This page documents what Limen intends to deliver over the next twelve months (2026-07 through 2027-06), what it intends not to do, and how this intent is kept current. Limen is a Bitcoin alpha research engine; the ranking below orders work by its contribution to credible alpha discovery per unit of research time, because that is the value the downstream Vaquum stack (Nexus, Praxis, Veritas) consumes.
Every planned item is anchored to an open GitHub issue. Items marked to be filed are committed intent without a filed issue yet; they enter delivery through the same PRD and slice process as everything else. A roadmap entry is intent, not a delivery date: work reaches main only through a filed slice, its gates, and review.
Horizon key
| Tag | Meaning |
|---|
| Near | targeted for 2026-H2 |
| Sustained | continuous through the whole horizon |
| Later | targeted for 2027-H1, sequenced behind Near work |
1. Research integrity and statistical rigor (Near)
The highest-value work: results that survive scrutiny. Benchmark today explicitly does not claim walk-forward validation, purged cross-validation, embargoed evaluation, or statistical acceptance gates; closing that disclaimed scope is the core of this theme.
| Item | Anchor |
|---|
| Per-round exception isolation in the UEL without losing research integrity | #602 |
Deterministic standard-path enumeration when random_search=False | #470 |
| Declarable experiment objective shared across threshold optimizer, feedback, and analysis | #533 |
| UEL experiment conclusion report | #211 |
| Sensitivity analysis for parameter sweeps | #213 |
| Leakage-safe evaluation: purged and embargoed walk-forward validation | to be filed |
| Statistical acceptance gates: multiple-testing-aware selection (deflated Sharpe ratio, probability of backtest overfitting) for sweep-scale results | to be filed |
| Item | Anchor |
|---|
source_columns attribute on the target interface for explicit column exclusion | #515 |
| Stationary shim layer for PCA-ready features | #518 |
| Deterministic regime classification columns | #504 |
| Explicit snapshot backtest support for multiclass and multilabel outputs | #490 |
| Explicit handling of probability-style snapshot inputs | #466 |
| Constraint checks on all indicator and feature input parameters | #252 |
3. Throughput and scale (Near)
| Item | Anchor |
|---|
Cache fitted manifest prep state deterministically when prep_each_round=True | #484 |
| Initialize and load historical data once per session | #461 |
| Saving the UEL object | #224 |
experiment_log as the sole canonical downstream surface, without per-permutation UEL state retention | #485 |
4. Data layer consolidation (Sustained)
| Item | Anchor |
|---|
| Replace remaining pandas usage with polars and drop pandas from dependencies | #643 |
| Move polars-to-sklearn handoff off the deprecated dataframe interchange protocol | #689 |
| Retype the pyarrow facade seam when the pin bumps past 23 | #680, #681 |
| Cohorts, indicator and feature taxonomy, and HistoricalData refactor | #369 |
| Unified market taxonomy for indicators and features | #346 |
5. Model and research surface (Later)
| Item | Anchor |
|---|
| Full TabPFN parameter surface, wider SFD parameters, and updated calibration path | #444 |
| Fold XGBoost research extensions into core Limen | #446 |
| Tradeline deferred remainder: multiclass and directional-conditional scope | #605 |
| Item | Anchor |
|---|
| Formalizing search strategies in experiments | #343 |
| Formalizing perturbation strategies in experiments | #341 |
| Formalizing pruning strategies in experiments | #344 |
| Formalizing linkage strategies in experiments | #342 |
| Comprehensive reference architectures and their foundational SFDs | #297 |
| Item | Anchor |
|---|
| Release provenance and supply-chain trust audit tracker | #626 |
| Test authority and required gates audit tracker | #625 |
| Manifest and YAML contract integrity audit tracker | #632 |
| Split the TabPFN model-download proof from the gate tracker | #628 |
| Master register burn-down for the 2026 audit findings | #693 |
| Property-based testing and fuzzing for parser and frame-contract surfaces | to be filed |
What Limen will not do
These are deliberate boundaries, not backlog. They mirror the product boundary in the README and the disclaimed scope in Benchmark:
- No trade execution and no downstream trade decisioning; Nexus, Praxis, and Veritas own those layers.
- No generic multi-asset research platform; Limen stays Bitcoin-native.
- No upstream source-of-truth market data infrastructure; Origo owns the data layer.
- No synthetic market data generation.
- No public leaderboard, benchmark corpus, or model card program.
- No hosted service or graphical interface; Limen remains a Python library and CLI.
How this roadmap stays current
The roadmap is reviewed whenever a PRD is filed and at least quarterly; changes land by pull request under the same documentation gates as every other page. Priority order may change between reviews; the anchored issues carry the authoritative current state of each item.
Read next