergence, Winter for discipline, Spring for proof, and Summer for closure.

QCG itself became a browser console.

A navigation rail controls a central workbench. Inspector, Console, WebMCP, Decisions, Sources, Receipts, and Activity each have a real job. A persistent evidence panel keeps integrity, recommendation, authority, participants, and review requests visible.

The same sanitized state can appear across three surfaces:

the website;
a dedicated QCG panel inside Chrome DevTools;
a browser companion side panel.

Accessibility became architecture during the redesign. Keyboard navigation, visible focus, semantic labels, responsive reflow, contrast, reduced motion, and an accessible splitter improve the experience for people while giving browser agents a clearer interface to understand.

Dark and Light became the product themes. Cyan identifies active technical state. Emerald marks safe primary actions. Gold identifies human authority. Red is reserved for refusal and error.

The console finally looked and behaved like QCG.

Conversation became useful because authority stayed separate

I wanted Codex, Gemini-oriented clients, the website, DevTools, and the side panel to contribute to one investigation.

The practical solution was shared, bounded application state.

Compatible MCP clients can target the same browser page, read sanitized context, post observations, challenge a hypothesis, and request human review. QCG records the declared actor, intent, confidence, references, sequence, and resolution state.

Gemini inside DevTools follows a visible human relay. QCG prepares a sanitized handoff package. I can copy it into Gemini, review the response, preview the structured result, and explicitly import it into the ledger.

That boundary gives the collaboration its strength. Agents can debug, search, find, brainstorm, and request a decision. The person applies the disposition. A collaboration message carries evidence; quantum consent comes from the active human workflow.

My goal is to give the human a much better loop.

Ten profiles, two executable paths

The quantum surface expanded through explicit capability contracts.

QCG now recognizes ten ecosystem profiles. Q# and OpenQASM earned executable status through the same bounded QDK/WebAssembly Worker. Eight additional profiles support static inspection: Qiskit, Cirq/TensorFlow Quantum, TorchQuantum, PennyLane, CUDA-Q Python, CUDA-Q C++, Amazon Braket, and textual QIR.

Recognition means QCG can identify relevant structure, calculate a real digest, expose capability facts, and produce a reviewable recommendation. Execution means the runtime has earned that status through limits, fixtures, tests, cancellation, receipts, and human consent.

I am building a preflight workbench. Every future adapter now has a clear graduation path: define the capability, implement the boundary, test the state, preserve the receipt, and identify the authority required for the next effect.

Then I made the numbers mean something

I finished Day 5 with a staged E2B campaign against the deterministic decision engine.

The campaign progressed from local validation to 10 sandboxes, then 50, then 100 concurrent sandboxes. Two complete million-operation passes produced matching digests across all 100 environments.

The consolidated record contains:

2.6 million validated decision-engine operations;
100/100 matching sandbox digests;
zero unauthorized effects;
zero missing receipts;
zero decision mismatches;
zero surviving sandboxes after cleanup.

These numbers measure reproducibility in QCG's decision and evidence engine. Quantum performance belongs to a different experiment.

I tested delivery separately. A deliberately small cPanel canary sent 80 requests to the public origin. All 80 returned HTTP 200, with zero errors and zero timeouts. That receipt measures delivery health.

Measurements become useful when I let them describe the system they actually measured.

The canonical application is now live. The regression suite remained green, the reviewed build matched the deployed origin, and the functional product entered feature freeze.

I still have ideas. Of course I do. My brain opened several imaginary repositories while I wrote this sentence.

They can wait.

This work matters to me because I think about my children and the learners whose hardware, geography, or budget narrows their access to advanced systems. A browser-native gate can reduce the knowledge and coordination cost of approaching quantum work responsibly. It can help a person understand what can run locally, what should be reused, what needs another target, what evidence exists, and who controls the next action.

Spring is often described as uncontrolled growth. I learned another version of it.

Growth becomes durable when the roots, branches, and limits develop together.

On Day 5, I stopped adding. I started proving.

Explore the project

Research-partner disclosure

I used Codex and Gemini in Browser to its fullest potential as a research partner—for code mapping, source comparison, evidence organization, consistency checks, editorial control, and deliverable preparation. I formulated the intent, defined the scope, interpreted the results, arbitrated the conclusions, and preserved every public decision. This collaboration expands my investigative capacity; judgment, responsibility, authorship, and final signature remain under my authority.

DEV Community