Roadmap

This roadmap records direction without delivery dates. It is not a compatibility promise; only the stability policy defines public compatibility commitments. Completed 1.0.0 work is recorded in the changelog, not here.

Study execution

Item	Status	Motivation	Measurable completion criterion	Dependency
Local study lock	Exploring	Prevent accidental competing mutators.	Two local processes cannot both acquire ownership; stale-lock recovery passes crash tests.	Portable lock semantics.
Durable task claims	Exploring	Make work assignment explicit before adding workers.	Every executable task transition carries one validated owner claim and rejects duplicate ownership.	Lock and journal design.
Leases	Planned	Bound ownership after worker loss.	Expired leases are reclaimed without allowing two valid commits for one attempt.	Durable task claims and a clock-skew policy.
Heartbeat	Planned	Distinguish slow work from abandoned ownership.	Fault-injection tests detect a stopped owner within a configured bound without disrupting a live owner.	Lease semantics.
Fencing	Planned	Reject writes from superseded owners.	A stale owner token cannot append events, checkpoints, attempts, or run references after reassignment.	Monotonic ownership epochs.
Bounded local worker pool	Planned	Use local CPU capacity without unbounded processes.	A configured worker limit is never exceeded and crash/retry tests preserve deterministic task identity.	Locks, claims, leases, heartbeat, and fencing.
Distributed/cross-host studies	Planned	Scale durable matrices beyond one machine.	A documented protocol passes duplicate-delivery, partition, reconnect, and recovery acceptance tests.	Stable local worker and ownership protocol.

Analysis

Item	Status	Motivation	Measurable completion criterion	Dependency
Statistical comparison	Exploring	Turn verified study outputs into attributable comparisons.	A versioned report records inputs, missingness, aggregation, and paired blocks for every statistic.	Stable study-to-run provenance.
Effect sizes	Planned	Report practical magnitude alongside significance.	Supported effect sizes have formula references, direction conventions, and independent numeric fixtures.	Statistical report schema.
Hypothesis tests	Planned	Make assumptions and multiplicity handling explicit.	Each test records hypotheses, assumptions, correction method, alpha, and deterministic fixture results.	Statistical report schema.
Attainment analysis	Planned	Compare stochastic fronts beyond scalar indicators.	Attainment surfaces reproduce reference fixtures within declared numerical tolerances.	Comparable objective scaling and verified fronts.
Publication reporting	Planned	Export reviewable tables and plots with provenance.	Every exported value links to a report field and ultimately to verified run IDs.	Stable analysis report.
Run checkpoints	Exploring	Support analysis of progress without weakening final-run integrity.	A versioned checkpoint format has bounded readers, integrity rules, and deterministic progress fixtures.	Public schema-evolution design.

Framework architecture

Item	Status	Motivation	Measurable completion criterion	Dependency
Declarative component descriptors	Exploring	Reconstruct components without importing arbitrary recorded code.	Built-ins round-trip through bounded JSON descriptors with schema and negative security tests.	Component identity model.
Machine-readable component catalog	Planned	Let tools discover verified capabilities.	The catalog reports stable IDs, encodings, dimensions, backends, and config schemas and matches registry tests.	Declarative descriptors.
Plugin-contract stabilization	Planned	Give third-party components an explicit compatibility boundary.	A versioned contract passes isolated conformance fixtures and documents deprecation rules.	Descriptors and catalog.
Backend-equivalence specification	Exploring	State which numerical comparisons are scientifically meaningful.	Each stable kernel has documented determinism/tolerance semantics and cross-backend oracle tests.	Independent numerical references.

Quality

Item	Status	Motivation	Measurable completion criterion	Dependency
Full-source mypy zero	In progress	Remove the frozen typing debt outside the strict/stable scopes.	Canonical mypy reports zero diagnostics for all src/vamos files and the ratchet is retired.	Incremental module cleanup without API churn.
Branch-coverage expansion	Planned	Exercise failure and recovery paths, not only statements.	A reviewed branch threshold covers critical run/study/security modules with documented exclusions only.	Stable, non-flaky coverage environment.
Broader notebook execution	Planned	Keep long-form learning material runnable.	Every primary notebook executes from a clean supported environment within a declared resource budget.	Curated notebook tiers and optional dependencies.
AGE-MOEA/RVEA/SMPSO oracles	Planned	Strengthen independent algorithmic evidence.	Each algorithm passes published small-case or independent implementation fixtures across multiple seeds.	Vetted reference data and tolerances.
Performance-regression framework	Exploring	Detect meaningful regressions without brittle timing claims.	Pinned benchmarks report distributions and flag preregistered effect-size thresholds on controlled runners.	Dedicated runners and baseline governance.

Artifacts and replay

Item	Status	Motivation	Measurable completion criterion	Dependency
Artifact authenticity/signing	Planned	Distinguish publisher identity from hash-based tamper detection.	Signed release and artifact verification succeeds against documented trust roots and fails for altered inputs.	Key-management and rotation policy.
Future public-schema evolution	Exploring	Evolve artifacts without stranding public data.	Every schema change has compatibility rules, golden old/new fixtures, and a tested reader transition.	Schema governance.
Explicitly non-exact reruns	Planned	Permit useful reruns when exact compatibility is impossible without mislabeling them.	A separately named mode records all substitutions and never reports exact; fixtures cover each downgrade reason.	Compatibility taxonomy and descriptor model.