Repository navigation
Solvi 1.6.0
This release adds bounded, checkable learning and decision workflows to solvi.experimental.experience.
New capabilities
- Learn for decision value. Train a small integer linear ranker on the loss of its downstream decisions, validate it against an explicit baseline and activate it separately. Evaluate one-step logged feedback with exact rational IPS, SNIPS and doubly robust estimates, with explicit refusals for missing support, unknown logging probabilities or insufficient effective sample size.
- Cooperate through checked progress. Finite constraint and objective solvers exchange verified exclusions, feasible candidates and bounds. Shared residuals, source revisions, dependency retraction and independently checked finite certificates keep partial progress distinct from a completed solution.
- Teach executable concepts. Supply examples and exceptions, answer distinguishing questions, admit concepts on separate finite points, compose version-pinned programs and revoke dependent uses. The host provides the typed hypothesis language; examples alone do not grant verified authority.
- Clarify the goal. Maintain finite utility alternatives, ask costed preference questions, return minimax-regret certificates and expose conflicting answers or proposed soft-requirement relaxations. Hard requirements remain enforced.
Other changes since 1.5
- Finite reasoning with joint alternative states and explicit unknown results, externally evidenced outcomes, budgeted acquisition and executable experience.
- Counterexample-guided refinement, a paid one-pass program committee and bounded selection of useful observations.
- Packed numeric observation frames and feedback-driven stratified sampling.
- Verified dependency computations, selective invalidation after source changes, versioned operators and journal recovery.
- Acquisition and independently checked use of recurring parameterized operations.
- Updated API reference, a combined workflow guide and five new runnable examples covering the four decision capabilities and their composition.
Scope and evidence
These APIs remain experimental and deliberately finite. They require host-owned formalizations, trusted outcome evidence, explicit budgets and separate checks. They neither train a foundation model nor require a GPU or external model service.
Engineering correctness and research benefit are separate. The finite comparisons performed for the new decision workflows did not establish their predefined joint quality-and-cost advantage over the strong simple controls. Exact certificates apply only within the declared models and checked domains. Logged-policy estimates do not establish causal identification for arbitrary field data, and simulated preference questions do not validate human usability.
Existing sessions are not automatically switched to new policies. Snapshot recovery requires the corresponding trusted host authority and a continuous work ledger; an untrusted payload cannot mint authority or refund spent work.
Start with the combined workflow guide and the experience API.