Versions
Ten versions to earn a candidate.
Each version added one capability and shipped its own report. v1.0 is a release candidate rather than a version bump, because the project could still name four things it had not established.
The arc
Prototype.
Hybrid route structure.
Adaptive selection.
Calibrated cost model.
Governance and signing; the release public key dates from here.
Bandit runtime.
20-seed LinUCB / Thompson / discounted Thompson comparison, the drift state machine, and the first MMLC conformance work.
Confidence-gated route portfolio, bootstrap stability, measured probe cost, and an adaptive one-step EVSI policy.
Strict outer-fold separation, split-conformal benefit lower bounds, and the measured batch-probe crossover.
Three stable modes, the CLI, machine calibration, release binding, and four ledgers that replay under the official MMLC Runtime.
Verification
From TEST_RESULTS_V1.md.
| Check | Result |
|---|---|
| Smoke tests | 4 of 4 PASS |
| CLI assurance factor | PASS |
| CLI release certificate verification | PASS |
| CLI machine-calibrated batch | PASS |
| Machine profile verification | PASS |
| Assurance model verification | PASS |
| Performance certificate verification | PASS |
| Assurance certificate verification | PASS |
| Python wheel build | PASS — mmpf-1.0.0rc1-py3-none-any.whl |
| Official MMLC replay | 4 of 4 PASS — first run 2026-08-01 |
| Synthetic machine profiles | format and sensitivity only |
Repository
- Repository
mmpf-runtime- Distribution
mmpf- CLI
mmpf- Licence
- Apache-2.0
- CI
- the official MMLC conformance replay, on Python 3.11, 3.12 and 3.13
The CI is the conformance replay rather than a unit-test run, which is a deliberate choice: the thing most worth defending on every push is that the ledgers still mean what they say to a runtime that did not produce them.