RSI4Science
English

This page is available in English. Choose a language to open its dataset guide.

Benchmarks and demos

Scientific verification · teaching fixture

Inspect a scientific reference and checker

Identify and forecast a damped oscillator from public synthetic observations. Use this small package to learn how inputs, a numerical reference and rejection checks fit together.

Run the teaching package

Public synthetic data, a numerical reference and rejection checks. Python 3.9+ and its standard library are sufficient; no network or paid model call is needed.

Download starter package

After extracting, run from the package directory: python3 -B test_smoke.py

Expected result: seven tests followed by OK. These check the numerical reference and rejection rules.

All answers are public. Passing this check demonstrates agreement with the declared teaching model, not agent learning or recursive self-improvement.

The question

Can a proposed output recover the oscillator parameters and forecast its trajectory under the declared noiseless model?

Where recursion would happen

A future research scenario could let a system revise its scientific diagnostics and use the revised tools to guide later self-modification. That scenario has not been constructed.

The current solver treats each episode independently. There is no editable improver, installation loop, future improvement test or recorded agent run.

Scientific feedback and evaluation

All observations and analytic answers are public. The reference calculates its estimates from the input trajectories; it does not read the answer file.

The checker tests parameter and trajectory errors and rejects malformed or non-finite output. Its strict tolerances apply to this noiseless analytic fixture, not arbitrary physical observations.

The comparison we need

The teaching package cannot compare recursive and fixed improvers. A research extension would need a justified scientific limitation, independent workspaces, matched controls and actual resource records.

Read the shared evaluation design.

Available now

  • Public synthetic observations and analytic answers
  • An input-only numerical reference solver
  • Output validation and deliberate failure checks

Needed for a future study

  • A research scenario with observed limits and independent assessment
  • An editable improver and verified later use of its revisions
  • Controlled agent runs and independent scientific review

A bounded contribution · Reproducibility and checker engineering

Reproduce the reference and inspect one rejection case

What to deliver

  1. Record the package version, environment and smoke-check result, including any failure.
  2. Explain one incorrect or malformed prediction and why the checker rejects it.
  3. Propose a bounded improvement to documentation or validation, with its scientific assumptions.

How it will be reviewed

Agree the scope and reviewer first. Review checks whether the evidence is reproducible and the explanation matches the declared model. No model experiment is required.

Sources and contributors

The resource page records the original source and acknowledges accepted reproduction, documentation and software contributions.

Source and reuse notes.

Send your reproduction notes

Agree a manageable scope and a reviewer before starting.

Project-led design. No independent scientific approval or completed RSI experiment is claimed. Accepted contributions and review decisions will be recorded as that work occurs.