Update without silently changing the contract.

This is the same reference used in Technical mode. Expand it whenever you need the exact numbers, assumptions or further explanation.

Improvement is a joint requirement

A model trained for a newer device can become worse on an older one. If the product promised to retain the older capability, a large new-task improvement is not permission to ignore that regression. A release decision needs a declared baseline, candidate, measurements and rule before looking at the preferred outcome.

This bench trains an actual thirteen-parameter tanh network. First train the old-task baseline for 100 updates and freeze its parameters. Then adapt a candidate using new-only or replay fitting. All updates use training coordinates. Separate monitoring grids measure old and new mean squared error and never enter gradients.

Calculate the gate

The versioned rule adaptation-release-v1 allows at most 0.02 old-task MSE regression and requires at least 0.01 new-task MSE improvement. Both conditions must hold. These teaching thresholds are fixed for this bench; they are not calibrated deployment safety standards.

Suppose baseline errors are 0.10 on old tasks and 0.80 on new tasks. A candidate produces 0.115 and 0.50. Its old regression is 0.015, below 0.02, and its new improvement is 0.30, above 0.01. Both pass. Another candidate with new MSE 0.20 but old MSE 0.14 fails because old regression 0.04 exceeds the protected limit.

The gate recomputes predictions from both saved parameter arrays. Imported scalar scores are not trusted. An eligible result records the exact checkpoints, suite version and thresholds, and a failed result retains its baseline instead. A frozen selection records a decision; it is distinct from automatic promotion.

Rollback preserves evidence

Rolling back makes the retained baseline the active model while keeping the candidate and release receipt available. Deleting the failed run would hide information needed to understand the next change. Compare the active-model label with the stored candidate; those objects answer different questions.

Changing replay strategy restores the original baseline and resets candidate updates. This makes a same-start comparison possible. Match update count, buffer capacity and frequency setting, and report the different new-versus-retained composition of fitting work. The no-adaptation control provides an explicit “do nothing” alternative.

What the record proves and what it does not

These monitoring outcomes are public and repeatedly visible. A gate built from them is development evidence, even after the candidate is frozen. A separate future evaluation needs a predeclared suite whose exposure is tracked; a new filename does not make old cases fresh.

The browser validates numerical bounds and recomputes metrics. It cannot authenticate that imported parameters were produced by the stated historical training sequence. Report that distinction alongside provenance. A text-presence check also cannot judge the quality of a decision brief or independently run learner code.

The cumulative adaptation project saves no-adaptation, new-only and replay checkpoints, asks for a retention/cost decision, and supplies separate objective transfer cases. Its implementation studio requires reproducible local outputs and all-run reporting. Independent human review and representative learner testing remain separate work.