Implementation roadmap

The roadmap is dependency-ordered. A later phase does not justify weakening an earlier contract.

Version progression

Roadmap phases are released as pre-1.0 minor versions because each phase adds public analytical capabilities. Patch versions are reserved for compatible corrections and hardening within a shipped phase; release candidates use Cargo/SemVer tags such as v0.2.0-rc.1 and PEP 440 package versions such as 0.2.0rc1.

Version	Roadmap scope	Release condition
0.1.x	Phases 0–3: foundations, signal diagnostics, validation core, audit/reports	Released
0.2.x	Phase 4: robustness, regimes, experiment lineage, basic multiple-testing correction	Released
0.3.x	Phase 5: trading realism, cost stress, liquidity, capacity	Released
0.4.x	Phase 6: point-in-time data correctness and bias detection	Released
0.5.x	Phase 7 plus deferred inference: permutation, Sharpe/PSR/DSR, CPCV/PBO, Reality Check/SPA	Released
0.6.x	Phase 8: separately optional adapters and extensions	Released
0.7.x	Portable, deterministic, independently verifiable evidence bundles	Released
0.8.x	Versioned cross-phase standardized audit profiles and evidence composition	Released
0.9.x	Persisted-artifact migration, integrated performance, diagnostics, and reference hardening	Released
0.10.x	Group-aware signal transformations, attrition, multi-lag stability, interactive evidence	Released
0.11.x	Validated decay inference, diagnostic portfolio projection, robust event studies	Released
0.12.x	Generic factor-panel interoperability and migration guidance	Released
0.13.x	PyPI-safe distribution identity, Trusted Publishing, and registry-install verification	Released
0.14.x	Admission-gated Rust migration and performance/memory hardening	Release-gated
1.0.0	Stable product contract in the technical specification	Every v1 definition item is evidenced, including independent use

This enumeration is a compatibility plan, not a schedule. A phase may receive multiple release candidates or patch releases, but Lacuna does not claim the next minor version until that phase's complete public contract is implemented and verified.

Phase 0 — foundations

Implemented foundation: packaging, typed configuration, result/finding models, Polars-first normalization, native bridge, baseline tests, CI, docs, and wheel build.

The audit-result v1 JSON Schema, persisted JSON/Markdown compatibility fixtures, Python end-to-end benchmarks, Criterion native-kernel benchmarks, conservative copy/materialization diagnostics, and domain-specific signal/label/price schema validation are implemented.

Phase 1 — signal diagnostics

v0.1 implemented: explicit forward labels, Pearson/Spearman IC, grouped IC time series, quantile returns and spread, monotonicity, rank turnover/autocorrelation, multi-horizon decay, and native grouped rank IC.

Deliver in this order:

1. Forward-return labels with explicit observation/entry/exit timing.
2. Pearson IC reference and grouped time series.
3. Spearman IC with documented ties and native grouped ranking.
4. Quantile assignment, returns, spread, and monotonicity components.
5. Turnover and signal autocorrelation.
6. Multi-horizon decay.
7. Structured Markdown report over the same results.

Exit criteria: Polars, NumPy, and optional pandas inputs agree; methodology is complete; native IC has differential tests and benchmarks.

Phase 2 — validation

v0.1 implemented: expanding/rolling walk-forward folds, interval-aware purging, observation-count embargo, and deterministic IID/moving/circular/stationary bootstrap confidence intervals.

1. Walk-forward split definitions and visualizable fold table.
2. Interval-aware purging and embargo.
3. IID reference bootstrap.
4. Moving block and stationary bootstrap with deterministic streams.
5. Later: permutation schemes.
6. Later: Sharpe uncertainty, then PSR and DSR with assumptions.

Exit criteria: no post-purge overlap in property tests; bootstrap calibration simulations pass; seeds reproduce across thread counts.

Phase 3 — audit and reports

v0.1 implemented: immutable audit contexts, independently executable versioned rules, explicit applicability, a 100-weight score with visible evidence coverage, deterministic JSON/Markdown/basic HTML, the SignalStudy facade, and the lacuna signal file workflow.

1. Rule protocol and applicability model.
2. Versioned finding codes and thresholds.
3. Audit context assembled from domain results.
4. JSON and Markdown renderers.
5. Optional escaped HTML renderer.
6. Versioned scoring policy with explicit missing-evidence penalties.

Exit criteria: absent evidence yields UNKNOWN; all rendered values trace to result fields; persisted audit fixture is versioned.

Phase 4 — robustness and experiments

v0.2 implemented: parameter surfaces, continuous perturbation, subperiods, regime analysis, universe perturbation, append-only experiment lineage, selection records, canonical fingerprints, and multiple-testing corrections.

