| # Runtime v0.0.2 |
|
|
| This release optimizes inference code; checkpoint weights and model intelligence are unchanged. |
|
|
| Feature encoding tokenizes each goal once, reuses lexical scores for retrieval, and |
| constructs tensors once per batch instead of repeatedly assigning tiny tensors. |
| Policy conversion no longer recursively copies unused DOM fields. |
|
|
| `LearnedPolicy.predict_batch([(goal, state, ticket), ...])` supports at most 64 |
| independent requests. Each returned decision retains its own ticket and observed |
| target references. Confidence abstention, sensitive-target exclusion, and value-copy |
| restrictions remain in effect. Empty batches return an empty list. |
|
|
| Measured on the development Windows CPU with two Torch threads and the mean |
| checkpoint: median single-request prediction 2.103 ms before / 0.884 ms after |
| (2.38x); p95 3.304 / 1.327 ms. A 16-request batch took 6.253 ms median, about |
| 0.391 ms per request. These are synthetic CPU policy timings, not browser-task |
| latency or target Linux VPS results. Loading, network, and browser time are excluded. |
|
|
| All encoded inputs, targets, and candidate maps were bit-identical across 1,440 |
| validation/test/novel-wording examples. Individual action/ticket parity was checked |
| on 160 requests; batch actions matched individual actions on those requests. |
| Novel-wording raw step match remains 55.83%; this release does not fix generalization. |
|
|
| Evidence: `reports/runtime-optimization.json`. Run `python -m unittest discover -s tests` |
| for runtime regression tests. The existing `baim.bench_policy` command can measure |
| the current feature encoder independently. Compare against Hub revision |
| `795f73703a1e33b82636b2167d7d9507581efd4d` for the prior implementation. |
|
|
