Public research preview

Precise edits. Honest evidence.

Repair a model.
Measure the change.

Inspect a small weight edit, preserve your original model, and compare the result with matched controls. Keep the evidence behind the decision.

Explore Weightspace ↗Read the measured result ↓

The browser displays research and experiment receipts and provides a downloadable toolkit. Model inference runs locally on a compatible Apple Silicon Mac; the browser does not run a model.

THE REPAIR WORKFLOW
01 / Preserve the originalInspect the exact edit.
BaseOriginal weights
ControlPrompt change
CandidateWeight change
02 / Measure the comparisonKeep or discard. With a receipt.

An illustration of the workflow. Published research, local model experiments, and deterministic examples have separate evidence.

A focused repair.
A fair comparison.

Weightspace turns weight space programming research into an inspectable workflow. The current toolkit replays a published sparse edit and tests bounded candidates against frozen tasks.

01 / INSPECT

Know what changes.

Read the source, target tensors, and repair provenance. Verify compatibility before applying the published edit to a separate model copy.

02 / RUN LOCALLY

Compare the same work.

Use the downloadable toolkit for matched local inference with fixed prompts, sampling, token limits, and a finite budget. Generated Python is checked in the macOS sandbox.

03 / DECIDE

Let failures count.

Review correctness, capped answers, and paired regressions. Development selects a candidate; fresh heldout tasks decide whether it earns the next step.

Measured local batch

90 generations.
No accepted edit.

Three goals tested the base model, a prompt control, the full published edit, and its two layer ablations. No candidate passed the development gate. Heldout tasks stayed sealed.

Inspect the batch receipts ↗
  • Three distinct jobs

    Python function correctness, strict JSON output contracts, and complete lists without repetition. Each job used six development tasks across five arms.

  • Original model retained

    The batch established no coding gain and promoted no edit. A published repair for one failure mode does not establish improvement on a new task.

  • CPU witness, separately

    A deterministic toy rule network demonstrates an exact, reversible weight change with executable checks. It is independent of the language-model experiments and does not show learned-model improvement.

  • A bounded research tool

    Experiment receipts and negative results are available to inspect. Autonomous discovery of useful new patches and general coding-bot improvement remain unproven.

Read it on your phone.
Run it on your Mac.

The public ChatGPT Site is the workbench. Local setup and model downloads are optional; viewing research and saved receipts does not start inference.

Open the public workbench ↗Research, repair provenance, experiment planning, and saved results. Get the local toolkit ↗Python source, the published patch, setup instructions, and notices. Follow the research ↗Primary sources and a pinned GitHub landscape with constraints. Read the release record ↗Public availability and the scope of the evidence.