Deliver for 0.2.0 in this order:

1. Canonical parameter encoding, fingerprints, immutable attempt records, and append-only local registry storage.
2. Explicit selection lineage over the complete eligible candidate set.
3. Parameter surfaces with visible failed points, mixed-type adjacency, boundary detection, and isolated-optimum evidence.
4. Deterministic continuous perturbation with rejection accounting.
5. Declared subperiod and universe perturbation tables with retained-sample/composition evidence.
6. Point-in-time or explicitly retrospective regime classification and conditional evidence.
7. Bonferroni, Holm, Benjamini-Hochberg, and Benjamini-Yekutieli adjustments over registered trial families.

Exit criteria: isolated optima and concentration are quantified; trial history is first-class provenance.

The exit criteria are covered by planted isolated/plateau and regime-concentration fixtures, deterministic perturbation tests, point-in-time threshold tests, append-only/concurrent registry tests, and the frozen 0.2.x public API contract.

Phase 5 — trading realism

v0.3 implemented: an immutable cost-model protocol and estimates; commission, observed/assumed spread, fixed/proportional/volatility-scaled slippage, participation/square-root impact, and borrow; deterministic grid/correlated stress scenarios; monotonic bracketed break-even solving; point-in-time or retrospective liquidity evidence; and multi-scenario capacity curves.

Exit criteria: cost monotonicity properties pass; outputs show assumptions and uncertainty rather than one false-precision capacity number.

The exit criteria are covered by hand-computed component fixtures, monotonicity/symmetry/ reconciliation properties, double-application guards, future-market-data tests, planted break-even/impact/capacity cases, explicit unknown liquidity and borrow evidence, Polars/pandas/Arrow equivalence, a frozen 0.3.x public contract, and a versioned end-to-end stress benchmark. The benchmark supports the NumPy/Polars path; no unevidenced native cost path is shipped.

Phase 6 — data correctness

v0.4 implemented: point-in-time as-of joins with deterministic revision ties, explicit availability/revision semantics, direct future-data checks and optional materiality, structural revision diagnostics, three-state survivorship evidence, half-open membership selection with an availability firewall, consecutive universe drift, and declarative dataset validation.

Exit criteria: safe joins never select unavailable records; unknown survivorship handling remains visible.

The exit criteria are covered by exact-boundary, one-nanosecond-future, timezone, staleness, revision, delisted-asset, overlap, future-known-membership, drift, and dataset-defect fixtures; generated join/membership invariants; Polars/pandas/Arrow equivalence; a frozen 0.4.x public contract; an end-to-end as-of benchmark; and clean-wheel exercise of the bias path.

Phase 7 — advanced inference

v0.5 implemented: combinatorial purged K-fold with inspectable reconstructed paths; explicit permutation nulls; Sharpe uncertainty, PSR, DSR, and minimum track-record length; symmetric CSCV/PBO with visible selection/ranks/logits and partition sensitivity; joint stationary bootstrap; White Reality Check; and Hansen SPA with lower, consistent, and upper recenterings.

Exit criteria: every method has an independent equation-level reference or deterministic stream fixture, a fixed-seed simulation for its statistical behavior, explicit assumptions and failure states, bounded combinatorial work, a frozen additive 0.5.x API, clean-wheel exercise, and a versioned public-call benchmark.

The exit criteria are covered by CPCV combination/path and no-overlap invariants; permutation and Sharpe null simulations; complete-family DSR checks; planted PBO stable-edge/forced-overfit cases; independent literal White and Hansen implementations; Reality Check/SPA size, power, and poor-alternative simulations; benchmark artifact v4; and the advanced-inference methodology contract. No native advanced-inference path is claimed without a future benchmark-justified, differentially tested implementation.

Phase 8 — extensions

v0.6 implemented: optional integrations are explicit boundaries rather than new research methodology. Core adds no mandatory runtime dependency.

Delivered in dependency order:

1. Immutable AdaptedFrame evidence and collision-safe canonical/source normalization.
2. DuckDB result-to-Arrow-stream ingestion without pandas or SQL construction.
3. A dependency-free scikit-learn CV protocol bridge over frozen Lacuna temporal folds.
4. Versioned vendor mappings with availability, revision, timezone, adjustment, and identity declarations.
5. Generic returns/trades/positions mappings that require complete backtest timing, gross/net, compounding, costs, borrow, calendar, session, missing-asset, and delisting semantics.
6. Domain-specific entry-point groups with metadata-only discovery, deterministic conflict handling, explicit trusted activation, and protocol/capability negotiation.
7. A separate lacuna-options 0.1 distribution with validated empirical chains, carry forwards, log-forward moneyness, absolute-delta buckets, and supplied-expectation IV residuals.

Exit criteria: optional packages do not load at core import time; real DuckDB and scikit-learn paths interoperate; discovery never imports targets; activation is explicit and evidenced; adapters reject ambiguous temporal/methodological semantics; core 0.6.x and extension 0.1.x public contracts are frozen independently; both distributions build, clean-install together, and enter the checksummed, attested release set.

The release criteria are covered by real DuckDB and scikit-learn interoperability, adapter/plugin unit and contract fixtures, extension property/adversarial tests, independent frozen API contracts, clean joint wheel installation, the complete cross-platform CI matrix, and the tagged checksum and provenance workflow.

