Recommend a task policy from recorded outcomes.
flux@store: ~/run
$ flywheel --run
✓ define task kind and policies
✓ record an observed outcome
✓ score win rate, reward, exploration, and decay
› recommend the next policy
done — output ready ▍
How it works
1
Define task kind and policies
2
Record an observed outcome
3
Score win rate, reward, exploration, and decay
4
Recommend the next policy
Why it beats the alternative
This product
A fixed policy choice
Evidence
Recorded outcomes
Initial preference
Exploration
Explicit bonus
Rarely revisited
History
Append-only ledger
Latest state only
Boundary
Recommendation follows supplied rewards
Bad metrics produce bad policy choices
How it actually works
Exactly what it does — and when.
1Record task kind, policy, and observed outcome
2Combine win rate, average reward, exploration, and time decay
3Return an inspectable next-policy recommendation
In practice · Two review strategies accumulate outcome history
A team records accepted-task outcomes for two review policies. FLYWHEEL recommends the stronger observed route while preserving an exploration term, and the operator audits the append-only history before allowing the recommendation to influence routing.
Use it to…
Choose among task policies
Record outcomes for several approaches and request the next recommendation for that task kind.
Keep exploring alternatives
Use the exploration term so an early winner does not permanently exclude other policies.
Audit a recommendation
Read the append-only history and component scores behind the choice.
The shift it creates
Without it
- Strategy choices rely on memory or intuition
- An early winner is never retested
- Outcome history is overwritten
→
With it
- Keep policy outcomes in an inspectable ledger
- Balance observed reward with exploration
- Apply time decay so older observations weigh differently
What's inside
Per-task policy learning: win-rate + avg-reward + explore + decay
record → recommend → policyFor: a compounding loop
Works on any task with a measurable outcome
Append-only ledgers you can audit
Free · zero-dep Node · MIT source
Who it's for
Developers comparing task strategies
have outcomes they can record consistently
Agent-router builders
need a small policy recommendation primitive
Experimenters
want transparent scores rather than a black-box optimizer
Questions, answered
✓ Free · no email✓ Yours to own✓ Use on every project✓ Built to ship results
Ready to put FLYWHEEL — Task-Level RL Brain to work?
Free download · no account needed · yours forever
FLYWHEEL — Task-Level RL Brain
Instant download · one-time price · yours forever
FreeDownload Free