SELECTIVE ONLINE LEARNING LAB · V1.0

Spend updates
where they matter.

A transparent test of whether a causal usefulness gate can lower prediction error without paying for every update.

01 / STREAM DESIGN

Configure a run

Browser preview: linear online regressor. Tiny recurrent learner runs in the local CLI only.

browser-sized

Stream length 480

60 steps1,200 max

Recurrence 0.82 Noise 0.08 memory in the targetlabel uncertainty

Update budget Seed per methodreproducible stream

Causal gate accountingi

Probe cost Probe every cost unitssteps

Ready. The example stream is deterministic.

Browser limit: 1,200 steps. Pause and cancel act at cooperative yield points.

02 / COMPARISON

What did the learner buy?

Import JSON

Lowest observed raw MSE—descriptive stream readout

Observed error / cost—descriptive; probes included

Total decisions—updates + probes

Evaluation rulebefore labelno future regime IDs

ERROR TRACE

Rolling squared error

Run the comparison to reveal the trace.

The chart shows rolling squared error for each method. Use the accessible table below for keyboard and screen-reader comparison.

Open accessible chart table

Descriptive browser-stream chart summary
Method	Mean squared error	First rolling error	Last rolling error
Run the comparison to populate this table.

METHOD LEDGER

Matched accounting

lower is better · update budget applies to every method

Per-method descriptive browser-stream accounting
Method	Mean squared error	Updates	Probes	Discarded	Total cost
No run yet. The seeded example is ready.

Inspect predictions, actions, probes and charged work

Method trace

Run a comparison to inspect its step trace.

Selected method step trace
Step	Input x	Prediction	Label y	Action	Squared error	Probe
Run the comparison to populate this trace.

03 / CAUSAL CONTRACT

What counts as a fair test?

✓

Predict before labels

Every prediction uses only the current feature and past observed targets.

✓

Probe cost is visible

The gate pays for counterfactual history checks; rejected updates stay in the ledger.

✓

No latent regime IDs

Recurring forcing stays inside the stream generator, hidden from every method.

✓

Held-out work is local

Use the CLI for repeated seeds; this browser run is an instrument, not a final claim.

04 / RESEARCH READOUT

Keep the result honest.

∿

This browser view is one seeded empirical instrument. Its readouts are descriptive only; they do not establish a scientific winner, equivalence, or negative result.

Evidence label EMPIRICALClassification EXPLORATORY