Deferred beyond the phase: DataFusion/query pushdown, framework-specific vectorbt/LEAN/Nautilus helpers, plugin isolation or marketplace behavior, universal IV/Greeks solvers, SVI calibration, arbitrage repair, and delta-hedged options attribution. They require their own semantic and numerical contracts and are not implied by the generic boundaries.

Integration and v1 hardening (0.7–0.9)

After Phase 8, work is driven by cross-phase use rather than unchecked feature count:

• run complete signal-to-audit workflows over vendor/backtester inputs and extension evidence;
• reconcile configuration, provenance, and error behavior across every public subsystem;
• profile realistic end-to-end workloads and optimize only evidenced bottlenecks;
• test migrations and persisted evidence across release lines;
• collect independent-user feedback and harden installation, diagnostics, and documentation;
• close every remaining v1 definition item in the technical specification or explicitly remove it from the stable contract.

0.7–0.9 do not promise one feature per minor version. A release needs a coherent compatibility milestone, migration story, and complete evidence gate.

The first coherent integration milestone is 0.7: deterministic .lacuna archives over immutable reports, configuration/environment summaries, and optional structured evidence. Bundle v1 has a published schema, byte-stable writer, strict non-executing verifier, SHA-256 manifest, privacy redaction/fail-closed behavior, CLI support, hostile-archive tests, frozen additive API, and clean-wheel exercise. It claims identifiable reproducibility only; recomputability and numerical or bitwise reproduction across independent executions are not implied.

The second coherent integration milestone is 0.8: standardized signal, strategy, and options profiles over every released evidence method family. The profile schema and API are frozen; required, optional, and not-applicable coverage is explicit by category; domain findings are propagated without reinterpretation; no universal score is emitted; strict v1 result JSON and CLI composition are bounded; vendor, backtester, options, bundle, installed-wheel, and release archive paths are executable.

The third coherent integration milestone is 0.9: an explicit identity-migration matrix and tagged-fixture corpus for persisted formats; strict profile/manifest readers; versioned, non-invasive installation diagnostics; benchmark artifact v5 with a profiled integrated strategy workflow; and an exact machine-checked reference route for every supported export. No speculative performance path was added when profiling found no justified semantic-preserving change.

The current status of every stable-release criterion and the remaining external 1.0 boundary live in the v1 readiness ledger.

Factor-research ergonomics (0.10–0.12)

These milestones adopt proven factor-research ergonomics without turning Lacuna into a notebook plotting package, cumulative backtester, or order simulator.

0.10 introduces immutable BucketSpec/SignalTransformResult contracts, explicit bucket-return analysis, availability-aware grouping, weighted least-squares neutralization, reconciled attrition, exact multi-lag turnover, retained study evidence, and an opt-in Plotly renderer over stored rows. Core rendering and schema-v1 audit JSON remain unchanged.

0.11 adds a validated positive exponential decay fit with dependence-aware joint resampling, explicit diagnostic portfolio weights without compounding or execution simulation, and event windows anchored to availability with clustered path inference. The portfolio boundary requires an ADR before implementation.

0.12 adds generic factor-panel schemas and fully declared timing semantics across Polars, Arrow, and optional pandas MultiIndexes. It publishes migration guidance and frozen compatible numeric fixtures pinned to a reviewed upstream commit without a runtime Alphalens dependency.

Excluded throughout: lookahead z-score filtering, implicit calendars or timing, mandatory pandas or plotting, cumulative backtesting, order/fill simulation, and plotting side effects.

Admission-gated performance (0.14)

0.14 preserves the complete 0.13 public API, schema-v1 result envelope, analytical method versions, c14n-v1 identities, and deterministic RNG streams. It profiles every registered migration candidate against an already optimized NumPy or Polars reference and ships native code only after correctness plus an end-to-end throughput, memory, or bounded-completion gate.

The implementation and candidate-admission phases are complete. Most candidates finish as OPTIMIZED_NON_NATIVE; shared resampling is NOT_MIGRATING, and deferred identity/public-carrier work is BLOCKED. Grouped-rank IC and PBO/CSCV are SHIPPED_NATIVE after pinned Linux evidence, the same cp311-abi3 wheel on Python 3.11–3.14, and every target-wheel smoke job reproduced their contracts. The exact-SHA non-publishing preflight remains a tag gate, and the tag workflow refuses any non-terminal decision.

Native work stays single-threaded. Python retains timing semantics, methodology, validation, findings, provenance, RNG, and public result construction. Arrow C Data, native RNG/scheduling, Rayon, public compact carriers, c14n-v2, approximate quantiles, and semantic method changes remain outside this milestone.

Historical v0.1 boundary

v0.1 turns a cross-sectional signal into a rigorous diagnostics report. It excludes live trading, full backtesting, distributed execution, GPU kernels, advanced impact simulation, and a plugin marketplace.

The release is ready when the narrow path is correct, reproducible, benchmarked, documented, and usable across supported dataframe inputs—not when every long-term module exists.