# Inside the r34 SaaS research workbench

A narrated walkthrough of BYOK, BYOS, a real Bell-benchmark run, private equation overrides, and the AI research assistant’s inspect–propose–validate–activate workflow.

## Watch the demo

[Open the video and chapters](/documentation/#demo). English captions and native full-screen controls are available. The video is an edited sequence of actual interface captures, with synthetic narration.

## Recording scope

The recording uses the actual r34 SaaS interface in a controlled local deployment. The AI provider responses are scripted for a repeatable demonstration. Numerical simulation and candidate validation execute locally. The demonstration executor does not establish production sandbox isolation. Storage setup controls are shown; no S3 or NAS transfer is demonstrated, and no customer cloud account or NAS is represented as connected.

The captured Bell-workflow candidate adds a reporting diagnostic without changing the Bell-state physics. Two additional chapters explain equation overrides and show original r34 acceptance-test evidence: a deliberately introduced Lorentz time-equation sign fault in an isolated test copy is corrected through a private source layer, validated and activated. A second dependent, comment-only layer is subsequently validated and run. The displayed event times come from the retained numerical artifacts. These are explanatory evidence cards, not new GUI screenshots.

This is a workflow and equation-override capability demonstration. It does not claim a live AI model evaluation, a discovered production solver defect, or experimental validation. A source layer can change predictions while still being required to pass the implemented reference checks. The shared installation and trusted validation infrastructure remain unchanged.

## Before you begin

- Obtain access to a PF application package or a hosted workbench. This public website does not run the research application.
- Confirm which provider, model, compatible endpoints and scientific runtimes your operator permits.
- Enter provider credentials only in the dedicated API key field. Saving the settings does not verify provider access.
- For BYOS, save the destination, test storage access, and select the default before creating new storage scopes. Existing work remains attached to its original destination.
- A Linux/POSIX computer or NAS connector must stay online. The selected AI provider can receive relevant evidence, source excerpts and attachments.

## Narration transcript

### From equations to evidence

Welcome to PF Simulation, release r34. The PF Research Workbench connects your scientific question to private enhancements and repeatable simulations. Those enhancements can change equation implementations, as well as analysis and reporting. This walkthrough combines actual interface captures with an equation-level acceptance test. The AI replies are scripted; the numerical calculations are real.

### Begin with your research project

Let's begin in the PF Research Workbench. Your project brings together configurations, uploaded inputs, and completed runs. Organization roles control who can view, edit, or administer the workspace. Open a study to inspect its fields and full configuration. Before running it, choose your AI and storage settings.

### Bring your own AI account

First, bring your own key. In the AI research assistant, choose an enabled provider and its model, then save your key in the dedicated field. These settings follow your account across projects. The server encrypts the key; saving it does not test provider access. Your chosen provider may receive relevant evidence and source excerpts during an investigation.

### Choose where new work is stored

Next, choose where new work is stored. Keep managed storage, or add an approved S3 destination or a Linux computer or NAS connector. Save the destination and test access before selecting it as the default. Existing work keeps its original destination, and the connector must stay online. Bulk files can use your storage; authorization, conversations, layer records, and trusted hashes remain platform managed. Here, we show the setup controls.

### Establish a simulation baseline

With those choices in place, establish a baseline in PF Simulation. Select a documented study, review its assumptions, validate the configuration, and run it. This example uses a Bell benchmark. The results distinguish successful execution from numerical checks and scientific interpretation. The saved plots, data, and logs give the assistant evidence to inspect.

### Inspect the evidence

Now open the AI research assistant from the completed run. Ask it to explain the recorded result and the equations behind it. The investigation can inspect permitted scientific source and saved evidence, with tool activity available for review. Attach supporting material when useful. That connects your next proposed change to a specific calculation.

### Propose a private enhancement

With the evidence in view, describe your enhancement and the behavior it should preserve. The assistant prepares a private Python layer, an explanation, and a reviewable code diff. The Bell example on screen adds a reporting diagnostic. The same mechanism also supports changes to equations and solver implementations. Let's look directly at that capability.

### Override equation implementations

An equation override changes the mathematical code that a private simulation executes. In permitted scientific Python files, a layer can replace an equation implementation or introduce a model change. Additional layers build on the exact preceding stack. PF Simulation applies them to a private source copy, while the installed shared engine remains unchanged. These are executable source changes.

### A demonstrated equation correction

The retained release tests demonstrate this with a deliberate sign error in the Lorentz time equation, introduced only in a test copy. A private layer restores the correct equation. Numerical probes record the changed event times, and the activated stack executes that correction. A second dependent layer is then validated and run. This verifies equation-override capability; the provider replies are scripted, and no production solver defect is alleged.

### Validate the candidate

Returning to the workbench, validate your proposed candidate before activating it. PF Simulation executes the baseline and candidate stacks, compares numerical outputs, and runs its required reference checks. Changed predictions are allowed; the reference checks still apply. Review the report and its scope. A numerical pass supports the tested cases. Experimental claims require experimental evidence.

### Activate. Run. Refine.

Once you have reviewed a passing candidate, explicitly activate it and run the study with your active profile. PF Simulation applies the selected layers in order. Their identities and source hashes connect the results to the exact implementation used. The original results remain available, so you can compare, refine, and continue your investigation.

### Keep the evidence reproducible

To complete the cycle, inspect the derived results and download the report and saved artifacts. Keep the configuration and enhancement stack with your interpretation. Saved history and conversations support the next investigation. The PF Research Workbench also provides documented satellite, numerical relativity, and quantum workflows, with specialist integrations where their required runtimes are available.

### Build around your next question

For academic and corporate research teams, this is a practical path from equations to evidence: inspect the implementation, propose a private change, validate its numerical behavior, and explicitly activate the chosen stack. Explore the guides and reproducible gallery studies to plan your next question. As your research evolves, PF Simulation evolves alongside it.

## Useful controls to revisit

| Task | Control in r34 |
| --- | --- |
| Configure personal provider access | AI research assistant → Your personal AI account → Save personal AI settings |
| Choose new-work storage | Account → Your storage → Add a storage destination → Test storage → Set default |
| Inspect a saved execution | Runs → View results → AI research assistant |
| Review a proposed change | Your personal override stack → Inspect changes |
| Evaluate and adopt a candidate | Validate → review Validation result → Activate |
| Execute the selected stack | Run this study with my active profile |
| Reuse an ordinary run | Runs → Use configuration |
| Export an ordinary run | Results → Download run ZIP |
| Export assistant-derived evidence | Derived run → Download report and individual artifact links |

## Continue with the guides

- [AI Research Assistant](assistant.md)
- [Bring Your Own Key](byok.md)
- [Bring Your Own Storage](byos.md)
- [Continuation and attachments](continuation-attachments.md)
- [Scientific integrations](workflow-integrations.md)
- [Reproducible Gallery studies](/gallery/)

## Media credits

Synthetic narration uses the stock Kokoro-82M v1.0 `af_heart` voice. No voice cloning was used. The revised narration uses explicit acronym pronunciations and continuous speech synthesis where supported. Captions retain the canonical sentence casing. The source package includes the storyboard, captured screens, equation-test evidence, timing data, rendering script and media provenance. Original interface and scientific outputs come from the supplied r34 PF Simulation package.