The Self-Driving Discovery Brain — propose → experiment → prove.
Mneme remembers; Melete discovers. v0.6.0
When running an experiment is the expensive part — a lab assay, a training run, a process batch, a pricing test — Melete finds the most informative next experiment so you reach the best answer in the fewest trials, and emits a signed, offline-verifiable trace of exactly how it got there.
Tune learning rates, architectures, RAG/serving configs, compiler flags — fewer GPU-hours to the best model, with a provable tuning record.
Find the reagent mix / conditions that maximise yield or potency in far fewer assays — and a tamper-proof discovery trail for patents & audits.
Tune deposition/etch/print parameters against real KPIs on-prem — air-gapped, data never leaves the fab, result still verifiable.
Search price points, configurations, and policies where each test is costly — converge faster than grid/manual search.
In every case: fewer expensive experiments to the best answer + a cryptographic proof of how the discovery was made.
No single optimiser wins on every landscape (No-Free-Lunch). So a bandit spends each experiment on whichever strategy — Gaussian-Process+EI, CMA-ES, trust-region, annealing, space-filling — is winning on your problem. One engine, no per-problem tuning.
| landscape | Melete portfolio | single Bayesian | random |
|---|---|---|---|
| smooth | 1.000 | 0.999 | 0.838 |
| rugged (many traps) | best 🏆 — beats every single algorithm | far behind | far behind |
| high-dimensional | 0.996 | 0.987 | 0.555 |
≈ 26 experiments vs ~95 for random to reach 99% of a hidden optimum (3.7×). Reproduce: melete bench --robust.
Zero runtime dependencies + local cryptographic signing ⇒ the whole brain runs fully offline on an isolated machine — yet its discovery trace is verifiable by anyone with the public key alone. Built for regulated work where the process must stay inside the air gap but the result must still be provable.
Melete is not a model you train on data. You only provide two things: ① SPACE = the dials you can turn (+ their ranges), and ② a way to score one try — your real process (brew the coffee → taste it; train the model → read accuracy; set the price → read revenue). Melete proposes dial settings, you (or your script) return the score, it learns — generating its own data one experiment at a time.
— real use: point it at YOUR process, no formula —
melete tune --cmd "python train.py --lr {lr} --depth {depth}" \
--space '[{"name":"lr","type":"real","min":0,"max":0.1},{"name":"depth","type":"int","min":1,"max":12}]'
The browser demo below uses a math formula as a stand-in for that process, just so you can try it here. (It does not predict the future / fortunes — it optimises things you can actually measure + repeat.)
Pick an example, then press Discover.
It's a service whose users are agents / pipelines. POST your space + objective; get the optimum + a signed trace.
curl -s https://melete.161.35.122.73.nip.io/discover -H 'content-type: application/json' -d '{
"space":[{"name":"lr","type":"real","min":0,"max":0.1},{"name":"depth","type":"int","min":1,"max":12}],
"objective":"-(lr-0.03)**2*1000-(depth-6)**2", "budget":40, "goal":"maximize" }'